Skip to main content
Glama

Two layers, one tool:

  • The notebook (coldstart kb) — durable, agent-written notes about how this codebase actually works: what a file is for, how a flow spans files, which invariants hold. Captured after real tasks, recalled when a later task matches, and kept honest by the index — every note is anchored to real files, and a note whose evidence drifted is flagged, not served as truth.

  • Navigation (coldstart find / coldstart gs) — a fast static index over file paths, symbol names, exports, and the import/call graph. It answers "which files are relevant to this task?" in milliseconds, with checkable evidence instead of a similarity score.

No embeddings, no model to run, no service to babysit. Agents are already good at reading and reasoning about code; what they waste tokens on is finding the right file and re-deriving what the last session already figured out. coldstart does those two parts and gets out of the way.


Install

Requires Node.js 18+.

npm install -g @cstart/coldstart
cd your-project
coldstart init   # coldstart.md + client wiring + notebook + background index warm-up

A single coldstart init does everything — navigation and the notebook. It asks two things — the experience (cli, recommended, or mcp) and the client — then writes the agent-facing guidance into the client's own rules file (as an imported coldstart.md for Claude Code; inlined directly for Cursor and Codex, which don't resolve @file references), wires the client, and sets up the notebook (skeleton, git wiring, and — for Claude Code, Codex, and Cursor — the capture/recall hooks). Pass --experience / --client to skip the prompts. The client is never auto-detected; you always pick it.

  • Claude Code → writes coldstart.md and ensures CLAUDE.md imports it via @coldstart.md, and registers both the find/gs search hooks (a PostToolUse nudge + a PreToolUse find-dedup guard) and the notebook recall/capture hooks (UserPromptSubmit + Stop/SubagentStop, plus a PreCompact hook that lets recall re-show notes after a compaction) in .claude/settings.json — merged into any existing settings, never overwriting them. The mcp experience also writes .mcp.json.

  • Codex → embeds the full coldstart guidance inline in a marked block in AGENTS.md (Codex has no @file include, so there's no separate coldstart.md), refreshed in place on re-run, and registers Codex-specific navigation plus notebook hooks in .codex/hooks.json (including PostCompact, so recall can re-show notes after a compaction). The capture hook understands Codex rollout and subagent transcripts. The mcp experience also writes [mcp_servers.coldstart] into .codex/config.toml.

  • Cursor → writes .cursor/rules/coldstart.mdc — an always-applied rule that carries the full coldstart guidance inline (Cursor doesn't reliably resolve @file references in rules), rewritten on every init — and registers Cursor-specific navigation plus notebook hooks in .cursor/hooks.json (a preToolUse find-dedup guard, a postToolUse nudge, beforeSubmitPrompt recall, a preCompact recall reset, and stop/subagentStop capture — merged into any existing hooks). The capture hook parses Cursor's own conversation transcript. The mcp experience also writes .cursor/mcp.json.

  • Other → writes coldstart.md only, and prints the wiring directions (plus the MCP server entry for the mcp experience).

On every client, the PostToolUse hook also delivers the "Edited together" signal without waiting to be asked. Once an agent has edited two different files, it names the files that git history says keep changing alongside them — the sibling implementation, the test in another language, the doc that drifts. Agents don't reliably run gs before editing, so the signal only reached them if they went looking; this puts it in front of them at the edit. It's advisory: the message says outright that the relation is a habit rather than a code dependency, and asks the agent to check rather than to change anything. Each file is named at most once per task, and the list resets whenever you send a new message.

init then warms the index in the background, so your first lookup is instant. Re-running init is safe — it never duplicates entries.

Upgrading

npm install -g @cstart/coldstart@latest
coldstart init   # re-run in each project to refresh coldstart.md

A version stamp in the keeper's lockfile makes the old background keeper shut down on the next lookup; a fresh one spawns from the new binary. No manual restart needed.

NOTE

Migrating from coldstart-mcp: the package was renamed coldstart-mcp → @cstart/coldstart at 2.0.0 (the CLI is now the primary surface). coldstart-mcp is deprecated but still installs; switch with npm uninstall -g coldstart-mcp && npm install -g @cstart/coldstart && coldstart init. The coldstart-mcp binary name is kept as an alias, so existing MCP configs keep working.

