Observed since 20 Oct, 21:49 UTC
No superseded metadata states collected. Unchanged observations do not create new revisions.
Bootstrap and maintain a Fumadocs-powered Project Atlas with agent skills, reusable templates, and validation.
Project Atlas documentation system for agent-assisted projects.
agent-docs-kit bootstraps a Fumadocs + MDX Project Atlas, project-local scaffolding scripts, reusable MDX templates, and agent skills for Codex, Claude Code, Copilot, Cursor, Gemini CLI, or a generic agent directory.
The main goal is a broad architecture map: modules, layers, core flows, and architecture direction. It is not a project-management platform and does not ask you to document every feature.
agent-docs-kit is skills-first. It installs SKILL.md workflow skills into the
target project instead of installing slash-command prompt files.
The model follows the same separation that makes tools like Spec Kit scale:
The CLI is a small Python bootstrapper so it can be run with uvx without
requiring Node in the source package. The generated documentation app is still
Node/Fumadocs: after init, previewing, type-checking, and building docs happens
inside the generated docs/ app with npm.
One-time usage:
uvx agent-docs-kit init
uvx agent-docs-kit web --open
Persistent install:
uv tool install agent-docs-kit
agent-docs-kit init
The PyPI distribution installs one public CLI command: agent-docs-kit.
agent-docs-kit init is interactive in a terminal:
? Target project [.]:
? Docs directory [docs]:
? Agent integrations
Up/Down or j/k moves. Space selects. a selects all. Enter confirms.
> [x] Codex (codex)
[ ] Claude Code (claude)
[ ] GitHub Copilot (copilot)
[ ] Cursor (cursor)
[ ] Gemini CLI (gemini)
[ ] Generic (generic)
? Style [atlas]:
? Overwrite existing managed files if needed? [y/N]:
Use explicit flags for scripts or CI:
agent-docs-kit init . --integration codex
agent-docs-kit init . --integration codex --integration claude --integration copilot
agent-docs-kit init . --integration codex --integration cursor --integration gemini
agent-docs-kit init . --integration codex --docs-dir docs --style atlas --yes
agent-docs-kit web
agent-docs-kit atlas --force
agent-docs-kit skills
The starter uses the atlas style by default:
agent-docs-kit init . --style atlas --interactive
agent-docs-kit styles
agent-docs-kit installs the same workflow skills into the best available
instruction surface for each agent. Native skill platforms get direct skill
folders; other agents get portable SKILL.md workflow files plus a project
context file that points to them.
| Integration | Context file | Workflow files |
|---|---|---|
codex | AGENTS.md | .agents/skills/agent-docs-kit-* |
claude | CLAUDE.md | .claude/skills/agent-docs-kit-* |
copilot | .github/copilot-instructions.md | .github/skills/agent-docs-kit-* |
cursor | .cursor/rules/agent-docs-kit.mdc | .agent-docs-kit/skills/agent-docs-kit-* |
gemini | GEMINI.md | .agent-docs-kit/skills/agent-docs-kit-* |
generic | .agent-docs-kit/AGENT_CONTEXT.md | .agent-docs-kit/skills/agent-docs-kit-* |
You can repeat --integration to install several surfaces in one project.
cursor, gemini, and generic share .agent-docs-kit/skills, so a project can
support multiple agents without duplicating the portable workflow files.
There is one bootstrap path: install the CLI, then run agent-docs-kit init in
the target project. The project-local skills are generated during init.
<project>/.agent-docs-kit/
<project>/docs/
<project>/.agents/skills/agent-docs-kit-* # Codex integration
<project>/.claude/skills/agent-docs-kit-* # Claude integration
<project>/.github/skills/agent-docs-kit-* # Copilot integration
<project>/.cursor/rules/agent-docs-kit.mdc # Cursor integration
<project>/GEMINI.md # Gemini CLI integration
<project>/.agent-docs-kit/skills/agent-docs-kit-* # Portable workflows
The per-project workflow skills live in the CLI assets under
src/living_docs_cli/assets/workflow-skills/ and are copied into each target
project during init, so they can use that project's
.agent-docs-kit/config.json, templates, scripts, and context file. There is no
separate global agent-docs-kit helper skill.
.agent-docs-kit/
config.json
templates/
atlas.mdx
architecture.mdx
change.mdx
plan.mdx
glossary.mdx
scripts/
create-doc.mjs
glossary.mjs
check.mjs
docs/
app/
components/
content/docs/
atlas.mdx
lib/
package.json
source.config.ts
.agents/skills/ # codex integration
.claude/skills/ # claude integration
.github/skills/ # copilot skills mode
.cursor/rules/agent-docs-kit.mdc # cursor integration
GEMINI.md # gemini integration
.agent-docs-kit/skills/ # portable workflow skills
AGENTS.md / CLAUDE.md / ... # managed agent-docs-kit context block
The managed context block is bounded by:
<!-- AGENT-DOCS-KIT START -->
...
<!-- AGENT-DOCS-KIT END -->
Re-running init replaces only that block and preserves surrounding project instructions.
