The fastest and most efficient code intelligence engine for AI coding agents. Full-indexes an average repository in milliseconds, the Linux kernel (28M LOC, 75K files) in 3 minutes. Answers structural queries in under 1ms. Ships as a native executable with a small verified runtime-asset set for macOS, Linux, and Windows — download, run install, done.

High-quality parsing through tree-sitter AST analysis across all 162 languages, enhanced with Hybrid LSP semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl — producing a persistent knowledge graph of functions, classes, call chains, HTTP routes, and cross-service links. 17 MCP tools. No language runtime, hosted service, or API key. Plug and play across 45 supported automatic/conditional client surfaces.

Research — The design and benchmarks behind this project are described in the preprint Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP (arXiv:2603.27277). Evaluated across 31 real-world repositories: 83% answer quality, 10× fewer tokens, 2.1× fewer tool calls vs. file-by-file exploration.

Security & Trust — This tool reads your codebase and writes to your agent configuration files. That is what it is designed to do. If you prefer to audit before running, the full source is here. For each release product, three behaviourally identical executable candidates (unstripped, debug-stripped, stripped) are submitted to VirusTotal before testing; the selected candidate is then packaged with its SHA-256 unchanged. Release notes link every measured candidate result. Publication permits only the narrowly documented single-Microsoft !ml tolerance in SECURITY.md. All processing happens 100% locally; your code never leaves your machine. Found a security issue? We want to know — see SECURITY.md. Security is Priority #1 for us.

Built-in 3D graph visualization — explore your knowledge graph at localhost:9749

• Extreme indexing speed — Linux kernel (28M LOC, 75K files) in 3 minutes. RAM-first pipeline: LZ4 compression, in-memory SQLite, fused Aho-Corasick pattern matching. Memory released after indexing.
• Plug and play — native executable plus authenticated release-owned assets for macOS (arm64/amd64), Linux (arm64/amd64), and Windows (amd64). The native install needs no Docker, language runtime, or API keys. Download → install → restart agent → done.
• 162 languages — vendored tree-sitter grammars compiled into the binary. Nothing to install, nothing that breaks.
• 120x fewer tokens — 5 structural queries: ~3,400 tokens vs ~412,000 via file-by-file search. One graph query replaces dozens of grep/read cycles.
• 45 supported automatic/conditional client surfaces — install configures detected clients and safely activates conditional clients only when their documented platform, marker, or explicit existing config path is present. See Multi-Agent Support for the complete matrix and manual/UI-only boundaries.
• Built-in graph visualization — 3D interactive UI at localhost:9749, served from the binary itself.
• Infrastructure-as-code indexing — Dockerfiles, Kubernetes manifests, and Kustomize overlays indexed as graph nodes with cross-references. Resource nodes for K8s kinds, Module nodes for Kustomize overlays with IMPORTS edges to referenced resources.
• 17 MCP tools — search, trace, architecture, impact analysis, targeted index-coverage checks, Cypher queries, dead code detection, cross-service HTTP linking, ADR management, and more.

One-line install (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash

With graph visualization UI:

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash

Windows (PowerShell):

# 1. Download the installer
Invoke-WebRequest -Uri https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.ps1 -OutFile install.ps1

# 2. (Optional but recommended) Inspect the script
notepad install.ps1

# 3. Unblock the downloaded file (removes Mark-of-the-Web restriction added by browsers/Invoke-WebRequest)
Unblock-File .\install.ps1

# 4. Run it
.\install.ps1

Note: If you see a script execution policy error, run Set-ExecutionPolicy -Scope Process Bypass first, or invoke with PowerShell -ExecutionPolicy Bypass -File .\install.ps1.

Options: --skip-config (binary only, no agent setup), --dir=<path> (custom location).

Antivirus note: Microsoft Defender may flag a release binary as Trojan:Script/Wacatac.B!ml. This is a known false positive — typically 61 of ~62 engines return clean, and the same detection family hits gh, llama.cpp, Godot and Microsoft's own Go toolchain. See Antivirus False Positives for the evidence, how to verify the artifacts yourself, and how to report it if you think we are wrong.

Restart your coding agent. Say "Index this project" — done.

The install command auto-detects installed coding agents and configures their documented MCP entries plus durable instructions, skills, and lifecycle hooks where supported.

CBM automatically shares one per-account coordination daemon across Claude Code, Codex, OpenCode, and every other configured client. There is no opt-in setting for MCP servers or hook clients: the first daemon-backed CBM session starts it, each session registers its own work, and the final session shuts it down. The daemon owns long-lived background services such as watchers, shared indexing jobs, and the optional UI. Closing one session cancels work owned only by that session, while work still needed by another session continues.

The detached daemon does not depend on an MCP frontend's stderr. It keeps owner-only durable records under the canonical ${CBM_CACHE_DIR}/logs directory (default ~/.cache/codebase-memory-mcp/logs):

Thin frontends still write immediate startup and session-specific errors to their own stderr; MCP JSON-RPC stdout remains clean.

All active CBM processes must run the exact same version, executable build, coordination ABI, and canonical cache root. Equivalent CBM_CACHE_DIR aliases resolve to the same root; a genuinely different root is rejected while any CBM process is active. MCP servers, hooks, one-shot CLI commands, temporary index workers, and the daemon share a crash-safe OS admission barrier; starting an ordinary conflicting process fails before doing work and records an explicit conflict in ${CBM_CACHE_DIR}/logs/daemon-conflicts.ndjson.

The native install, update, and uninstall commands are the deliberate exception to that conflict rule. Download, verification, and private same-filesystem staging happen first so a bad candidate never disrupts active work. Activation then publishes account-wide maintenance intent, asks the daemon and every temporary local operation to cancel, and waits to a finite deadline for all coordinated CBM processes to exit. It holds the admission and lifetime barriers exclusively while changing the active binary, configuration, PATH, or indexes. New CBM work cannot enter during this window. Activation progress and results are recorded in ${CBM_CACHE_DIR}/logs/activation-events.ndjson, and a successful command tells you to restart open coding-agent sessions so they launch the activated build.