Removing coldstart

init writes per-repo wiring that a global npm uninstall can't reach (npm fires no reliable uninstall hook, and it has no record of which repos you init'd). So — like husky — coldstart ships an explicit reverse:

coldstart unwire          # strip coldstart's wiring from this repo (notebook kept)
coldstart unwire --purge  # also delete .coldstart/notebook/ and its git plumbing

unwire removes only coldstart-owned markers from the files init touched — hook entries, the @coldstart.md import, the AGENTS.md block, the MCP server entry, and files coldstart fully owns (coldstart.md, .cursor/rules/coldstart.mdc) — never your own content in shared files. It sweeps all four clients, is idempotent (a second run reports everything already gone), and keeps the notebook by default since it's committed, shared data. Run it in each project first, then npm uninstall -g @cstart/coldstart to remove the package.


Related MCP server: knos

The notebook

A repo-local knowledge base written and read by agents, in .coldstart/notebook/:

coldstart kb search tile save lifecycle   # plain task words, symbols, or file names
coldstart kb lookup src/models.py Tile    # everything known at one exact address
coldstart kb write spec.json              # the write gate (two-phase dedup)
coldstart kb commit                       # publish notes to git, nothing else rides along
coldstart kb view                         # open a single-file HTML browser of the notebook
coldstart kb repair                       # worklist of notes that are written but unfindable
coldstart kb status / lint / render / init / migrate

What a note is. Three shapes: a file note (what a file is for — a single summary, or per-symbol facets for hub files), a flow note (a cross-file story: ordered steps, invariants), and a lesson (a trap, rule, bug-cause, rationale, or confirmed absence). Every note carries anchors — concrete file paths and symbols its claims rest on.

Where notes reach the agent. Three surfaces, no new habits required:

  1. Summary: lines on find results — a past agent's verified overview of a file, right where the file ranks. [fresh] means the file is byte-identical to when the summary was verified — the agent can rely on it without re-reading the file.

  2. Recall at prompt time (optional hook) — the incoming prompt is ranked against every note's title and text; when one note clearly leads, the top three are surfaced as a compact title + gist + path block for the agent to pick from, framed as reference data. No clear leader → nothing injected.

  3. kb search / kb lookup — a search engine over the notebook for mid-task vocabulary changes, and an exact-address lookup (path [symbol]) before editing a file.

Why it can be trusted. This is the part that took the design work:

  • Freshness is mechanical, not hoped-for. Every anchor is stamped with a content hash at write time; the index re-checks stamps as the code changes. A drifted note renders [evidence changed: <path>] and the guidance says re-verify — stale knowledge degrades into a labeled hypothesis instead of a confident lie.

  • The log is the truth. Notes live in an append-only .raw event log (commit it — merges are unions, so parallel branches of notes reconcile without conflicts). The Markdown notes are derived, regenerated mechanically, and gitignored.

  • Writes go through a gate. A new note's concept is first searched against existing notes — the agent must explicitly merge into a match (--into <id>) or declare it new (--new). Duplicates are gated at write time, not cleaned up later.

  • Concurrent sessions are safe. Multiple agents can write at once: per-note append-only logs, exclusive creation for new note ids (a same-moment duplicate becomes two visible notes, never a silent merge), lossless merging for shared file notes, and atomic renders (a reader never sees a half-written note).

  • Corrections happen in-session. If an agent finds a note wrong while the evidence is in its context, the guidance tells it to fix or retract the note right then — no better-placed future agent exists.

Setup: the notebook comes with coldstart init — no separate step. It creates the notebook skeleton, sets union-merge for the logs, and (on Claude Code, Codex and Cursor) wires the two hooks — capture at session end, recall at prompt time. (coldstart kb init still exists as an alias if you want to (re-)wire just the notebook.) Other hosts can drive the notebook without the hooks: via the full kb CLI, or — for no-shell clients — the kb_search / kb_lookup / kb_write / kb_status / kb_repair MCP tools.

Language-agnostic. The notebook's freshness machinery is content-hash based, so it works on any codebase — including languages the navigation index doesn't parse. Where the index does parse, notes additionally get symbol-level freshness.