Start the local docs site from the project root:
agent-docs-kit web
The command installs docs dependencies when node_modules is missing, finds an
available local port starting at 3333, and prints the URL. Add --open to
open the browser automatically.
Use generated skills from your agent:
agent-docs-kit-write for routing a general docs request to the right workflowagent-docs-kit-architecture for Project Atlas and current architecture docsagent-docs-kit-change for shipped change recordsagent-docs-kit-plan for future design plansagent-docs-kit-glossary after terms changeagent-docs-kit-check before committing docs changesCommon commands:
agent-docs-kit skills
agent-docs-kit web --open
agent-docs-kit atlas --force
agent-docs-kit create-doc change api auth-flow "Auth Flow Update"
agent-docs-kit glossary
agent-docs-kit check
The generated .agent-docs-kit/scripts/ files are implementation details used
by the project-local skills.
Use agent-docs-kit atlas --stdout after init to preview an editable Project
Atlas draft from repository structure. It writes docs/content/docs/atlas.mdx
by default. New projects already include a starter atlas.mdx, so pass
--force only when you want to replace that starter page.
agent-docs-kit ships reusable MDX components in the generated Fumadocs app. Use
them directly in .mdx pages:
<SystemMap
title="System map"
nodes={[
{ title: 'Web app', body: 'User-facing routes.', tone: 'blue' },
{ title: 'API', body: 'Server boundary.', tone: 'green' },
]}
links={[{ from: 'Web app', to: 'API', label: 'calls' }]}
/>
<LayerMap
title="Project layers"
layers={[
{
title: 'Surfaces',
body: 'User-visible or agent-visible entry points.',
tone: 'blue',
items: ['app/', 'api/', 'cli/'],
},
]}
/>
<FlowMap
title="Primary runtime path"
flows={[
{ title: 'Entry', body: 'Input and routing boundary.', tone: 'blue', steps: ['Request', 'Route'] },
{ title: 'Execution', body: 'Core modules and state changes.', tone: 'green', steps: ['Validate', 'Persist'] },
]}
/>
<RoadmapMap
items={[
{ title: 'Current shape', body: 'What the repo and runtime prove today.', status: 'now', tone: 'green' },
{ title: 'Next structural move', body: 'The next architecture-level change.', status: 'next', tone: 'blue' },
]}
/>
<ArchMap
tiers={[
{
title: 'Entry points',
boxes: [
{ title: 'Web app', body: 'User-facing route.', tone: 'blue' },
{ title: 'API', body: 'Server boundary.', tone: 'green' },
],
},
]}
/>
<FlowSteps
title="Request flow"
steps={[
{ title: 'Receive input', body: 'Validate payload.', tone: 'blue' },
{ title: 'Run service', body: 'Execute domain action.', tone: 'green' },
{ title: 'Return result', body: 'Send status and telemetry.', tone: 'amber' },
]}
/>
<StateFlow
title="Lifecycle"
states={[
{ name: 'draft', description: 'Work in progress.', tone: 'amber' },
{ name: 'active', description: 'Current runtime truth.', tone: 'green' },
]}
/>
The generated docs include /docs/components with copyable examples. Use
mermaid for complex sequence diagrams, branches, or graphs that do not fit the
built-in components.
The global CLI can also validate the project:
agent-docs-kit check
agent-docs-kit web
Fumadocs reads MDX from docs/content/docs. If you prefer to run the generated
app directly, use cd docs && npm install && npm run dev -- --port 3333.
Releases are published from GitHub Actions with PyPI Trusted Publishing. This avoids long-lived PyPI API tokens in the repository or on a local machine.
One-time PyPI setup for a new package:
agent-docs-kitwuyuxiangXagent-docs-kitpublish.ymlpypiAfter that, publish a release by tagging the commit:
git tag -a v2.2.2 -m v2.2.2
git push origin v2.2.2
The workflow builds the Python package with uv build and uploads the
distribution files to PyPI.
terms.agent-docs-kit init
agent-docs-kit init [target] [--integration codex|claude|copilot|cursor|gemini|generic] [--docs-dir docs] [--force] [--yes]
agent-docs-kit init . --style atlas --interactive
agent-docs-kit web [target] [--port 3333] [--host 127.0.0.1] [--open] [--no-install]
agent-docs-kit atlas [target] [--output docs/content/docs/atlas.mdx] [--force] [--stdout]
agent-docs-kit create-doc <atlas|architecture|change|plan> <domain> [slug] ["Title"]
agent-docs-kit glossary
agent-docs-kit check
agent-docs-kit skills
agent-docs-kit styles
agent-docs-kit version
agent-docs-kit self check
MIT
Download counts and popularity will appear here as the collector receives them.
| Version | Published | Direct dependencies | Open version |
|---|---|---|---|
| 2.2.2 | 22 Jun 2026 | 0 |
This describes our collected data; it is not a security assessment.
Observed since 20 Oct, 21:49 UTC
No superseded metadata states collected. Unchanged observations do not create new revisions.
package:699b3b1d21376da9f06661dc96cce595ab5d47c06c2a09f5f74763d474200a9d