Package-manager setup (npm, PyPI, or Go) verifies and publishes a coherent private cached runtime set. Sidecars are replaced before the executable with per-file atomic renames; an interrupted multi-file publication is detected and repaired on the next launch rather than being described as one crash-atomic filesystem transaction. It does not replace the active native installation and therefore does not stop running CBM sessions. When that cached binary is executed, it still enters the same exact-build admission barrier. The shell and PowerShell installers invoke the verified candidate's native install command, so they do receive the full account-wide activation guarantee.

The ordinary cli mode runs one command locally, but it is not disconnected from the daemon above: it connects to the shared coordination daemon — starting one if none is already running — for that same admission barrier plus per-project locks on graph mutations. Its connection is a cli_session: it never registers with the background watcher and never starts a UI, and it holds the admission lease only for the command's lifetime. If the command's own connection is what started the daemon, that daemon exits again once the command closes and no other session is attached; if a daemon was already running, the command simply joins and leaves it exactly as found. While the command is running, a temporary monitor separately lets activation cancel that operation and its supervised worker safely; the monitor exits with the command and never becomes a standing daemon. See CLI Mode for details.

The graph UI is built into the binary — every install on every channel has it. Then run it:

codebase-memory-mcp --ui=true --port=9749

Open http://localhost:9749 in your browser. The UI is owned by the shared coordination daemon, so concurrent agent sessions do not start duplicate HTTP servers.

Enable automatic indexing on MCP session start:

codebase-memory-mcp config set auto_index true

When enabled, new projects are indexed automatically on first connection. Previously-indexed projects are registered with the background watcher for ongoing git-based change detection. Configurable file limit: config set auto_index_limit 50000.

Watcher registration is controlled separately by auto_watch (default true). Set config set auto_watch false to keep a session from registering its project with the background watcher — useful when working across many projects and you want each session contained to explicit indexing.

To turn the watcher off entirely, set config set watcher_enabled false (default true): the background poll thread never starts and no project is registered, while auto_index and manual index_repository keep working. Unlike auto_watch — which is consulted per session — watcher_enabled is read once when the background daemon starts, so run codebase-memory-mcp daemon stop after changing it; reconnecting your MCP client alone will not restart the daemon. See docs/CONFIGURATION.md.

Updates run from the install script on every platform, not from inside the running binary. codebase-memory-mcp update validates your flags and then prints the exact command to run:

# macOS / Linux
bash "<install-dir>/install.sh"

# Windows
powershell -ExecutionPolicy Bypass -File "<install-dir>\install.ps1"

The install script is placed next to the binary at install time, so the printed path resolves beside the executable. It is idempotent, so re-running it is the update: it stops the daemon, retires the running binary, installs the new one, and cleans up.

Why it works this way. On Windows it is a hard requirement — a running executable cannot replace its own image, so the swap has to happen from a process that is not the binary being replaced. On macOS and Linux it is a deliberate choice: an in-process updater is structurally a downloader (fetch an archive, verify it, unpack it, mark a file executable, run it), and shipping that composite in every binary to serve a command most people run a handful of times is a poor trade. The release archives now carry no download URLs at all, and cbm makes no network request of its own accord — it does not check for new versions in the background, and nothing phones home. You find out about releases from the install script, your package manager, or GitHub.

If PowerShell refuses to run the script because the file came from the internet, Unblock-File it first.

Installed through npm or pip? Update with your package manager on every platform (npm install -g codebase-memory-mcp@latest / pip install -U codebase-memory-mcp).

codebase-memory-mcp uninstall

Removes owned agent config entries, skills, hooks, instructions, and the installed binary. Existing graph indexes are listed and deleted only after confirmation.

The install script placed beside the binary is reported, not deleted — uninstall prints its path and the rm command for it. It is left alone on purpose: it may be your own copy, a symlink into a checkout, or managed by a package manager, and an uninstaller should not delete a file it cannot prove it owns.

• Architecture overview: get_architecture returns languages, packages, entry points, routes, hotspots, boundaries, layers, and clusters in a single call
• Architecture Decision Records: manage_adr persists architectural decisions across sessions
• Louvain community detection: Discovers functional modules by clustering call edges
• Git diff impact mapping: detect_changes maps uncommitted changes to affected symbols with risk classification
• Call graph: Resolves function calls across files and packages (import-aware, type-inferred)
• Dead code detection: Finds functions with zero callers, excluding entry points
• Cypher-like queries: MATCH (f:Function)-[:CALLS]->(g) WHERE f.name = 'main' RETURN g.name

• Semantic search (the semantic_query parameter of search_graph, not a separate tool): vector search across the entire graph, powered by bundled Nomic nomic-embed-code embeddings (40K tokens, 768d int8) compiled into the binary — no API key, no Ollama, no Docker. 11-signal combined scoring (TF-IDF, RRI, API/Type/Decorator signatures, AST profiles, data flow, Halstead-lite, MinHash, module proximity, graph diffusion).
• BM25 full-text search via SQLite FTS5 with cbm_camel_split tokenizer (camelCase / snake_case aware)
• Structural search (search_graph): regex name patterns, label filters, min/max degree, file scoping
• Code search (search_code): graph-augmented grep over indexed files only

• HTTP route ↔ call-site matching with confidence scoring
• gRPC, GraphQL, tRPC service detection with protobuf Route extraction
• Channel detection (EMITS / LISTENS_ON) for Socket.IO, EventEmitter, and generic pub-sub patterns across 8 languages with constant resolution

• CROSS_* edges link nodes across multiple repos indexed under the same store
• Multi-galaxy 3D UI layout for cross-repo architecture visualization
• Cross-repo architecture summary combining services, routes, and dependencies across the indexed fleet

