agent-kit

A provider-neutral agent kit and a local artifact-memory substrate for a personal multi-repo workspace.

Two things live here:

1. A knowledge kit — skills, agents, entrypoints and references maintained once, compiled into per-provider packs (Claude Code, Codex CLI, OpenCode) by generators, and served to a Claude Code session over MCP.

2. An artifact-memory substrate — a local pipeline that catalogs the workspace's canonical Markdown, chunks it into an immutable outbox, embeds it locally, and serves semantic search from a loopback Qdrant behind a resident Unix-socket service. Nothing leaves the machine.

Design rules

• Canonical Markdown is the source of truth. The search index is derived — a discovery hint, never an authority. Retrieve the real file before acting.

• Derived state is rebuildable; canonical state is never rewritten to fix an index.

• Fail closed. A hash mismatch, a missing build manifest, an incomplete backup: each stops the pipeline and says so rather than shipping something unverified.

• Local embeddings only. A pinned ONNX model (all-MiniLM-L6-v2, 384-dim) loaded from a digest-checked local snapshot; runtime downloads are refused.

Related MCP server: mcp-service

⚠️ The rebuild rule — read before editing src/

The registered server command runs the compiled output (node dist/index.js), not the TypeScript source. An edit under src/ is invisible until you npm run build and restart the Claude Code session (the MCP server is spawned per session). npm run dev runs the source directly, but that is not what Claude Code launches.

Content under data/ is loaded once at server startup — no hot reload, but no rebuild needed either. Restart the session after editing it.

Architecture

canonical Markdown
      ├─────────────► artifact_catalog.py ──► catalog (sqlite, exact sha256)
      │                                            │
skills │ emit receipts                             ▼
      │                        artifact_ingestion.py ──► immutable gzip-JSONL outbox
      ▼                                            │
 skill-events/  ──► artifact_event_consumer.py ────┤ (checkpointed, dead-lettered)
                                                   ▼
                                    Qdrant (loopback, API-keyed)
                                                   │
                                artifact_memory_service.py (resident, UDS)
                                                   │
                                       MCP tools ──┘  search_artifacts / get_artifact

Support plane: an external watchdog (a dead consumer cannot announce its own death), a composite health watermark, an age-encrypted off-device backup with a fail-closed --verify, a reversible quarantine, and a retention classifier that defaults to RETAIN.

MCP tools (17)

All are read-only. The server exposes exactly one transport (stdio); a static tripwire test fails if an HTTP/SSE transport is ever imported.

What the default profile writes to disk. Under the default SERVER_PROFILE=personal, the token log keeps a 120-character raw preview of every tool call's arguments (summarizeArgs), and the response-cache snapshot keeps raw argsKeys. Both default to $WORKSPACE_ROOT/.claude/mcp-token-log.jsonl and $WORKSPACE_ROOT/.claude/mcp-cache-snapshot.jsonl. SERVER_PROFILE=shared switches the log to hashArgs and writes to distinct *.shared.jsonl files. Both patterns are gitignored, but if you point TOKEN_LOG_PATH or CACHE_SNAPSHOT_PATH somewhere else, that is on you. Set either to "" to disable it.

Under SERVER_PROFILE=shared the personal-memory tier is excluded from both the search index and the memory tools, and the artifact-memory tools are not registered at all.

Quick start

git clone https://github.com/chris-dare-dev/agent-kit
cd agent-kit
npm ci && npm run build

claude mcp add agent-kit -- node "$PWD/dist/index.js"

No environment variables are required: with none set, the server serves the content bundled in data/ and resolves everything else relative to the package root. PLATFORM_ROOT (extra content root), WORKSPACE_ROOT, MEMORY_ROOT, CONTEXT_GUIDES_DIR and CLAUDE_MD_GLOBS are optional overrides — see the header of src/config.ts.

That block is executed by node scripts/verify-quickstart.mjs, so if it drifts from what actually works, the check fails.

Optionally, plant the bundled skills, agents, commands and hooks into a workspace's .claude/, and check your machine:

node scripts/cli.mjs init      # symlink on POSIX, copy on Windows
node scripts/cli.mjs doctor    # one PASS/FAIL/SKIP row per prerequisite

init refuses to touch anything it did not create unless you pass --force, never writes outside the clone without --install-to, and records every path it touched in .agent-kit/install-receipt.json. Run init --dry-run to see the plan first.

