Cartograph

Cartograph is task-shaped repo context for coding agents. It ships a CLI, an MCP server, user-scope install adapters for Claude Code and OpenClaw, and packaged agent assets for documentation-heavy workflows.

• npm: @anthony-maio/cartograph

• MCP Registry: io.github.anthony-maio/cartograph

• Quickstart: cartograph.making-minds.ai

Instead of dumping an entire repository into context, Cartograph ranks the files that matter, maps dependencies, caches structured artifacts, and lets the next tool or agent pick up from those artifacts.

The primary workflow is:

1. analyze to map the repo

2. packet to prepare the work

3. context to load the minimum files

wiki, host installs, and benchmarks are secondary surfaces built around that core path.

Install

npm install
npm run build

For global use from npm:

npm install -g @anthony-maio/cartograph

For MCP host discovery via the official registry:

• io.github.anthony-maio/cartograph

• Registry listing

For Claude Code plugin install from this public repo:

/plugin marketplace add anthony-maio/cartograph
/plugin install cartograph@making-minds-tools

For global use from a local checkout:

npm install -g .

See CONTRIBUTING.md for the development workflow and SECURITY.md for vulnerability reporting.

Related MCP server: MCP Filesystem Server

Claude Code Plugin

This repo also acts as a Claude Code plugin marketplace. The cartograph plugin bundles:

• a plugin-scoped Cartograph MCP server

• slash commands: /cartograph:analyze, /cartograph:context, /cartograph:wiki

• the use-cartograph and repo-surveyor skills

• the documentation-agent bundle: repo-scout, dependency-tracer, context-picker, api-surface-writer, and wiki-writer

Install it with:

/plugin marketplace add anthony-maio/cartograph
/plugin install cartograph@making-minds-tools

Primary Workflow

If you only remember one thing, remember this:

cartograph analyze <repo> --static
cartograph packet <repo> --type <type> --task "<task>"
cartograph context <repo> --task "<task>" --json

• analyze maps the repo and tells you what matters.

• packet turns a concrete job into a reusable working artifact.

• context gives the next agent the smallest useful file set.

Command Surface

cartograph analyze <repo> [options]
cartograph packet <repo> --type <type> --task "<task>" [--changed <paths...>]
cartograph context <repo> --task "<task>" [options]
cartograph wiki <repo> [options]
cartograph export <run-id> --to <path> [--artifact <name>]
cartograph install <claude|openclaw|mcp>
cartograph uninstall <claude|openclaw|mcp>
cartograph doctor [target] [--json]
cartograph mcp

Legacy compatibility still works:

cartograph <repo> --static
cartograph <repo> -c "trace auth flow"

CLI Usage

# Map the repo
cartograph analyze ./my-project --static

# Prepare a concrete job
cartograph packet ./my-project --type bug-fix --task "fix auth refresh bug" --changed src/auth/service.ts tests/auth/service.test.ts

# Load the minimum file set for that job
cartograph context ./my-project --task "add user authentication" --json

# Force embedded snippets when you really want them
cartograph analyze ./my-project --static --json --include-contents

# Full wiki output
cartograph wiki ./my-project -p gemini -k $CARTOGRAPH_API_KEY -o wiki.md

# Export a cached artifact to an explicit path
cartograph export run-abc123 --to ./artifacts/wiki.md

# Run the MCP server directly
cartograph mcp

For small repos, analyze --static --json now defaults to compact output instead of embedding top-file contents. That keeps tiny repos readable and lets direct file reads stay cheaper than a giant JSON blob. Use --include-contents when you explicitly want embedded snippets.

The default markdown output from analyze --static is human-first: it highlights what matters, surfaces dependency hubs, and recommends the next commands instead of dumping raw file contents immediately.

Providers

• gemini

• openai

• openrouter

• ollama

Set the API key with --key or CARTOGRAPH_API_KEY. Ollama does not require a key.

Cache Model

Cartograph writes successful runs into the user cache by default:

• Windows: %USERPROFILE%\\.cartograph\\cache

• POSIX: ~/.cartograph/cache

Each run gets a manifest plus named artifacts, which keeps agent handoffs lightweight and makes cartograph export deterministic.