• CALLS — a callable is invoked at the source site
• CALL_REFERENCE — a callable is used at a supported reference site (for example, a direct value argument) and resolves to one exact target
• USAGE — an identifier is used, but a unique callable target is not proven (including ambiguous or complex expressions)
• IMPORTS, DEFINES, IMPLEMENTS, INHERITS
• HTTP_CALLS, ASYNC_CALLS (cross-service)
• EMITS, LISTENS_ON (channels)
• DATA_FLOWS with arg-to-param mapping + field access chains
• SIMILAR_TO (MinHash + LSH near-clone detection, Jaccard scored)
• SEMANTICALLY_RELATED (vocabulary-mismatch, same-language, score ≥ 0.80)

• 158 vendored tree-sitter grammars compiled into the binary
• Generic package / module resolution — bare specifiers like @myorg/pkg, github.com/foo/bar, use my_crate::foo resolved via manifest scanning (package.json, go.mod, Cargo.toml, pyproject.toml, composer.json, pubspec.yaml, pom.xml, build.gradle, mix.exs, *.gemspec)
• Infrastructure-as-code indexing — Dockerfiles, Kubernetes manifests, Kustomize overlays as graph nodes
• Hybrid LSP semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl — a lightweight C implementation of language type-resolution algorithms, structurally inspired by and compatible with major language servers including tsserver / typescript-go, pyright, gopls, Roslyn, Eclipse JDT, and rust-analyzer (parameter binding, return-type inference, generic substitution, JSX component dispatch, JSDoc inference for plain JS files, namespace + trait + late-static-binding resolution for PHP, file-scoped namespaces + records + LINQ method syntax for C#, class-hierarchy + overload + lambda resolution for Java, extension-function + scope-function resolution for Kotlin, trait-method + UFCS resolution for Rust)
• RAM-first pipeline: LZ4 compression, in-memory SQLite, single dump at end. Memory released after.

• Native runtime set, zero infrastructure services: SQLite-backed, persists to ~/.cache/codebase-memory-mcp/
• Auto-sync: Background watcher detects file changes and re-indexes automatically
• Route nodes: REST endpoints are first-class graph entities
• CLI mode: codebase-memory-mcp cli search_graph '{"project": "my-project", "name_pattern": ".*Handler.*"}'
• Available on: npm, PyPI, Homebrew, Scoop, Winget, Chocolatey, AUR, go install

Commit a single compressed file to your repo and your teammates skip the reindex.

.codebase-memory/graph.db.zst is a zstd-compressed snapshot of the knowledge graph that lives next to your source. The artifact is opt-in: index_repository writes it only with persistence: true (the default is false), and once it exists later indexes refresh it; when a teammate clones the repo and runs codebase-memory-mcp for the first time, the artifact is decompressed and incremental indexing fills in their local diff.

• Format: SQLite database, indexes stripped, VACUUM INTO compacted, then zstd 1.5.7 compressed (8–13:1 ratio typical)
• Two tiers:
  • Best (zstd -9 + index strip + VACUUM INTO) — written by index_repository with persistence: true
  • Fast (zstd -3) — refreshes an existing artifact on every other index (including the watcher's low-latency incremental updates)
• Bootstrap: when no local DB exists but the artifact is present, index_repository imports the artifact first, then runs incremental indexing — avoiding the full reindex cost
• No merge pain: a .codebase-memory/.gitattributes line with merge=ours is auto-created on first export, so concurrent edits don't produce conflicts on the binary artifact
• Commit it deliberately: the artifact is rewritten on every index, including the watcher's Fast tier, and git stores each rewrite as a full new blob. Committing every refresh is what turns a 20 MB file into gigabytes of history — one team reached ~6 GB across ~350 commits of this single path. Pick a cadence (a release, a milestone, a nightly job) rather than committing every save.
• Git LFS, if it must move on every commit: track it from the repo-root .gitattributes and leave the auto-created .codebase-memory/.gitattributes in place — the nearer file goes on supplying merge=ours, and only filter comes from the root:

  .codebase-memory/graph.db.zst filter=lfs diff=lfs merge=lfs -text

  Track only the .zst; artifact.json is small and carries the schema version. The attribute applies to future commits only, so a repo that already has the blobs in history needs git-filter-repo to rewrite them first. Two costs to weigh before adopting it: GitHub meters LFS storage and bandwidth, and its objects cannot be pruned without contacting support; and every teammate needs git lfs install — without it their checkout leaves a pointer file where the artifact should be, the integrity-checked import refuses it, and they fall back to a full reindex.
• Optional: never committed unless you want it. Add .codebase-memory/ to .gitignore if you prefer everyone to reindex from scratch.

The result is similar in spirit to graphify's graphify-out/ directory, but as a single compressed file with explicit two-tier export, integrity-checked import, and zero merge friction.

codebase-memory-mcp is a structural analysis backend — it builds and queries the knowledge graph. It does not include an LLM. Instead, it relies on your MCP client (Claude Code, or any MCP-compatible agent) to be the intelligence layer.

You: "what calls ProcessOrder?"

Agent calls: trace_path(function_name="ProcessOrder", direction="inbound")

codebase-memory-mcp: executes graph query, returns structured results

Agent: presents the call chain in plain English

Why no built-in LLM? Other code graph tools embed an LLM for natural language → graph query translation. This means extra API keys, extra cost, and another model to configure. With MCP, the agent you're already talking to is the query translator.

Benchmarked on Apple M3 Pro:

RAM-first pipeline: All indexing runs in memory (LZ4 HC compressed read, in-memory SQLite, single dump at end). Memory is released back to the OS after indexing completes.

Token efficiency: Five structural queries consumed ~3,400 tokens via codebase-memory-mcp versus ~412,000 tokens via file-by-file grep exploration — a 99.2% reduction.

