oomol-lab/open-connector

OpenConnector – A Gateway for AI Agents to Use External Services

What it is | OpenConnector is an open‑source “connector gateway” that lets AI agents (or any app) call thousands of third‑party services – GitHub, Gmail, Notion, Slack, BigQuery, etc. – without giving the agent direct access to the user’s credentials. It provides a catalog of 1 000+ providers and 10 000+ pre‑built Actions (individual API calls) that agents can discover, inspect, and invoke.

Why it matters for AI | Modern LLM‑powered agents need to act on real‑world data and perform tasks (send email, update a spreadsheet, query a database). Managing OAuth flows, API keys, and policy enforcement for each agent is complex and a security risk. OpenConnector centralises credential handling and exposes a stable, inspectable API that agents can call safely.

Key components

• Connector catalog – a shared registry of providers and Actions with request/response schemas, required scopes, and executable code.
• Runtime gateway – a Node.js/Cloudflare‑compatible service that stores credentials, enforces policies, logs runs, and executes Actions.
• Developer tools –
  • Connector SDK (TypeScript client) for programmatic use.
  • oo CLI for local agent relay and debugging.
  • MCP endpoint for agents that understand the MCP protocol.
  • Direct HTTP/OpenAPI access.
• Web console – UI for browsing the catalog, configuring credentials, creating runtime tokens, and inspecting logs.

---

Deployment options

• Managed (OOMOL) – hosted OAuth apps and runtime, no infrastructure to run.
• Self‑hosted – Docker Compose or plain Node.js; stores state in SQLite (default) or PostgreSQL; can use S3‑compatible storage for file transit.
• Platform‑agnostic – works on Cloudflare Workers, Fly.io, RepoCloud, nibrun, NEXUS AI, etc.

---

Typical workflow

1. Connect a user account – via API key or OAuth2 (e.g., PUT /api/connections/github).
2. Agent discovers an Action – queries the catalog or OpenAPI spec for github.get_current_user.
3. Agent calls the Action – POST /v1/actions/github.get_current_user with the runtime token.
4. Gateway executes – uses stored credentials, applies any allow/block policies, logs the run, and returns the result.

---

Who should use it

• Teams building AI‑driven products that need reusable, auditable access to users’ SaaS tools.
• Developers who want a plug‑and‑play catalog of actions without writing individual OAuth integrations.
• Organizations that prefer to keep credentials behind a controlled runtime while still allowing agents to act.

---

Getting started

# Quick local launch (Docker Compose pulls the pre‑built image)
 docker compose up
 # Verify with a no‑auth action
 curl -X POST http://localhost:3000/v1/actions/hackernews.get_top_stories \
      -H 'content-type: application/json' -d '{"input":{}}'

Then open http://localhost:3000 for the console or http://localhost:3000/docs for the OpenAPI reference.

License & community

• Apache‑2.0 license for the code; third‑party provider assets are not covered.
• Open to contributions (see CONTRIBUTING.md), with a friendly code of conduct.

OpenConnector therefore serves as the plumbing that lets LLM agents safely and reliably interact with the vast ecosystem of web services.