NOTE

The notebook is young. What's verified today: notes written by agents in real sessions checked out accurate against the code; the stale-note loop closes end-to-end (flag → re-read → correction); capture, recall, and concurrent writes hold up under stress. The bet — stated as a bet — is that a corpus like this compounds over a repo's lifetime: the second time any question comes up, the answer is one Read away instead of a re-derivation.


Navigation: the two operations

What it answers

Replaces

find <terms>

"Which files are about this?" — ranks files by how many of your query terms they cover (filenames, path segments, exported symbols, plus a repo-wide name-reference pass).

a flurry of grep/glob while orienting

gs <file>

"What is this file?" — top-level symbols with line ranges, who imports it, who calls each symbol, and name-related neighbors.

reading a whole file just to learn its shape and usage

The intended flow: find a concept → pick the best path → gs that file for its shape and who uses it → Read only for the implementation inside a method body. Notebook summaries ride along on find results, so often the orientation step answers itself.

flowchart LR
    A["coldstart find<br/>which files?"] --> B["coldstart gs<br/>what is it? who uses it?"] --> C["Read<br/>just the method body"]
    class A,B cold
    class C warm
    classDef cold stroke:#16708f,stroke-width:2px
    classDef warm stroke:#c26714,stroke-width:2px

find — locate the files for a concept

coldstart find auth session cookie
TIP

Pass every salient identifier from your task — the symbol, the domain noun, the rare token you half-remember — not one distilled keyword. find ranks files by how many of your terms each one covers and shows, per file, which terms it defines vs. imports and a preview of the lines where they cluster. Often that's enough to answer without opening anything.

Speed-wise, find competes with raw grep: its repo-wide reference pass runs on ripgrep — yours from PATH, the bundled copy, or an editor's (COLDSTART_RG overrides) — with git grep/grep fallbacks, and the ranked page comes from the pre-built index, not a scan.

Flags: --path GLOB (scope; comma-combine, ! excludes) · --tests (include test files) · --via (show name-reference relations) · --json

gs — drill into one file

coldstart gs src/auth/service.ts

Returns the file's symbols (with line ranges), its 1-hop internal imports, who imports it, and per-symbol cross-file callers — in one call. This is the answer to "who uses this file / who calls this symbol"; it is not a grep.

Flags: --symbol a,b (deliver named method bodies inline) · --match TERM (filter a god-file to one area; a|b = OR, /regex/ = regex) · --view symbols|imports|importers|callers · --json

graph — see the whole repo at once

coldstart graph

For humans, not agents. Writes one self-contained HTML file and opens it: every indexed file is a point on a sphere, positioned by directory, and clicking one opens a 2D view of everything it is connected to — with each relation named (imports, calls save(), load(), edited together · 13 of 42 commits, same note · <title>). Click a neighbour to open its connections; click one of those and the view slides along, so you always see two levels rather than an ever-growing hairball.

No dependencies. The page is HTML, CSS, and a canvas script with your repo's data baked in — no server, no build step, and nothing leaves your machine. Mail it, drop it in a gist, open it on a plane. Try it on coldstart's own codebase.

Flags: --out PATH (default .coldstart/graph.html, gitignored for you) · --no-open · --json

Batch independent lookups in one shell call

coldstart find auth; coldstart find 'session cookie'; coldstart gs src/auth/service.ts

Two ways to call it, identical output

coldstart ships as one binary with two front doors:

  • CLI (primary) — coldstart find … / coldstart gs … / coldstart kb …. For any shell-capable agent (Claude Code, Cursor, terminal use). This is the fast path.

  • MCP (for no-shell clients) — the find and gs tools, plus the notebook as kb_search / kb_lookup / kb_write / kb_status / kb_repair, all byte-identical to the CLI. For clients like Claude Desktop that have no shell. (kb commit stays CLI/human-only — publishing notes to git is never an agent action.)

Same engine, same index, same results. Pick whichever your agent can reach.

It works best with Claude Code, Codex, and Cursor: all three get platform-specific find/gs hooks and notebook recall/capture hooks from coldstart init. Any other client gets coldstart.md plus printed wiring directions.