To measure comparable quality, latency, and agent-efficiency metrics on your own workload, see Measuring quality, latency, and agent savings. Exact reproduction of the figures above requires the original inputs and raw artifacts.

codebase-memory-mcp runs 100% locally and collects no telemetry — your code, queries, environment, and usage never leave your machine. That privacy guarantee also means that when you hit something we can't reproduce on our side (a slow memory climb over hours, a performance regression, a leak that only appears after days of real use), we have no data at all unless you choose to send it. Here is how to capture it yourself.

Set CBM_DIAGNOSTICS=1 before the first daemon-backed MCP session starts, then reproduce the problem (let it run as long as it takes — a slow leak needs time to show in the trend). The shared daemon captures this setting from the session that starts it. If it is already running, close all daemon-backed sessions so it exits before changing the setting. The daemon creates a fresh owner-private cbm-diagnostics-<pid>-<random> directory below the system temp directory ($TMPDIR or /tmp on macOS/Linux, %TEMP% on Windows). The exact paths are recorded by the diagnostics.start event in ${CBM_CACHE_DIR}/logs/cbm-daemon.log:

The private randomized directory prevents another local account from pre-placing a link or special file at a predictable diagnostics path. Its <pid> component is the shared daemon's process ID, also recorded by the daemon.start event. Set the variable consistently in the env block of each agent's MCP server config, or export it before launching the first session.

When you open a memory/performance issue, attach the .ndjson trajectory — it contains no source code or query text, only resource counters. If you'd rather not attach a file, paste it (or an agent's summary of it) into the issue: your assistant can read the NDJSON directly and report whether rss/committed grow monotonically, how fast, and relative to query count — which is exactly what we need to find the cause.

Every release includes checksums.txt with SHA-256 hashes. The executable is self-contained — no adjacent data file is required. Linux -portable archives contain the fully static builds; ordinary platform archives use their native system ABI.

Windows note: SmartScreen may show a warning for unsigned software. Click "More info" → "Run anyway". Verify integrity with checksums.txt.

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/scripts/setup.sh | bash

irm https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/scripts/setup-windows.ps1 | iex

yay -S codebase-memory-mcp-bin

paru -S codebase-memory-mcp-bin

The codebase-memory-mcp-bin package is available at: https://aur.archlinux.org/packages/codebase-memory-mcp-bin

The flake exposes two server packages plus the standalone frontend:

Run directly without installing:

# Standard server
nix run github:DeusData/codebase-memory-mcp

# Server with the embedded graph UI
nix run github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui -- --ui=true --port=9749
# then open http://127.0.0.1:9749

Or build a binary into ./result/bin/codebase-memory-mcp:

nix build github:DeusData/codebase-memory-mcp                          # standard
nix build github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui   # with UI

Working in a clone? Use . in place of the flake URL, e.g. nix run .#codebase-memory-mcp-ui -- --ui=true, or drop into a shell that puts the binary on PATH with nix shell .#codebase-memory-mcp-ui.

Note: launched by hand (not from an MCP client) the server exits as soon as stdin closes — that's normal MCP behaviour. Keep stdin open while testing the UI, e.g. sleep infinity | codebase-memory-mcp --ui=true --port=9749. The codebase-memory-mcp-ui package embeds the UI at build time; nix run'ing the standard default package with --ui=true will refuse to start the HTTP server.

You: "Install this MCP server: https://github.com/DeusData/codebase-memory-mcp"

git clone https://github.com/DeusData/codebase-memory-mcp.git
cd codebase-memory-mcp
scripts/build.sh --with-ui          # the shipped composition (graph UI embedded)
scripts/build.sh                    # without the UI (development only)
# Binary at: build/c/codebase-memory-mcp   (codebase-memory-mcp.exe on Windows)

Every platform ships one self-contained executable: the graph UI and the agent integration templates are linked into the binary, so an extracted archive is immediately complete.

Run the test suite (8,060 tests across 141 suites):

scripts/test.sh                     # full: clean sanitizer build + all suites + guards
scripts/test.sh --suites <name>     # one suite, incremental, seconds
build/c/test-runner --list-suites   # what is available

scripts/test.sh is the same entry the CI gates run, so a local pass means the same thing a CI pass does. The canonical local artifact-flow check builds both stripped/unstripped candidates, defaults to the stripped candidate for this explicitly unscanned local run, packages those exact bytes, extracts the archive, and smokes it:

scripts/ci/smoke-artifact.sh <linux|darwin|windows> <amd64|arm64>

scripts/package-release.sh is intentionally a lower-level immutable boundary: it accepts only an already-final --selected-binary plus its --expected-sha256; it never builds, strips, signs, or relinks the executable.

{
  "mcpServers": {
    "codebase-memory-mcp": {
      "command": "/path/to/codebase-memory-mcp",
      "args": []
    }
  }
}

install configures 45 client surfaces: 39 detected automatically and 6 conditional or explicit. “Conditional” means the installer writes only when the documented platform or an explicit, already-existing config path proves the target is active. It never flips experimental feature flags, enables plugins, YOLO modes, global permission bypasses, or third-party instruction trust.

Where a client has a documented custom-agent format, the installer creates three exact-owned definitions from one canonical contract:

• Scout (Tier 1) — about 3–4 narrow calls for fast positive, provisional discovery; no absence, exhaustive-impact, or dead-code claims.
• Verify (Tier 2, default) — task-directed graph evidence, exact source checks, path coverage for every cited file, and scope coverage before negative claims.
• Auditor (Tier 3) — bounded scope, current index generation, complete relevant pagination, broader relationship checks, and explicit unresolved limitations.

Every direct tier batches check_index_coverage for its evidence paths and reads flagged ranges or skipped/excluded files directly. A clean coverage result means only “no recorded gap,” never proof of completeness. Clients without safe child MCP access receive the same three tiers as parent-handoff agents; the parent must supply project, generation, pagination state, graph evidence, and coverage results. Updates migrate only byte-identical prior Verify definitions and never overwrite user-modified agents.

