ContextLib

io.github.willykeenanPublished by GitHub account willykeenanv0.1.0Updated Oct 25, 2026

A project's long-term memory as plain Markdown files, with an append-only, tamper-evident log.

Overview

AI-generated overview

ContextLib gives an assistant a project's long-term memory stored as plain Markdown files, with an append-only, tamper-evident log.

What you can do
Read the project brief and search, fetch and record decisions, requirements, facts, lessons, glossary entries, sources and returned work as Markdown records with front matter and body hashes. Supersede, review, accept or reject records, and list records whose review date is due. Capture a result file into a return record with proposed lessons, and import or export projects as bundles. Report library status and health checks.
When to use it
Use it when a project needs durable context that survives agent sessions and stays readable, diffable and portable as files, rather than living only in chat logs or a model-only store. Suited to teams that want an auditable record of decisions and lessons shared between people and several agents.
Requirements
A local process started with the ctxlib CLI (Python 3.9+, standard library only) or the published Docker image, with the library mounted at /library. The library path comes from --library, the CONTEXTLIB_ROOT environment variable, or ~/ContextLib. An identity is fixed at startup with --identity. Git is optional for sync.
Before you install
The assistant writes, supersedes, reviews and imports records in the library, so changes are real and should be reviewed. The identity is fixed per process and tools cannot write as another identity. Absolute home paths and secrets are refused at write time, but avoid putting sensitive material in records. Export bundles may include private records unless excluded.

Installation

In SourceWeft

  1. Open ContextLib in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

ContextLib

A project library that outlives every agent session.

[The sample library rendered read-only: project brief, current decisions, open requirements and key facts]

Real screenshot of the sample library (an invented project).

[A ctxlib session: an owner's decision lands current, an agent's lesson waits in the inbox, a secret is refused, and doctor passes]
A ctxlib session: an owner's decision lands current, an agent's lesson waits in the inbox, a secret is refused, and doctor passes.

Each project's knowledge — decisions, requirements, facts, lessons, glossary, sources and returned work — is stored as plain Markdown files with a small front-matter block, in one folder per project. The folders can sit on an external SSD. A person can understand a project from Finder alone. Agents read and write the same files through a CLI (ctxlib) and an MCP server (context_* tools). Any database is a search index rebuilt from the files, never the source of truth.

It is the open core of the context plugin of AgentBrain.

The Python package is agentbrain_contextlib. The CLI is ctxlib. Python 3.9+, standard library only. Git is optional.

Why

Agent sessions end. Chat logs drift. A store that only a model can parse is a dead end the first time you open Finder on an airplane.

ContextLib keeps the durable layer as files you can read, diff, copy, and take to another Mac. The search index is disposable. The Markdown is the product.

60-second quickstart

bash
pip install "git+https://github.com/willykeenan/agentbrain-contextlib"
ctxlib init ~/ContextLib --project lighthouse-app \  --name "Lighthouse App" \  --purpose "Harbor dashboard for lamp status, fog signals, and traffic"
ctxlib add lighthouse-app decision "Use keeper logs as lamp status" \  --body "**Decision:** Lamp status comes from keeper logs, not AIS." \  --author mara \  --tag lamp --tag keeper-log
ctxlib brief lighthouse-appctxlib doctor lighthouse-app --ten-minute

The library path is --library PATH, else env CONTEXTLIB_ROOT, else ~/ContextLib.

A finished sample you can open in Finder without installing anything lives at examples/sample-library/. Start at that README, then the project BRIEF, then decisions/INDEX.md.

Full walkthrough: docs/QUICKSTART.md.

CLI (ctxlib)

ctxlib [--library PATH] COMMAND
CommandWhat it does
init PATH [--project SLUG --name NAME --purpose TEXT]Create a library (and optionally the first project)
project add SLUG NAME [--purpose]Add a project folder
project listList projects
add PROJECT TYPE TITLE (--body TEXT | --body-file F) --author ID [options]Write a record (--evidence, --tag, --review-by, --approved-by)
get PROJECT IDPrint one record, with evidence state and supersession chain
search PROJECT QUERY [--type] [--all]Search the rebuildable index
brief PROJECTPrint the always-loaded summary (≤ 8192 bytes)
supersede PROJECT OLD NEW --reason TEXT --author IDMark old superseded, move the file, keep the body hash
review PROJECT ID confirm|retire --note TEXT --author IDConfirm or retire a record
inbox PROJECTList proposed records
accept PROJECT ID --author IDMove inbox → current
reject PROJECT ID --reason TEXT --author IDMove inbox → rejected
due PROJECTRecords whose review_by is due
regen PROJECT|--allRewrite BRIEF, INDEX, GLOSSARY, TIMELINE, README
rebuildRebuild the SQLite search index from files
capture PROJECT RESULT.jsonTurn a result file into a return (and proposed lessons)
import-md PROJECT FILE --type TYPE --author IDMarkdown file → record
export PROJECT OUT_DIR [--include-private]Zip a project
import BUNDLEImport a zip
doctor [PROJECT] [--ten-minute]Plain-English health check
syncgit pull / git push when the library is a git repo; otherwise a no-op with a message
mcp --identity IDStdio MCP server; the identity is fixed for the process