Bring your own semantics

coldstart has no embeddings, no generated summaries, no semantic layer computed at index time — on purpose. The semantic layer is the agent. Every consumer is already a frontier model; pre-computing meaning at index time only duplicates that, worse and stale. So the index keeps what's cheap to keep exact — paths, symbols, exports, the import/call graph — and returns why each file ranked.

The notebook is the same philosophy applied to memory: coldstart still computes no meaning of its own. It stores, anchors, and freshness-checks the meaning agents author — written at task time, by the reasoner that had the full context, about the question that actually mattered. The full argument is in PHILOSOPHY.md.


How the index stays fresh

coldstart is one keeper, thin readers:

flowchart TD
    K["keeper — coldstart --daemon<br/>watches repo, patches/rebuilds, saves cache<br/>serves nothing"] -->|debounced save| C[("on-disk cache")]
    C --> F["coldstart find<br/>reads cache, prints"]
    C --> G["coldstart gs<br/>reads cache, prints"]
    C --> M["MCP server<br/>reads cache, stdio"]
    class K cold
    classDef cold stroke:#16708f,stroke-width:2px
  • A single keeper process per repo watches the filesystem and keeps the on-disk cache current. It does not answer queries.

  • The CLI readers (find/gs) and the MCP server are stateless readers over that cache. The first reader for a repo lazily spawns the keeper, so even uncommitted edits stay live.

  • Readers never build the index. On a cache miss they wait for the keeper's build (progress to stderr) instead of silently kicking off a multi-minute build inline — or three of them concurrently.

  • No HTTP, no ports, no bridge. The keeper logs to ~/.coldstart/daemon/<root>.log and exits when its lockfile is removed.

There is no cache TTL. The index is never discarded for being old — it's kept correct instead:

  • While the keeper runs: edits are debounced (400 ms), then patched incrementally (~2–5 ms/file, up to 30 files or 20% of the repo, whichever is larger) or trigger a background full rebuild above that (served from the last good index until the swap). The cache re-saves ~5 s after edits settle, in atomic generations — a reader can never load a half-written mix of old and new.

  • When the keeper starts: it reconciles — stat-checks every indexed file against its stored fingerprint (~150 ms even at 16k files) plus a git diff against the indexed HEAD — and patches exactly what changed while nothing was watching. A branch switch that used to force a 96-second rebuild on a 16k-file repo is now a ~3-second patch.

  • As a backstop: every patch is lint-checked against index invariants (a violation triggers an automatic rebuild and lands in a repair log that status shows), and a rotating fingerprint audit after each save catches watcher-missed events.

The keeper also stamps the notebook's anchor freshness (a small sidecar, derived single-flight) — the notebook never loads the code index to answer a query.

Lifecycle commands

coldstart status              # keepers on this machine: alive? fresh? last patch/rebuild/save? repairs?
coldstart restart             # kill the current repo's keeper (respawns on next lookup)
coldstart restart --root DIR  # kill a specific repo's keeper from anywhere
coldstart restart --all       # kill every keeper
coldstart index               # build + save the cache once, up front (single-writer prep)

restart is the right move whenever anything feels stale — a fresh keeper reconciles on start, so it comes back correct, not just alive. status answers "is my index fresh, and why?": liveness, cache age, the keeper's last reconcile/patch/rebuild/save stamps, and the tail of the repair log — no network probe.


Supported languages

Navigation index: TypeScript, JavaScript, JSX/TSX, Vue, Svelte, Astro, AngularJS 1.x, Java, Kotlin, Ruby (Rails-aware: has_many/belongs_to associations, routes.rb resources, controller↔view edges), Python (Django convention edges), Go, Rust, C#, PHP (Laravel convention edges), C++, Groovy (incl. Gradle DSL), GraphQL, YAML, TOML, XML, and .env files.

Not indexed: Swift, Dart — no extension mapping; these files are not walked or parsed.

The notebook works regardless — its freshness stamps are content-hash based, so notes on a Swift repo are as trustworthy as notes on a TypeScript one (they just lack symbol-level freshness detail).


When not to reach for it

  • A literal string / phrase / regex inside file bodies → Grep.

  • Reading an implementation → Read, after gs gives you the shape.

  • find says "no indexed file contains any of […]" → those identifiers aren't in the repo. Don't grep spelling variants.


Development

npm install
npm run build
npm test

# run a query from your build:
node dist/index.js find auth --root .

# run the MCP server in a single process (no background keeper) for debugging:
node dist/index.js --root . --no-daemon

See PHILOSOPHY.md for why coldstart computes no semantics of its own, ARCHITECTURE.md for the index pipeline, process model, and notebook internals, and TROUBLESHOOTING.md for recovery procedures.


Limitations

  1. It's a routing layer plus an agent-written notebook — no semantic analysis or generated code summaries. This is deliberate: the consuming agent is the semantic layer (see PHILOSOPHY.md).

  2. gs callers are one-hop and file-scoped. Member-expression calls (this.method(), api.method()) aren't cross-file resolved; named function/constant calls are. Chase further hops by calling gs on the caller files.

  3. Dynamic/computed imports (import(variable)) and runtime-DSL references (polymorphic associations, gem/reflection-backed models) stay unresolved.

  4. Hidden directories and files over 1 MB are skipped by the index.

  5. The keeper is per-repo and per-machine — no sharing across projects or hosts. The notebook does travel: its .raw logs are committed and union-merge across branches and machines.

  6. Notebook quality is bounded by what writing agents actually read — notes are accurate about what they state, but a note is not a proof of completeness.

Measured

Two repositories, each run more than once, with the CLI. Each question is asked in two arms that differ only in whether coldstart is installed: the baseline is plain Claude Code using the search tools it ships with. Sonnet 5 on both sides, one fresh session per question.

Repository

Language

Tokens vs. baseline

Recall

Questions

Arches

Python / Django

−64%

+2 pts

27

JMRI

Java

−31%

parity

25

The questions come from closed issue reports, which are written before anyone knows where the fault lives, so they describe a symptom rather than naming a file. The correct answer is the set of source files in the commit that closed the issue, so ground truth comes from git history rather than from anyone's judgement.

Those two numbers are not averaged. Per-turn cost differs by language, and the saving comes from removing turns, so a question an agent can settle in two or three turns has little to give back. An earlier benchmark on Kafka, Django and Mastodon was discarded: the model has read all three, and that contaminated both arms. The full method, including what was thrown out and why, is at coldstartmcp.dev/benchmark. The harness itself is a separate project, coldbench, and works on any repository with git history, including private ones. Both question sets are published with their gold file lists and the issue number behind each question, at coldbench/examples.

Writing

Longer pieces on the problems behind this tool — what agent sessions actually cost, and what happened to the design when the measurements disagreed with the plan.

License

MIT — see LICENSE.


coldstart MCP server

Available Tools

7 tools
findA

Locate the files relevant to a task. Pass query = EVERY salient identifier from the task (symbol names, domain nouns, the rare token you half-remember) — not one distilled keyword. Recall is bounded by the terms you give: a one-token query cannot out-rank lookalikes, so over-supply rather than under-supply. Reach for find BEFORE Read/Grep/Glob.

HOW IT WORKS: find greps every term across the repo body AND matches declared names (filenames, path segments, exported symbols), then ranks files by DISTINCT-TERM COVERAGE — the file that covers MORE of your query rises above its lookalikes. This catches body-level matches (nested defs, dynamic refs, string literals) that a declared-name index misses.

OUTPUT: a ranked page. Top files get an inline preview — their indexed symbols (with line ranges) plus the body lines where your rare terms CLUSTER (def/class/assignment lines first), so you often answer WITHOUT a follow-up Read. Lower-ranked files list as bare paths. Prose/doc and stylesheet matches are partitioned into secondary lists so they do not crowd out source. Related files (sharing a rare identifier with a top hit, no import edge between them) are surfaced as first-class neighbors.

NAMING: case- and separator-insensitive (LoadStaging ≡ load_staging). It does NOT expand synonyms or plurals for you — that is your job: if the concept could be named two ways, pass both tokens.

AFTER THE RESULT:

  1. A path + its inline symbols/preview answer the question → done, no Read needed.

  2. Path looks right but you need shape/usage → gs on it (symbols + imports + per-symbol callers + importers in one shot).

  3. "no indexed file contains any of [...]" → those identifiers do not exist in the repo; reformulate or grep for a phrase/regex find cannot index. Do not grep spelling variants of a token find already reported absent.