For Codex, install keeps only a tiny managed activation pointer in global $CODEX_HOME/AGENTS.md; all detailed behavior lives in the installed codebase-memory skill. Fresh installs create the pointer, upgrades replace the legacy full managed block while preserving all user-owned bytes, and uninstall removes only the managed pointer.

Hooks installed by this project are fail-open and context-only. Claude Code's PreToolUse observes Grep/Glob/Bash and injects matching graph symbols as additionalContext; PostToolUse on Read adds targeted coverage context when the graph could not fully parse or index that file. It never denies or replaces the requested tool call.

Claude Code, Codex CLI, Qwen Code, GitHub Copilot CLI, and VS Code's Copilot runtime receive paired session/subagent context where the vendor exposes a documented context-output contract. Codex users must review and trust installed hooks through /hooks; changing a hook definition changes its trust hash, so an update can require re-trust. Qoder uses SessionStart, SubagentStart, and post-Read coverage, including its documented PowerShell executor on Windows. Kimi uses UserPromptSubmit, while Hermes uses pre_llm_call; both retain their documented Windows execution paths. Devin installs UserPromptSubmit and PostCompaction on macOS/Linux and adds SessionStart only when Claude's equivalent managed hook is not present. GitLab Duo gets a narrowly scoped macOS/Linux user SessionStart entry on its experimental hook surface. GitLab Duo, Devin, and Factory hooks are withheld on Windows because those vendors do not document a deterministic shell/executor contract there. Gemini CLI, Factory Droid, and Augment also add documented post-read/view coverage context but expose no equivalent documented child-start context.

For runtimes without a stable context-producing lifecycle event, durable files carry the contract across fresh sessions and compaction: verify the graph project and index freshness, query structural facts in the parent, then pass the project, qualified symbols, paths, and call-chain evidence in every delegated task. Claude, Codex, Gemini, Kiro, Qwen, Copilot, CodeBuddy, OpenCode, Kilo, Vibe, Qoder, Junie, Factory, and Grok Build receive Scout, Verify, and Auditor graph profiles. Kiro embeds this MCP server with --tool-profile scout for Scout and --tool-profile analysis for Verify/Auditor. Junie registers equivalent named server aliases because its subagent schema filters by server rather than by individual tool. Both process profiles use positive allowlists: Scout exposes seven fast inspection tools, Analysis exposes eleven, and future or mutating tools remain unavailable until explicitly reviewed. If either Junie alias collides with user configuration, the installer preserves it and installs parent-handoff profiles instead. Qoder combines its documented named-server selection with exact tier-specific MCP tool IDs. Factory uses exact registered MCP tool IDs without its additive mcpServers field, which would expose the whole server. Codex, Kilo, Vibe, and other capable formats likewise enumerate the narrowest supported tool set. Rovo, Cursor, Augment, Pochi, and Cline use parent handoff where direct child MCP is unavailable or unsafe; Pochi is limited to readFile, and Cline child agents cannot use MCP.

Cline's file hooks auto-activate when present, and current Cline does not reliably consume their context output, so automatic adapters are withheld and older owned adapters are cleaned up. CodeBuddy's beta, version-gated hooks are not auto-installed. Junie's EAP SessionStart output is documented as ignored, so no context hook is installed. Junie custom agents remain EAP-dependent. Qoder can resolve higher-priority project or plugin agents before user agents with the same name; reload the client after installation or profile changes. Cursor context hooks are withheld: session context injection has a known race, subagentStart is control-only, and read-only subagents cannot safely receive MCP access. Grok Build's passive hook events (SessionStart, SubagentStart, PostToolUse) discard stdout and PreToolUse honors only deny/rewrite decisions, so its context hooks are withheld; Grok also reads Claude and Cursor MCP, skill, and hook files through its compat layer, and the native config.toml entry shadows that copy by name. Rovo has no documented session context-output hook, and Bob documents neither a suitable hook nor a custom-agent surface. Those surfaces are not approximated with invented augmentation. Kimi plugins, Amp plugins, and GitLab experimental global skills remain opt-in.

OpenClaw reinjects the Codebase Knowledge Graph (codebase-memory-mcp) AGENTS section after compaction and places the same guidance in TOOLS.md, the bootstrap files inherited by its subagents. Automatic augmentation covers the active/default workspace. Separate agents.list[].workspace directories require making that workspace active for installation or copying the managed block there.

The installed Claude shim is named cbm-code-discovery-gate for backward compatibility; despite the legacy name, it never gates or blocks.

These are intentionally not counted as automatic installs: Qodo MCP is added through its UI and may be governed by enterprise allowlists; Warp MCP is managed through Warp Drive/UI or per invocation (only the shared skill is automatic); JetBrains AI Assistant / ACP is IDE-managed; GitHub Copilot coding agent, Jules, and CodeRabbit are cloud/repository-managed; Replit exposes a remote/service integration rather than a stable local user-global client; BLACKBOX AI does not document a stable arbitrary user-global MCP/instruction/agent schema; Plandex has no stable global registry safe to mutate; and SWE-agent uses explicit YAML and is no longer a suitable automatic global target.

Every MCP tool can be invoked as a local, one-shot command. CLI tools connect to the shared coordination daemon — starting one if none is already running — to get a crash-safe exact-build admission lease for the command's lifetime; they never register with the background watcher, and a daemon a command had to start itself exits again once the command's connection closes and no other session is using it. index_repository is the only exception internally: it starts a temporary, exact-build supervised worker for the index, then stops that worker before the CLI command exits; the worker holds its own lease until exit.

Commands that mutate graph data use shared OS-backed, per-project locks. This serializes conflicting work from CLI and MCP sessions on the same project while allowing unrelated projects to proceed independently.