The substrate is separate and optional — see workspace-tooling/README.md for provisioning (Docker + Qdrant + a Python 3.12 venv from the pinned lockfile).

Supported platforms

This table is what is measured today, not what is intended. Every entry was run on Windows 11 and on Linux; macOS entries are marked as inferred where no host was available to check them.

native — runs here, no workaround. WSL2 — Windows users run it inside a WSL2 guest. partial — runs and reports, with known residue that is tracked. unsupported — declines by name; see the linked milestone.

Substrate suite, measured. The suite used to be un-importable off POSIX; M2 routed fcntl, os.geteuid and the resource/AF_UNIX dependencies through workspace-tooling/platform_compat.py, so it now collects everywhere. It does not yet pass everywhere:

workspace-tooling/run-substrate-tests.py records these as baselines and tells you whether a run is at, above or below them — so "did I break something" is answerable without a clean tree to diff against. On POSIX, ~30 of the errors are cases that import qdrant_client at call time and clear once the provisioned venv is used. The Windows residue (SQLite handles held across tempdir teardown, macOS-only launchd fixtures) is #70.

macOS is inferred, not measured. No macOS host was available. Its rows assume POSIX parity with Linux, which is what the code implies but not something anyone has run.

Unsupported means it says so. artifact_memory_service.py on Windows exits 2 in under a second naming the platform, the missing AF_UNIX support, and the WSL2 alternative — rather than emitting a cascade of import errors that read like substrate defects.

On Windows the MCP server starts and serves the 13 non-substrate tools, and says on stderr exactly which group is unavailable and why.

Testing

npm run test:all        # everything below that runs without a provisioned substrate

npm run test:unit       # all 10 TypeScript suites (globbed) + init + uninstall
npm run gates           # every generator gate, consistency check and shell harness
npm run verify:quickstart

# substrate suite — runs on every OS and reports against a recorded baseline
# (Linux 627 collected / 1 fail / 30 error; Windows 625 / 31 / 337). Exits
# non-zero whenever anything failed, matching baseline or not.
npm run test:substrate

# The substrate suite needs the provisioned venv to clear ~30 of those errors:
# they are cases that import qdrant_client at call time. It is NOT part of
# test:all for that reason. Windows residue is tracked as #70.

# `npm run gates` above runs all of these and prints one named line per gate;
# reach for them individually only when you want one gate's own diagnostics.
# No PYTHONUTF8 needed: every text I/O in data/scripts names its encoding, and
# catalog provenance is POSIX-separated, so these are byte-identical on Windows,
# macOS and Linux. Both were verified on Windows AND Linux.
python3 data/scripts/catalog-generate.py --check
python3 data/scripts/generate-adapter-packs.py --check
python3 data/scripts/generate-root-contract.py --check
python3 data/scripts/model-policy-apply.py --check
python3 data/scripts/mcp-server-name-check.py --check
python3 data/scripts/denylist-check.py --check
python3 data/scripts/template-settings-check.py --check

The last one is why a rename is survivable: every mcp__<server>__ tool grant in data/**/*.md must name a server that all three shipped registration templates register. Change the server name and it tells you every file still carrying the old grant.

agent-kit doctor runs all of the above prerequisite checks in one pass, with the right encoding, and prints a fix command per failure.

AGENTS.md and CONTEXT.md are generated — edit the coverage map, then regenerate.

Layout

Provenance

This kit began as a personal fork of an internal tool I built at work, then was genericized: the employer-specific integrations, infrastructure facts, and operational tooling were removed, and the remaining machinery renamed. What is left is the general architecture — the generators, the guard, the memory substrate — plus content I use on my own projects.

src/security/aidefence-rules.ts transcribes detection patterns from ruflo (MIT) as data; see that file's header for the attribution and the one deliberate deviation.

Requirements

Server: Node ≥ 20. Nothing else — macOS, Linux and Windows alike.

Substrate (optional): Python 3.12 · Docker (Qdrant). The Python modules and their test suite import and run on macOS, Linux and Windows; the resident memory service needs a Unix-domain socket, so on Windows it runs inside WSL2 — see Supported platforms and docs/platforms/windows-wsl.md.