Inputs: query (required).

gsA

Drill into a known file. Returns these sections as compact text:

  • Symbols — top-level + per-class methods (name, kind, line range, extends/implements). With cross-file callers attached per exported symbol (inline if 1 caller; newline-per-caller block if ≥2). For huge files (>20 symbols, no match), symbols are reordered by caller count (most-used first) and truncated to top 15.

  • Imports — 1-hop internal outbound dependencies (library imports stripped).

  • Importers — 1-hop reverse: files in this repo that import this one. With match, additionally lists EVERY indexed file (importer or not, any language) whose CONTENT references the matched term even when its filename does not (a registry, admin, or config file using the symbol — or a frontend file referencing a backend name). That subsection IS the complete "who uses " answer: it is exhaustive over indexed content, so a subsystem absent from it does NOT use the symbol — do not grep to enumerate or re-verify use-sites, and do not keep hunting in subsystems the section rules out.

  • Edited together — files that keep changing in the SAME COMMITS as this one, from git history, shown with the count as evidence ("changed together in 14 of this file's 17 commits"). This is a HABIT, not a dependency: no import or call edge connects them, so do not go looking for one. It is frequently what an edit here forgets — a sibling implementation with a shared interface, a test in another language, a doc or template that has to stay in sync. Files connected to this one by an import edge are omitted here, so everything shown is a file the graph did NOT give you. Rendered in EVERY view, including a narrowed one. Absent = no git history yet (a new file, a shallow clone) or nothing pairs strongly enough.

  • Related — files sharing rare identifier/string-literal tokens with this file (with match: with the matched symbols' code region), shown only when NO import edge connects them. These are name-reference relations the import graph cannot see — Django migrations↔models, config-by-name registration, cross-language (JS↔Python) pairs. Treat them as first-class neighbors: the shared token shown is the reason they are related. Rendered in EVERY view, including a narrowed one. Use this AFTER find surfaces a candidate file. This is the right tool for "who uses this file" / "who calls this symbol" — no separate call needed.

view controls which sections you get (default full = all of them). symbols, imports, importers, callers each return one section in isolation when you want a byte-light answer.

For god-files (large classes, large routers, large config modules), pass match to filter symbols/imports/importers/callers to one area — e.g. match: "auth" or match: "/^handle/". Substring is case-insensitive; wrap in slashes for regex.

Prefer this over Read when you need shape, neighbors, or usage. Reach for Read only for implementation details inside a method body. If you have called gs on 5+ files for one question, you are enumerating — go back to find with a sharper path glob or a different concept token instead.

Inputs: file_path (required), view, match, symbol.

kb_lookupA

Read everything the NOTEBOOK knows at ONE exact address — the file you are about to edit. Address-keyed, not concept-keyed: unlike kb_search (fuzzy, ranked), this filters notes by exact anchor path, so it is exhaustive at that path. Returns the file note's facets, every cross-file flow that passes THROUGH this file, and lessons/absences anchored here — each with a live freshness stamp against the current code.

Reach for this the moment you have DECIDED on a file and are about to modify it: it surfaces the flow you might break, the absence you are about to violate, the rationale for the code's shape. A clean result ("nothing known here") is itself a positive signal — proceed. Anything marked [evidence changed: ] drifted since it was verified; re-verify before relying on it, and correct the note with kb_write if it proved wrong. Pass an optional symbol to narrow to one top-level symbol at that path.

Inputs: path (required), symbol.

kb_repairA

List the notebook notes that are WRITTEN BUT UNFINDABLE — missing the fields a note cannot be retrieved without (anchor symbols, a flow's verified paths). Notes written before those fields were required are correct but unreachable, and this is how they get found.

Returns a worklist, never a change: repairing is your work, because every gap needs a judgement about the code (which words a reader would search for, which symbols the note is actually about, which files you can honestly claim to have read). Fix each one with kb_write, passing the note's id — fields merge, so nothing already in the note is lost.

A clean notebook returns "Nothing to repair here." Run it when the user asks to repair, fix, or clean up the notebook.

Inputs: no named parameters.

kb_statusA

Notebook overview: how many notes exist (by type: file/flow/lesson), how many are flagged stale (their anchored files drifted since verification), and how many are superseded/retracted. Pass paths (array of repo-relative paths) to instead list the notes anchored at each of those exact paths with their freshness state — a quick "is anything known here?" check across several files at once.