When stderr is an interactive terminal, the CLI automatically shows lifecycle and indexing progress. Pass --progress to force the same feedback when stderr is redirected or the command is run non-interactively. Pass --quiet to disable automatic terminal progress and ordinary diagnostics while retaining errors; it cannot be combined with --progress or outer cli --verbose. Routine informational logs are quiet by default; pass outer cli --verbose to include them. Progress and logs use stderr while stdout remains reserved for the command result. Read tools return a compact tree by default; pass a tool's --format json for machine-readable payload JSON, or outer --json for the full MCP envelope.

Large compact-tree tables may start with a response-local <section>_refs directory and an explicit <section>_ref_rule. A cell such as @0+handler.go reconstructs to ref 0's prefix plus handler.go. References are local to that sibling <section> table and expansion is non-recursive: entries inside <section>_refs are always literal prefixes. This is limited to declared path and qualified-name columns and activates only when the exact rendered table is at least 15% and 64 bytes smaller and a conservative model-neutral token-shape proxy also improves by at least 1%. Search and trace likewise render direct and prefix-grouped tree shapes and keep the smaller complete representation, so singleton or scattered answers do not pay directory overhead. Keys are declared once per table but never cryptically abbreviated, and --format json keeps stable literal strings for machine consumers. Both gates are deterministic; exact token counts still depend on the caller's tokenizer.

Lean responses truncate semantically, never by cutting arbitrary bytes from code or identifiers. Ranked graph rows are retained ahead of raw grep rows and diagnostic summaries; omitted rows/sections report totals, has_more, and a strictly advancing continuation offset or cursor. If even the first whole row cannot fit, CBM asks for a higher budget and emits no self-looping cursor. max_output_tokens is model-neutral sizing guidance: CBM enforces a deterministic ceiling of four UTF-8 bytes per requested token, so it is not a tokenizer-exact count. Detail flags such as diagnostics, source_mode, and detail opt into heavier fields. search_code pages ranked rows with result_limit/result_offset (limit remains a compatibility alias), raw rows with raw_limit/raw_offset, and directory summaries with directory_limit/directory_offset. Raw lines default to a UTF-8-safe match-centered preview; each row reports content_start_byte, returned/total byte counts, match byte bounds when known, and a content continuation offset. Pass raw_content_offset to page the original line without moving the raw-row cursor. match_limit and source_max_lines bound per-result details, with exact omission metadata. detect_changes pages changed files, impacted symbols, and module summaries independently; prefer its snapshot-bound *_cursor continuations, which reject changed commits, worktree bytes, graph generation, or semantic arguments instead of silently skipping or duplicating rows.

Every response is standard UTF-8. Identifiers, paths, and raw search previews preserve POSIX byte-string identities: a preserved value containing malformed UTF-8 is emitted reversibly as @bytes:<lowercase hex of every original byte>. A valid preserved value that literally begins with the reserved @bytes: or @utf8: prefix is emitted as @utf8:<original value>, so decoding is unambiguous: strip one @utf8: prefix for literal UTF-8, or hex-decode one @bytes: prefix for original bytes. Ordinary valid UTF-8 is unchanged and pays no output-token overhead. To keep code readable, source bodies replace malformed UTF-8 with U+FFFD; use the pageable raw search preview when byte-exact source inspection is required.

Use cli <tool> --help to see the flags generated from that tool's input schema:

codebase-memory-mcp cli index_repository --repo-path /path/to/repo
codebase-memory-mcp cli list_projects

# Use the "name" returned by list_projects as the project value.
codebase-memory-mcp cli search_graph --project my-project --name-pattern '.*Handler.*' --label Function
codebase-memory-mcp cli trace_path --project my-project --function-name Search --direction both
codebase-memory-mcp cli query_graph --project my-project --query 'MATCH (f:Function) RETURN f.name LIMIT 5'

# Force human-readable progress without contaminating stdout.
codebase-memory-mcp cli --progress index_repository --repo-path /path/to/repo
# Suppress automatic terminal progress and non-error diagnostics.
codebase-memory-mcp cli --quiet list_projects --format json
codebase-memory-mcp cli search_graph --project my-project --label Function --format json
codebase-memory-mcp cli list_projects --format json --detail stats | jq '.projects[].name'

JSON arguments can also be piped on stdin, for tools that take arguments. A tool whose input schema declares none — list_projects — never reads stdin, so it stays responsive when it inherits a pipe the caller never closes (the default for child_process.spawn and similar wrappers). Inline JSON remains accepted for backward compatibility but is deprecated in favor of flags, --args-file, or stdin.

Long indexes and client call deadlines. A synchronous index_repository on a large repository can take longer than an MCP client allows one tool call (some IDE clients give up after a fixed deadline). When a client cancels or disconnects, the daemon cancels an index that nobody else is waiting for, so retrying the same blocking call never finishes. Use the async mode instead:

1. index_repository(repo_path="/abs/path", async: true) starts the index in the daemon (or joins the one already running for that project) and returns immediately with state (queued/running). The job keeps running even if the client cancels, times out or disconnects; only stopping the daemon ends it.
2. index_repository(repo_path="/abs/path", status: true) reports state (queued, running, cancelling, succeeded, failed, cancelled), started_at, finished_at and an error summary. Poll it until the state is succeeded, failed or cancelled. Pass the same repo_path (and name, if the index call used one).

async and status are exclusive; async does not apply to cross-repo-intelligence. Both need the daemon-backed server (the default codebase-memory-mcp entry point). A temporary daemon (started on demand rather than by codebase-memory-mcp daemon start) stops, and cancels its jobs, when its last client disconnects. A connected MCP session keeps it alive, so async from your editor works. A one-shot codebase-memory-mcp cli index_repository --async that is the daemon's only client is refused with a clear error, because the job would die the moment the command exits: run codebase-memory-mcp daemon start first, keep an MCP session open, or call without --async. status works everywhere. When a synchronous call was cut short, the next index_repository or status call for that project carries a notice suggesting the async mode. index_status keeps describing the published graph and its freshness.

manage_adr(mode='set_sections') writes one or more sections by name and splices them into the stored document, so text outside the named sections — including a preamble, code fences and section ordering — is preserved byte-for-byte. Any ## Heading works, not just the conventional PURPOSE / STACK / ARCHITECTURE / PATTERNS / TRADEOFFS / PHILOSOPHY set; names match exactly, including case. Writing the same section twice is a no-op, so a retry after a lost response cannot duplicate content.

manage_adr query modes (get and sections) use the server's cached query store so they can proceed while a same-project reindex is running. If another process publishes a replacement store during reindexing, they can return the pre-publication ADR until idle eviction refreshes that cache. Updates remain serialized through the project mutation guard.