Types: decision, requirement, fact, lesson, glossary, source, return, brief-note.

MCP (context_* tools)

ctxlib mcp --identity ID is a stdio JSON-RPC server. Protocol versions: 2024-11-05, 2025-03-26, 2025-06-18.

Tools: context_brief, context_search, context_get, context_record, context_supersede, context_review_due, context_review, context_capture, context_export, context_import, context_status.

The identity is fixed at start. Tools cannot write as another identity. Errors come back as isError results.

Claude Code

json
{  "mcpServers": {    "contextlib": {      "command": "ctxlib",      "args": ["mcp", "--identity", "claude:project"],      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }    }  }}

Codex

json
{  "mcpServers": {    "contextlib": {      "command": "ctxlib",      "args": ["mcp", "--identity", "codex:project"],      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }    }  }}

Cursor

In .cursor/mcp.json (project) or ~/.cursor/mcp.json (user):

json
{  "mcpServers": {    "contextlib": {      "command": "ctxlib",      "args": ["mcp", "--identity", "cursor:project"],      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }    }  }}

Details: docs/MCP.md.

Record format

A record is UTF-8 Markdown. It starts with a YAML-like front-matter block (scalars, quoted strings, lists of scalars, lists of flat maps) and a Markdown body. The body hash is stored in the front-matter; bodies are not rewritten.

---id: dec-20260926-7f3atype: decisiontitle: Use plain files as the source of truthstatus: currentproject: lighthouse-appscope: {lane: workstream, team: owner}created: 2026-09-26T14:03:00Zauthor: maraapproved_by: mara — "files first, database is only an index"evidence:  - {ref: "project:BRIEF.md"}  - {ref: "url:https://example.com/x"}supersedes: []superseded_by: nullreview_by: 2027-03-25tags: [storage, library]sensitivity: normalbody_sha256: "…"---**Decision:** …**Why:** …**Consequences:** …**How to verify:** …

Evidence refs are project:<relative path>, ssd:<relative path>, room:<name>#<seq>, url:<https url>, or note:<text>. Absolute home paths and secrets are refused at write time.

Ids: <prefix>-<YYYYMMDD>-<4 hex> with prefixes dec, req, fact, les, glo, src, ret, note.

Files: <YYYY-MM-DD>_<title-slug>_<id-suffix>.md. Superseded records move to superseded/ with the body hash still valid.

Library layout

<library root>/├── README.md              how to read this in 10 minutes├── INDEX.md               one line per project├── library.json           {"format": "contextlib/1", "library_id", "created"}└── projects/<slug>/    ├── README.md BRIEF.md GLOSSARY.md TIMELINE.md project.json    ├── decisions/ requirements/ facts/ lessons/ glossary/ sources/ returns/    │   ├── INDEX.md    │   ├── <YYYY-MM-DD>_<title-slug>_<id-suffix>.md    │   └── superseded/    ├── inbox/             proposed records    ├── exports/    └── .contextlib/       ledger + rebuildable index.sqlite

Keeping the library on an SSD, using it unplugged, and moving it between Macs: docs/SSD.md.

Develop

bash
PYTHONPATH=src python3 -m unittest discover -s testspython3 tools/check.py packaging

Python 3.9+, standard library only. See CONTRIBUTING.md and CHANGELOG.md.

Apache-2.0, copyright KE Studios. SECURITY.md for reports.

ContextLib is the open library layer of AgentBrain (agentrooms.io).

Run with Docker

The image is published at ghcr.io/willykeenan/agentbrain-contextlib for Apple silicon and Intel. Mount your library at /library:

bash
claude mcp add contextlib -- docker run -i --rm --user "$(id -u):$(id -g)" -v ~/ContextLib:/library ghcr.io/willykeenan/agentbrain-contextlib mcp --identity you

Source: README.md at commit f373be1

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.0LatestSep 27, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts