connectorgen

Connectors for any HTTP API or MCP server, generated from what already exists, kept working as APIs change.

Point connectorgen at an OpenAPI or Swagger spec, an API's documentation pages, or an MCP server, and it generates a connector: actions with input schemas, auth, pagination and triggers. Your app runs any connector the same way, with one execute(action, inputs) call. When the API changes, connectorgen detects it, regenerates the connector, keeps your customizations, and proposes the update for approval.

A connector is data, not code (one JSON file), so it can be validated, diffed, tested and repaired. It's built as a library for workflow and integration tools, with a CLI on top.

OpenAPI / Swagger 2.0 ─┐
API docs (+ LLM) ──────┼─► Connector (JSON) ─► one runtime: execute · paginate · triggers · OAuth
MCP server ────────────┘         │                └─► or serve it as an MCP server
                                 └─► contract tests · live smoke tests · check/diff/apply when APIs change

Tested on 740 real-world API specs (99.8% of 43,000 generated actions pass their contract tests), plus the full Stripe and GitHub APIs and the official MCP reference servers.

Works with any LLM, and needs none. Generating, running, testing and repairing connectors use no model. The optional AI features (filling gaps, reading docs, diagnosing rejected requests) work with Claude, OpenAI, or any OpenAI-compatible model, including local ones via Ollama.

Install

pip install connectorgen              # core: generate, run, test, repair
pip install "connectorgen[mcp]"       # + MCP client/server
pip install "connectorgen[ai]"        # + the Anthropic SDK (other LLM providers need no extra)

Quick start

CLI

# 1. Generate a connector from an OpenAPI 3.x or Swagger 2.0 spec (URL or file)
connectorgen generate https://example.com/openapi.yaml -o example.json
connectorgen convert swagger.json -o openapi.yaml     # or just convert Swagger 2.0 to OpenAPI 3

# 2. See what it found
connectorgen inspect example.json

# 3. Call an action (credentials come from env vars)
export CONNECTORGEN_API_KEY=...        # or CONNECTORGEN_TOKEN / _USERNAME + _PASSWORD
connectorgen call example.json get_pet -i petId=42
connectorgen call example.json list_pets -i limit=50 --paginate --max-pages 3
connectorgen call example.json create_pet --body pet.json
connectorgen call example.json delete_pet -i petId=42 --allow-destructive

# 4. Test it offline, and keep it current
connectorgen test example.json
connectorgen check example.json        # re-reads the spec; proposes an update if it changed

Python

from connectorgen import connector_from_openapi, ConnectorClient, Credentials

connector = connector_from_openapi("openapi.yaml")
connector.save("petstore.json")           # store / version / diff this

with ConnectorClient(connector, Credentials(api_key="...")) as client:
    pet = client.execute("get_pet", petId=42)
    client.execute("create_pet", body={"name": "Rex"})
    for pet in client.paginate("list_pets", page_size=100):
        print(pet["name"])

Inputs are validated against each action's JSON Schema before any request goes out (InputValidationError); non-2xx responses raise ActionHTTPError with status and body.

MCP, both ways

# Use an MCP server as a connector (--env names a secret the server needs; only the name is saved)
connectorgen from-mcp --command "npx -y @modelcontextprotocol/server-filesystem /data" -o fs.json
connectorgen call fs.json read_text_file -i path=/data/notes.txt

# Serve any connector as an MCP server (destructive actions are left out unless allowed)
connectorgen serve github.json --tag issues

Embedding in a workflow tool

Everything a host needs comes from the connector file, whatever its source:

Errors share one base class, ConnectorError, with specific subclasses (InputValidationError, ActionHTTPError, RateLimitError, McpToolError, …) so a host can map them to step failures, retries, or user-facing messages.

What's supported

On the full Stripe (612 actions) and GitHub (1,242 actions) specs, every action passes its contract test, and pagination is detected for 133 and 274 actions respectively.

Documentation

• MCP: use MCP servers as connectors, and serve connectors as MCP servers
• Running connectors: retries, rate limits, response validation
• Testing connectors: offline contract tests and live smoke tests
• Triggers: polling and webhooks
• OAuth2: sign-in, token storage and refresh
• AI enrichment and docs ingestion: any LLM, guardrails, building connectors from docs
• Repair loop: keeping connectors current as APIs change

Contributing

Contributions are welcome, especially reports of specs that don't work. See CONTRIBUTING.md for setup, tests and design principles, and SECURITY.md for reporting vulnerabilities privately. Notable changes are listed in CHANGELOG.md.

License

Apache License 2.0. Copyright 2026 Ahmad Karawash.