Project, Package, Folder, File, Module, Class, Function, Method, Interface, Enum, Type, Route, Resource

CONTAINS_PACKAGE, CONTAINS_FOLDER, CONTAINS_FILE, DEFINES, DEFINES_METHOD, IMPORTS, CALLS, CALL_REFERENCE, HTTP_CALLS, ASYNC_CALLS, IMPLEMENTS, HANDLES, USAGE, CONFIGURES, REFERENCES_FILE, WRITES, MEMBER_OF, TESTS, USES_TYPE, FILE_CHANGES_WITH

get_code_snippet uses qualified names: <project>.<path_parts>.<name>. Use search_graph to discover them first.

query_graph is a read-only openCypher subset:

• Clauses: MATCH, OPTIONAL MATCH, multiple MATCH, WHERE, WITH (+ WITH … WHERE), RETURN, ORDER BY, SKIP, LIMIT, DISTINCT, UNWIND, UNION / UNION ALL, CASE.
• Patterns: labelled nodes, label alternation (n:A|B), relationship types/direction, variable-length paths [*1..3], inline property maps.
• WHERE: = <> < <= > >=, AND/OR/XOR/NOT, IN, CONTAINS, STARTS WITH, ENDS WITH, IS [NOT] NULL, regex =~, label test n:Label, and EXISTS { (n)-[:TYPE]->() } (single-hop existence — great for dead-code, e.g. WHERE NOT EXISTS { (f)<-[:CALLS]-() }).
• Aggregates: count (+DISTINCT), sum, avg, min, max, collect.
• Functions: labels, type, id, keys, properties; toLower/toUpper/toString/toInteger/toFloat/toBoolean; size, length, trim/ltrim/rtrim, reverse; coalesce, substring, replace, left, right.

Anything outside this subset (write/MERGE/CALL clauses, unsupported functions, list/map literals, comprehensions, path functions, parameters) fails with a clear unsupported … error rather than returning empty results.

Layered: hardcoded patterns (.git, node_modules, etc.) → .gitignore hierarchy → .cbmignore (project-specific, gitignore syntax). Symlinks are always skipped.

See docs/cbmignore.md for the full .cbmignore how-to: syntax, precedence across the ignore layers, and negation semantics.

codebase-memory-mcp config list                          # show all settings
codebase-memory-mcp config set auto_index true           # auto-index on session start
codebase-memory-mcp config set auto_index_limit 50000    # max files for auto-index
codebase-memory-mcp config set auto_watch false          # don't register background git watcher (default: true)
codebase-memory-mcp config set watcher_enabled false     # stop the watcher thread entirely (default: true)
codebase-memory-mcp config set index_max_files 250000    # optional per-index source-file limit
codebase-memory-mcp config set index_max_source_mb 16384 # optional per-index source-size limit
codebase-memory-mcp config reset auto_index              # reset to default

The two index_max_* settings default to off. Exceeding one fails the complete index attempt rather than publishing a partial graph; an existing serving index is preserved. See Index resource limits.

Environment used by daemon-owned components—such as diagnostics, daemon logging, and process-wide indexing resource limits—is captured from the first daemon-backed session that starts the daemon. Later sessions join that process and cannot replace those values. To change them, close all daemon-backed sessions, update the relevant agent configurations consistently, and restart a session. CBM_ALLOWED_ROOT remains session-specific, a conflicting CBM_CACHE_DIR is rejected, and a one-shot CLI command is not exempt from the rule above: it connects to the coordination daemon like any other session, starting one if none is running, so its own environment becomes the captured daemon-owned environment only when its invocation is the one that starts the daemon — joining an already-running daemon, it inherits that daemon's already-captured values instead.

# Store indexes in a custom directory
export CBM_CACHE_DIR=~/my-projects/cbm-data

The JSON config files support a single key, extra_extensions, which maps additional file extensions to supported languages. Useful for framework-specific extensions like .blade.php (Laravel) or .mjs (ES modules). (For other tunables, see Environment Variables and the config subcommand above.)

Need the full config-file reference? See docs/CONFIGURATION.md.

Per-project (in your repo root):

// .codebase-memory.json
{"extra_extensions": {".blade.php": "php", ".mjs": "javascript"}}

Global (applies to all projects):

// ~/.config/codebase-memory-mcp/config.json  (or $XDG_CONFIG_HOME/...)
{"extra_extensions": {".twig": "html", ".phtml": "php"}}

Each entry maps an extension (which must start with .) to a language name. Language names are matched case-insensitively. Accepted values (aliases in parentheses) are:

bash (sh), c, c++ (cpp), c# (csharp), clojure, cmake, cobol, common lisp (commonlisp, lisp), css, cuda, dart, dockerfile, elixir, elm, emacs lisp (emacslisp), erlang, f# (fsharp), form, fortran, glsl, go, graphql, groovy, haskell, hcl (terraform), html, ini, java, javascript, json, julia, kotlin, lean, lua, magma, makefile, markdown, matlab, meson, nix, objective-c (objc), ocaml, perl, php, protobuf, python, r, ruby, rust, scala, scss, sql, svelte, swift, toml, tsx, typescript, verilog, vimscript, vue, wolfram, xml, yaml, zig.

Project config overrides global for conflicting extensions. An entry whose language name is unknown, or whose extension does not start with ., is skipped and a warning is logged to stderr (shown at the default info log level). Missing config files are ignored.

