codebase-memory-mcp

A workflow for using codebase-memory-mcp, a tool that builds a searchable graph of a Git repository. It resolves the correct checkout, checks that the graph works, and manages linked worktrees and duplicate indexes.

In plain words
What is it for?
Use it to initialize, refresh, verify, search, or troubleshoot a repository graph, trace code paths, retrieve snippets, and audit duplicate worktree caches.
Why use it?
It helps agents find and trace code reliably without indexing the same repository worktree multiple times or relying on an unverified graph.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/vincentkoc/dotskills/codebase-memory-mcp
Any agent
npx skills add vincentkoc/dotskills --skill codebase-memory-mcp
Clone the repo
git clone --depth 1 https://github.com/vincentkoc/dotskills

Made for: Claude Code, Codex.

Per session 62 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,981 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00062 $0.01981
Opus 5 $0.00031 $0.00991
Sonnet 5 $0.00012 $0.00396
Haiku 4.5 $0.00006 $0.00198

Measured 2d ago against content hash 7b3fd7be7dd6, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

codebase-memory-mcp scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

The scan reads SKILL.md. This mod also ships 2 executable files (scripts/codebase_memory_cache.py, scripts/codebase-memory-graph.sh), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/codebase-memory-mcp/SKILL.md · 90 lines

How it starts

The opening of the file, as written. The whole thing — 90 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Codebase Memory MCP

Purpose

Bring up codebase-memory-mcp for the owning Git checkout and prove the graph is usable before relying on it for code discovery. Keep linked worktrees on the owner's graph instead of creating one graph per branch.

When to use

  • Initialize, re-index, refresh, or troubleshoot a repository graph.
  • Use search_graph, trace_path, get_code_snippet, or query_graph.
  • Verify that a repository is indexed before graph-backed exploration.
  • Audit or prune duplicate linked-worktree indexes without deleting cache files directly.

Workflow

  1. Verify the repository and binary.
    • git rev-parse --show-toplevel
    • git status -sb
    • command -v codebase-memory-mcp
    • codebase-memory-mcp --version
  2. Resolve the canonical owning checkout.
    • scripts/codebase-memory-graph.sh canonical --repo "$(git rev-parse --show-toplevel)"
    • Linked worktrees resolve through their absolute Git common directory to the one checkout that owns it.
    • Separate clones remain separate projects.
    • Independent roots under ~/.codex/worktrees, ~/GIT/_Worktrees, any .worktrees component, /tmp, or /private/tmp are never indexed. Linked worktrees under those paths may only rewrite to one existing nonreserved owner.
    • Missing, invalid, bare, ambiguous, ownerless, reserved-owner, or NUL-containing repositories fail closed.
  3. Prefer exposed MCP graph tools for discovery.
    • Installer or client configuration must separately disable the MCP index_repository tool because it cannot enforce the canonical indexing boundary. For Codex installs, render the private disabled_tools configuration accordingly.
    • When a graph is missing, run scripts/codebase-memory-graph.sh index --repo "$(git rev-parse --show-toplevel)" --mode full.
    • Use search_graph, trace_path, and get_code_snippet before broad text scans.
  4. Use the helper for CLI indexing.
    • scripts/codebase-memory-graph.sh init --repo "$(git rev-parse --show-toplevel)" --mode full
    • The helper always sends the canonical owning checkout to index_repository.
    • Use --mode fast for a smoke index.
    • Installer integrations render scripts/codebase-memory-gateway.py.tmpl with an approved pinned backend path. Replace @@PYTHON_PATH_SHEBANG@@ with the raw absolute interpreter path and replace @@PYTHON_PATH_JSON@@, @@BACKEND_PATH_JSON@@, and @@RESOLVER_PATH_JSON@@ with JSON string literals containing the exact absolute interpreter, backend, and codebase_memory_cache.py paths. The rendered gateway has no upgrade logic and uses execve for pass-through.
    • The gateway guards raw CLI calls only. Zero-argument MCP stdio startup intentionally passes through to the approved backend, so the gateway is not an MCP tool-filtering proxy and does not replace the separate disabled_tools control.
  5. Verify the graph.
    • codebase-memory-mcp cli list_projects
    • scripts/codebase-memory-graph.sh schema --repo "$(git rev-parse --show-toplevel)"
    • Run one focused graph query before declaring success.
  6. Treat the UI as unavailable.
    • start-ui and keepalive fail closed before configuration or process mutation.
    • UI startup remains disabled until upstream /api/index canonicalization can enforce the same boundary. status may report an already-running grandfathered listener.
  7. Audit cache cleanup before applying it.
    • Freeze a host-specific manifest: scripts/codebase-memory-graph.sh cache-audit --manifest /secure/path/cbm-cache.json
    • Built-in reserved roots are ~/.codex/worktrees, ~/GIT/_Worktrees, /tmp, /private/tmp, and any exact .worktrees path component. The audit records normalized lexical and resolved boundary evidence; .worktrees component matching is case-insensitive on Darwin.
    • Missing roots under a reserved boundary are reserved_missing_root candidates. Live valid graphs whose physical canonical root and Git common directory remain reserved are reserved_live_root candidates, including shallow, promisor, and partial clones. Dangling symlinks, bare repositories, non-Git roots, and invalid reserved mappings block as live_root_unmapped.
    • Other missing roots are protected unless the audit names an explicit narrow --ephemeral-prefix.
    • Reserved aliases or linked worktrees that resolve to a nonreserved owner retain duplicate cleanup rules. They require a mapped, registered, final-protected healthy full canonical graph; candidate graphs, reserved graphs, shallow/promisor/partial owners, missing alternate graphs, and ambiguous owners cannot preserve them.
    • When legacy home symlinks name the same physical checkout, preserve the exact canonical graph and treat only the symlink-named graph as a duplicate. Preserve a sole symlink-named graph.
    • Review the manifest, then dry-run it: scripts/codebase-memory-graph.sh cache-prune --manifest /secure/path/cbm-cache.json
    • Manifests with host blockers fail closed by default. After reviewing every relationship, explicitly add --allow-blocked-manifest to preflight and prune only independent candidates while preserving all protected and blocked projects.
    • To retain an exact manifest candidate at runtime, repeat --protect-candidate NAME on cache-prune. Each named candidate must still exist in the unchanged snapshot and pass its root, prefix, classification, canonical mapping, and clone-health rules. It remains in the expected snapshot but is excluded from cache inventory, holder sweeps, DB integrity preflight, final holder probes, and deletion.
    • Runtime candidate protection does not bypass host blockers; add --allow-blocked-manifest independently when blockers were reviewed. Unknown, duplicate, or non-candidate names fail before preflight.
    • Apply only the unchanged manifest: scripts/codebase-memory-graph.sh cache-prune --manifest /secure/path/cbm-cache.json --apply
    • Dry-run and apply freeze regular DB/WAL/SHM fingerprints for every eligible candidate in manifest order, rejecting a missing or size-mismatched DB, nonregular files, and nonzero WAL. Existing absolute paths are swept with /usr/sbin/lsof -nP -F0pfn -f -- in deterministic, NUL-parsed batches before any SQLite open and again after every exact mode=ro&immutable=1 quick_check; global fingerprint equality is required between phases. An absent or zero-byte WAL and a stable regular SHM are allowed. Use prune-only --lsof-timeout-seconds SECONDS to override the 300-second holder timeout within the guarded 30-900 range.
    • Batch holder sweeps are point-in-time checks: they do not prevent a legacy index or server surface from reopening a graph after the check. Unprivileged lsof may not see root-owned or other-user holders, so pruning trusts the point-in-time visibility available to the cache owner; running as root provides stronger holder visibility. Apply processes eligible candidates in manifest order, at most eight per deletion batch and within the same 128 KiB path budget. Each batch revalidates the snapshot and live fingerprints, performs one holder sweep, proves post-sweep fingerprint equality, then revalidates every remaining candidate relationship immediately before the first delete child. The batch launches only codebase-memory-mcp cli delete_project children and verifies registrations plus DB/WAL/SHM absence before continuing; spawn errors, timeouts, nonzero exits, retained registrations, residue, drift, or ambiguous state stop future batches and report launched, verified-deleted, failed, and ambiguous names and bytes.
    • Before applying any manifest containing reserved_live_root, deploy the reserved-root indexing prevention and quiesce every legacy index/server surface outside this helper. The helper reports this operational precondition but does not probe or kill processes.
    • Deletion uses codebase-memory-mcp cli delete_project only. Never remove project databases directly.
  8. Report exact proof.
    • Indexed project name.
    • Node and edge counts when available.
    • Any grandfathered UI listener reported by status.
    • For cleanup: manifest path and digest, manifest/eligible/preflighted candidate totals, runtime protected names/reasons/bytes, required operational preconditions, before/after project and byte totals, deleted project names, and any stop condition.
    • Missing binaries, unavailable MCP tools, or incomplete proof.

