agent-kit

A shared toolkit of AI agent utilities for the team. At its centre are the three witan packages — witan-council, witan-code, and witan-core, the shared memory, work-coordination, and code-graph layer for coding agents — published to PyPI alongside agent-config-kit and the ol-agent-kit meta-package, for five in all. The repo also holds reusable skills, custom agent definitions, MCP server install helpers, and sample configurations.

Documentation for the witan packages lives in docs/ — tutorials, guides, a generated reference for every MCP tool, CLI command and environment variable, and the architecture notes. Build it locally with just docs-serve. It publishes on Read the Docs as witan-context. An agent consuming the published site rather than a person browsing it can start at /llms.txt — a link index into every page's raw markdown — or fetch /llms-full.txt for the whole corpus in one file; both are generated by bin/gen_llms.py as part of the site build, not committed.

Repository Structure

.
├── mcp/servers/witan/         # witan-council — memory, tasks, workflow; the `witan` umbrella CLI
├── mcp/servers/witan-code/    # witan-code — tree-sitter code graph + cross-repo bridge
├── packages/witan-core/       # witan-core — shared internals for the two servers above
├── packages/agent-config-kit/ # agent-kit — the CLI that installs the skills/MCP servers declared below
├── packages/agent-kit/        # PyPI meta-package (ol-agent-kit): agent-config-kit[cli] + witan + witan-code
├── skills/                    # Reusable skills, installed via agent-config-kit
├── custom-agents/             # Custom agent definitions for GitHub Copilot and Claude Code
├── mcp/                       # Install helpers and configuration for other MCP servers
├── configs/                   # Sample / reference agent configurations
├── docker/                    # Deployment images: the witan MCP tier and the omnigraph data tier
├── docs/                      # The witan-context documentation site (Zensical → Read the Docs)
├── bin/                       # Repo maintenance scripts (version/pin checks, docs generation)
└── agent-config.toml          # This repo's own manifest for agent-kit

Quick Start

Setup for this repo's own MCP servers and skills is driven by agent-config-kit (agent-kit) and the manifest at agent-config.toml — a single declarative file that installs MCP servers, skills, and hooks into whichever coding-agent platforms it detects (Claude Code, Pi, GitHub Copilot, OpenCode, ...).

Installing agent-kit

uv tool install 'agent-config-kit[cli]'

To also pull in the witan and witan-code MCP servers in one shot, install the ol-agent-kit meta-package instead — it depends on all three and carries no code of its own (agent-kit was already taken on PyPI, hence the ol- prefix on this one; the console script is still agent-kit either way):

uv tool install ol-agent-kit

Applying this repo's manifest

agent-kit apply agent-config.toml                   # everything, every detected platform
agent-kit apply agent-config.toml --profile python  # just the python profile (+ universal)
agent-kit apply agent-config.toml --dry-run         # preview without writing
agent-kit validate agent-config.toml                # check for drift
agent-kit profiles agent-config.toml                # list profiles + entry counts

agent-config.toml's [profiles.*] tables mirror the skills/ category layout — pick the profile matching your specialty (python, infrastructure, containers, dagster, process, ...) to install the universal baseline plus just the skills relevant to your work. Selecting no profile installs the whole catalog, including the toolhive-swe MCP servers below — they aren't part of any profile, so a --profile run skips them.

Run agent-kit apply agent-config.toml --scope project instead of the default global scope to register servers/skills in the current project only rather than user-wide.

This registers the skill catalog plus the toolhive-swe remote MCP server (one entry per environment tier). It does not register witan or witan-code — those own their own registration lifecycle via witan setup (see their READMEs), so they aren't duplicated in this manifest.

See skills/ for the full skill catalog and packages/agent-config-kit/README.md for the full manifest schema, remote skill/hook sources, and profile composition.

Setting up an MCP server

See mcp/ for per-server install scripts and configuration snippets, including witan and witan-code.

Using a custom agent

See custom-agents/ for agent definitions and setup instructions for GitHub Copilot and Claude Code.

Using agent-config-kit for your own project or team skills

agent-kit isn't specific to this repo — any project or team can write its own agent-config.toml manifest to declare MCP servers, skills, and hooks, then run agent-kit apply to install them into whichever coding-agent platforms are detected locally. For local, personal use, point [skills] at paths on disk; for team use, point at a shared repo (this one, or your own) via a git+https://... source so everyone applies the same manifest:

[skills]
my-local-skill = "skills/my-local-skill/SKILL.md"
team-skill     = "git+https://github.com/mitodl/agent-kit.git#subdirectory=skills/process/dependency-updates/SKILL.md"

See packages/agent-config-kit/README.md for the full manifest format, apply/validate/profiles command reference, and how remote (https:///git+) skill and hook sources are fetched and cached.

Contributing

1. Add new skills under skills/<category>/<skill-name>/ — follow the skill authoring guide or use the creating-skills skill.
2. Register the skill in agent-config.toml's [skills] table (and add it to the relevant [profiles.*] entry) so agent-kit apply picks it up.
3. Add new custom agents under custom-agents/<platform>/ — follow the agent authoring guide.
4. Add MCP install helpers under mcp/servers/<server-name>/ — follow the MCP guide.
5. Open a PR with a brief description of what the addition does and why it's useful.

License

BSD-3-Clause