😤 DopeForge

Give your AI agent every public API on the internet. One command.

pip install "dopeforge[mcp]"
forge mcp    # ← your agent now discovers and calls real public APIs

Ask Claude "what's the weather in Tokyo, the current Bitcoin price, and a cat fact" — and watch it discover and call three public APIs it had never heard of, live, through one MCP server. No per-API setup. No keys. Nothing leaves your machine except the API calls themselves.

What this is

Every awesome-list on GitHub is discovery for humans: a giant markdown file you Ctrl-F through, full of links that died silently years ago. DopeForge is the opposite — a living catalog built for agents (and impatient humans):

1. One registry — 6,623 entries. The biggest developer catalogs on GitHub, 838k+ combined stars of curated markdown, parsed, validated, and merged into one searchable database:

   Source	Entries	Brings
   public-apis ⭐474k	1,721	Public APIs + auth/HTTPS/CORS metadata
   awesome-mcp-servers ⭐93k	3,482	MCP servers tagged by language/scope
   free-for-dev ⭐136k	1,304	Dev services with free tiers
   awesome-llm-apps ⭐135k	116	LLM apps, agents, RAG examples

2. Actually callable. Catalogs give you a link. DopeForge gives you blueprints — small, hand-verified specs (base URL, endpoints, params) that make APIs invocable right now, from the CLI or from your agent:

   forge try public-apis-open-meteo forecast --param latitude=35.68 --param longitude=139.69

3. Proven alive. forge check calls every blueprinted API for real, and a nightly CI run does it in the cloud. Awesome lists rot — while building v1 we found entire categories of "popular" APIs that are already dead (expired TLS certs, no HTTPS listener, deprecation stubs — receipts in docs/blueprint-notes.md). Those never made it in. Your agent is never handed a dead tool.

Related MCP server: cocoon

For humans: the CLI

pip install dopeforge           # registry snapshot ships inside — works offline, instantly

forge search "sentiment analysis"            # search all 4 catalogs at once
forge search image --kind api --no-auth      # public APIs needing zero keys
forge search database --kind mcp-server      # MCP servers only
forge info public-apis-open-meteo            # full details for one entry
forge try public-apis-cat-facts-2            # call an API right from your terminal
forge stats                                  # what's in the registry
forge sync                                   # refresh from the live sources
forge check                                  # liveness-test every blueprint
forge random                                 # feeling lucky

Kinds: api · mcp-server · free-service · llm-app

Until the first PyPI release lands, install from source: pip install "dopeforge[mcp] @ git+https://github.com/DreadpiratePickles/dopeforge"

For agents: the MCP server

{ "mcpServers": { "dopeforge": { "command": "forge", "args": ["mcp"] } } }

Three tools, safe by construction:

• discover — search the whole registry with filters

• inspect — full metadata + call blueprint for any entry

• call_api — invoke a blueprinted API (GET-only, no-auth-only, HTTPS-only, validated params, bounded responses, explicit timeouts — the agent can't spend your money or leak your keys, because there are no keys)

The 15 launch blueprints cover weather, crypto prices, IP/geo, dictionary, books, exchange rates, trivia, quotes, nationality prediction, ZIP lookup, random users, and an irresponsible number of jokes and cat facts. Every one is live-verified at build time and re-verified nightly.

Make an API callable (the contribution flywheel)

A blueprint is ~20 lines of YAML:

id: public-apis-open-meteo        # must match a registry entry id
base_url: https://api.open-meteo.com/v1
auth: none
endpoints:
  - name: forecast
    path: /forecast
    description: Weather forecast for coordinates
    params:
      - {name: latitude,  required: true, type: number, example: 52.52}
      - {name: longitude, required: true, type: number, example: 13.41}

Drop it in src/dopeforge/data/blueprints/, run forge check, open a PR. Every merged blueprint makes every agent using DopeForge more capable. That's the flywheel.

Roadmap

Phase 1  ✅ Living Registry     4 catalogs → one validated, searchable database
Phase 2  ✅ Agent-native layer  forge mcp · blueprints · forge try · self-testing registry
Phase 3  🚧 Blueprint factory   LLM-assisted blueprint generation from API docs (human-verified)
Phase 4  ⏳ Auth'd blueprints   opt-in local keyring for api-key APIs (keys never touch the agent)
Phase 5  ⏳ Visual composer     chain discovered APIs on a canvas, export Python

Endgame: any API → AI-powered capability in one command — local-first, your data stays yours.

How it works

Each catalog gets a narrow parser adapter (src/dopeforge/sources/) that turns its particular markdown dialect — tables, nested bullets, emoji legends — into one pydantic-validated schema. Parsers are pure functions tested against real snapshots of each source (tests/fixtures/), extracting 99.6–100% of entries. forge sync reports per-source success/failure explicitly — a source that breaks its format shows up as a failure, never as silently missing data.

Registry resolution: $DOPEFORGE_DATA_DIR → ~/.local/share/dopeforge → packaged snapshot

Development

git clone https://github.com/DreadpiratePickles/dopeforge && cd dopeforge
python3 -m venv .venv && .venv/bin/pip install -e ".[dev,mcp]"
.venv/bin/pytest -q      # 130 tests, all against real fixture snapshots
.venv/bin/ruff check .

Adding a catalog source = one adapter + one fixture + one test. Adding a callable API = one YAML file. Specs live in docs/ and are the binding contracts for each phase.

License

MIT