Read the full file on GitHub · 90 lines

Files

What ships with it

5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 90 lines · 62 tokens per session scan A 7b3fd7be7dd6

Subscribe to this mod's changes

codebase-memory-mcp is a skill published in the GitHub repository vincentkoc/dotskills (102 stars, last pushed 11d ago), licensed MIT. It adds 62 tokens to every session and 1,981 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

panel-app-creator

Create or update the Panel App UI part of a NextClaw lightweight app. Use after nextclaw-app-creator selects Panel-only or Panel + Service, or when the user explicitly asks for a right-side Panel App UI, folder-based static panel, Service Actions UI, Agent-powered Panel App, React/Vite/Tailwind Panel App, or asks what…

Peiiii/nextclaw · 93 tokens

service-app-creator

Create or update the Service App backend action part of a NextClaw lightweight app. Use after nextclaw-app-creator selects Service-only or Panel + Service, or when the user explicitly asks for workspace service-apps, MCP-compatible backend helpers, file access, external API calls, local commands, or privileged actions.

Peiiii/nextclaw · 69 tokens

nextclaw-dead-code-governance

当用户要求扫描、识别、清理、常态化治理 NextClaw 仓库里的死代码、无用代码、unused files、unused exports、unused dependencies,尤其是核心包越来越臃肿、想删旧实现或做定期减债时使用。.

Peiiii/nextclaw · 71 tokens

nextclaw-http-agent-runtime-integration

Use when integrating or debugging a Hermes-style external HTTP agent runtime in NextClaw, including runtime config, session selection, readiness, streaming, and adapter behavior.

Peiiii/nextclaw · 41 tokens

nextclaw-product-visual-assets

当 NextClaw 产品更新后需要生成、替换、挑毛病或检查官网、GitHub README、用户文档或社交传播中的真实截图、AI 宣传视觉、整页 HTML 宣传预览、社区二维码等对外视觉资产时使用;也用于“更新截图”“重新截一批图”“做宣传页”“生成 campaign 页面”“视觉审稿”“五星挑刺法”或发布前检查视觉资产。普通站点布局开发或只写文章不触发。.

Peiiii/nextclaw · 115 tokens

agent-browser

Use Agent Browser for real browser navigation, dynamic pages, and web interaction; also use it as a distinct browser-based path when websearch is unavailable or fails.

Peiiii/nextclaw · 35 tokens