Inputs: paths.

kb_writeA

Save or correct NOTEBOOK notes after finishing real work here — you have the files in context, so no future agent is better placed to record what you learned. Write a file note (what a file is for), a flow note (how a task spans files), or an absence lesson (a confirmed "there is no X"). Pass specs (an ARRAY of specs) to write ALL your notes in ONE call — malformed notes are reported together instead of one hiding the rest, and the result ends with a coverage line naming any worked files still without a note; pass spec (a single object) for a one-off. Also the tool to FIX or RETRACT a note you used that proved wrong: op: "put" replaces; op: "retract" with target: {kind: "note"} removes the whole note (or target: {kind: "anchor"|"facet"|"invariant"|"behavior"|"feature", key: "<value>"} removes one part).

TWO-PHASE reuse gate: a flow/lesson spec sent WITHOUT an id first searches the notebook for the same concept. If plausible matches exist, kb_write returns {status:"candidates", candidates:[...]} INSTEAD of writing — re-call with into: "<id>" to merge into an existing note, or is_new: true to declare a genuinely new one. This makes note identity reliable (matching, not guessing an exact title). File notes skip the gate (id derives from the path).

The spec shape: type ("file"|"flow"|"lesson", or sugar "file-hub"/"file-single"), title, summary, anchors ([{path, symbols?}] — the addresses the note is about, which drive freshness), plus type-specific fields (file: facets/character; flow: steps/verified; lesson: kind:"absence"/scope/body). Call with NO arguments to get the full spec guide.

REQUIRED for the note to be findable at all — file-single: anchors[].symbols; file-hub: facets[].symbol; flow: steps + verified; lesson: scope.terms. Search ranks a note by its title and prose, plus its anchors in kb search: anchor symbols are what answers a query typed as an identifier.

This tool WRITES to the repo notebook; it never commits to git — publishing notes is a human-only step (coldstart kb commit).

Inputs: into, spec, agent, specs, +2 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 7 tools

    Disambiguation4/5

    The core split between find (search files) and gs (drill into a known file) is very clear, and the notebook tools are mostly well differentiated by axis: fuzzy search, exact-address lookup, status, repair, write. The main overlap is kb_lookup and kb_status with paths, which both operate on exact paths and freshness, though one returns note contents while the other returns a listing.

    Naming Consistency3/5

    The notebook subset follows a clean kb_verb pattern (kb_search, kb_lookup, kb_status, kb_write, kb_repair), but the code tools break it with terse names like find and the abbreviation gs. The set is still readable, but the conventions are visibly mixed rather than one predictable pattern.

    Tool Count5/5

    Seven tools is well scoped for a codebase-exploration plus notebook server. Each tool has a distinct role, and there are no redundant filler tools.

    Completeness4/5

    The notebook lifecycle is well covered: search, exact lookup, status, repair, write, and retract via kb_write. The code side covers ranked discovery and file drilling, but lacks a direct raw-file or regex/phrase search tool, though the descriptions imply external Read/Grep/Glob tools fill that gap.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to keep persistent, verifiable memory of a codebase, including the reasons behind code, prior rejected approaches, and invariants, anchored to the code and carried along as the code moves.
      -
    • A
      license
      A
      quality
      A
      maintenance
      One local memory every coding agent on your machine shares — Claude Code, Cursor, Claude Desktop, OpenCode. Answers from past sessions, commits, CLAUDE.md/AGENTS.md and ADRs with the exact source, and withholds work another agent has claimed so two agents don't rewrite the same thing.
      3
      4,718 PyPI
      2
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Gives AI coding agents persistent cross-session memory by storing atomic markdown knowledge cards with bidirectional wiki-links that agents recall before starting work. Syncs across clients and devices via git without any vector database or embeddings.
      9 npm
      ISC
    • A
      license
      Not graded
      quality
      A
      maintenance
      Gives AI coding agents persistent, cross-session memory by saving insights as atomic markdown cards with bidirectional links and recalling relevant ones before new work. Requires no vector database or embeddings, and cards can be shared across clients or synced between devices via git.
      9 npm
      ISC