SQLite databases stored at ~/.cache/codebase-memory-mcp/. Persists across restarts (WAL mode, ACID-safe). To reset: rm -rf ~/.cache/codebase-memory-mcp/.

Semantic type resolution beyond tree-sitter.

Tree-sitter alone gives a syntactic AST. That handles naming, structure, and call sites well, but it can't tell you that user.profile.display_name() resolves to Profile.display_name declared three modules away — tree-sitter doesn't track imports, generics, inheritance, or stdlib types.

codebase-memory-mcp ships a lightweight C implementation of language type-resolution algorithms, structurally inspired by and compatible with major language servers (tsserver / typescript-go, pyright, gopls, Roslyn, Eclipse JDT, rust-analyzer), embedded directly into the native executable. No language server process, no per-project setup, no API key. We call this layer Hybrid LSP: it runs alongside tree-sitter on every parse and refines invocation resolution (CALLS / RESOLVED_CALLS) and callable-value resolution (CALL_REFERENCE, with ambiguous values retained as USAGE) using type information, so the resulting graph mirrors what an IDE "Go to Definition" would resolve.

Languages with full Hybrid LSP:

Two-layer architecture:

1. Tree-sitter pass — fast, syntactic, runs for every one of the 162 languages. Extracts definitions, calls, imports.
2. Hybrid LSP pass — type-aware, runs above the tree-sitter pass per-language. Refines call edges using the import graph plus a per-file or pre-built cross-file definition registry. Languages without a Hybrid LSP pass yet fall back to textual resolution, so you always get some answer.

The result is a knowledge graph accurate enough to drive trace_path across packages, inheritance hierarchies, and stdlib calls — without paying for a language server process per project.

162 languages, all parsed via vendored tree-sitter grammars compiled into the binary. Benchmarked against 64 real open-source repositories (78 to 49K nodes):

Also supported (not yet benchmarked): Ada, Agda, Apex, Assembly (NASM), Astro, AWK, Beancount, BibTeX, Bicep, Bitbake, Blade, Cairo, Cap'n Proto, Clojure, CMake, COBOL, Common Lisp, Crystal, CSV, CUDA, D, Devicetree, Diff, .env, Elm, Emacs Lisp, F#, Fennel, Fish, FORM, Fortran, FunC, GDScript, .gitattributes, .gitignore, Gleam, GLSL, GN, Go module, Go template, GraphQL, Hare, HLSL, Hyprlang, INI, ISPC, Janet, Jinja2, JSDoc, JSON, JSON5, Jsonnet, Julia, Just, Kconfig, KDL, Lean 4, Linker Script, Liquid, LLVM IR, Luau, Magma, Makefile, Markdown, MATLAB, Mermaid, Meson, Move, Nickel, Nim, Nix, Odin, Pascal, Pkl, PO (gettext), Pony, PowerShell, Prisma, .properties, Protobuf, Puppet, PureScript, Racket, Regex, requirements.txt, ReScript, RON, reStructuredText, Scheme, Slang, Smali, Smithy, Solidity, SOQL, SOSL, Squirrel, SSH config, Starlark, Svelte, Sway, SystemVerilog, TableGen, Tcl, Teal, Templ, Thrift, TLA+, Typst, Verilog, VHDL, Vim script, Vue, WGSL, WIT, Wolfram, XML, Zsh.

src/
  main.c              Entry point (MCP stdio server + CLI + install/update/config)
  daemon/             Per-account session coordination, IPC, lifecycle, shared jobs/watchers
  mcp/                MCP server (17 tools, JSON-RPC 2.0, session detection, auto-index)
  cli/                Install/uninstall/update/config (45 client surfaces, hooks, instructions)
  store/              SQLite graph storage (nodes, edges, traversal, search, Louvain)
  pipeline/           Multi-pass indexing (structure → definitions → calls → HTTP links → config → tests)
  cypher/             Cypher query lexer, parser, planner, executor
  discover/           File discovery (.gitignore, .cbmignore, symlink handling)
  watcher/            Background auto-sync (git polling, adaptive intervals)
  traces/             Runtime trace ingestion
  ui/                 Local HTTP server + verified external 3D-UI asset pack
  foundation/         Platform abstractions (threads, filesystem, logging, memory)
internal/cbm/         Vendored tree-sitter grammars (162 languages) + AST extraction engine

Every release is verified through a multi-layer pipeline before publication:

• VirusTotal — all 24 executable candidates (unstripped, debug-stripped and stripped) across the eight release products are scanned before smoke/soak (clean is preferred and only the single-Microsoft !ml tolerance documented in SECURITY.md may pass; the number of engines that returned a decisive result is recorded as evidence but is VirusTotal fleet availability, not a pass condition). The selected executable is packaged without changing its SHA-256, release notes link the verdict for the exact bytes shipped, and the full per-candidate evidence is published alongside the release as TSVs for anyone auditing the selection. Every distinct object extracted from the shipped containers — install.sh, install.ps1, LICENSE, THIRD_PARTY_NOTICES.md, the MCPB manifest.json and the unpacked UI assets — is then scanned as well, so the full published surface is covered, not just the executables.
• SLSA Level 3 — cryptographic build provenance generated by the trusted GitHub Actions build workflow; verify with gh attestation verify <file> --repo DeusData/codebase-memory-mcp --signer-workflow DeusData/codebase-memory-mcp/.github/workflows/_build.yml
• Sigstore cosign — keyless signatures on all artifacts; bundles included in every release
• SHA-256 checksums — checksums.txt published with every release; verified by both install scripts before extraction
• CodeQL SAST — blocks release pipeline if any open alerts remain
• No language-runtime dependency chain — libraries are vendored at compile time; the small release-owned runtime assets are checksum-verified and content-addressed

Scan links for every release are also included in the GitHub Release notes automatically.

MIT