Host Installs

Cartograph uses an explicit hybrid install model. Installing the package does not modify Claude Code, OpenClaw, or MCP host configs automatically.

Instead, install only the integration you want:

cartograph install claude
cartograph install openclaw
cartograph install mcp

What each target installs:

• claude: user-scope skills plus the bundled documentation agents under ~/.claude

• openclaw: user-scope skill pack under ~/.openclaw

• mcp: a Cartograph MCP config snippet under ~/.cartograph/mcp

Check status at any time:

cartograph doctor
cartograph doctor --json

Choose Your Surface

• CLI: best default path when you want deterministic local analysis and explicit artifacts.

• Claude Code plugin: best when you want plugin install, bundled skills, slash commands, and a plugin-scoped MCP server.

• OpenClaw skills: best when you want the same Cartograph workflow inside an OpenClaw skill-first environment.

• MCP Registry: best when your host already prefers tool wiring over shell commands.

MCP Server

Cartograph's MCP server exposes static repo analysis directly to hosts that prefer MCP over shell commands.

Tools:

• analyze_repo: score files, map dependencies, and return compact analysis output or embedded top-file contents for a local repo or GitHub URL

• get_file_contents: fetch full contents for specific files after analysis

• build_task_packet: return a typed task packet with key files, dependency hubs, validation targets, risks, and task-specific details

Bug-fix packets are tuned to stay focused when you provide --changed or changed_files: they keep explicit change surfaces in view, prefer exact validation targets, and bias toward shared dependencies over peripheral utility scripts.

If you want Cartograph's packaged MCP snippet, run:

cartograph install mcp

That writes a reusable config file that points at:

{
  "mcpServers": {
    "cartograph": {
      "command": "cartograph",
      "args": ["mcp"]
    }
  }
}

MCP Registry

Cartograph is published in the official MCP Registry with the server name io.github.anthony-maio/cartograph.

• View the registry entry

Repo-side metadata lives in:

• package.json via the mcpName field

• server.json for registry metadata

The current install artifact published to npm is:

• @anthony-maio/cartograph

For repeatable registry releases, this repo also includes a GitHub Actions workflow at publish-mcp.yml. It is set up for npm trusted publishing on GitHub Actions, then authenticates to the MCP Registry with GitHub OIDC and publishes server.json.

Packaged Agent Assets

The package currently ships:

• two Claude skills: use-cartograph and repo-surveyor

• five Claude documentation agents: repo-scout, dependency-tracer, context-picker, api-surface-writer, and wiki-writer

• two OpenClaw skills: use-cartograph and repo-surveyor

• bundled OpenProse templates inside both OpenClaw skills for repo-doc and task-context workflows

Skill roles:

• use-cartograph: tool-first path when the CLI or MCP server is available

• repo-surveyor: manual fallback path when Cartograph is unavailable or needs verification

Both skills are designed to produce the same downstream contract:

• key files

• dependency hubs

• minimal task context

• doc-ready summary

These assets are meant to pass run IDs and artifact paths between steps instead of copying large prose into the main context.

The Claude plugin marketplace in this repo ships the same skills and agents, plus plugin-first slash commands and a bundled Cartograph MCP server.

Task Packet Benchmark

This repo includes a public benchmark pack for task packets under benchmarks/task-packets.

Use it to compare packet quality across large visible repos with a stable set of task prompts:

npm run benchmark:task-packets -- --list
npm run benchmark:task-packets -- --case llama-cpp-bug-fix --dry-run

The benchmark runner writes packet artifacts to benchmarks/task-packets/output/, which is gitignored.

Curated scorecards and public artifact links are also published on the site:

• cartograph.making-minds.ai/examples/benchmarks.html

Example Artifacts

Tracked examples live under docs/examples:

• llama-cpp-task-packet.md

• llama-cpp-task-packet.json

• llama-cpp-deepwiki.md

• task-packet-benchmark-scorecards.md

These are useful when you want to show what a focused bug-fix packet and a curated repo brief look like on a large public codebase.

Development

npm install
npm test
npm run check
npm run build
npm run pack:smoke

License

MIT