MCP server

Every action in Odexy — as an AI tool.

We run a Model Context Protocol server that auto-generates a typed tool for every action in the platform an agent may take — all but the ones marked dangerous. Point Claude, Cursor, Zed, Sourcegraph, or your own MCP-speaking agent at it, and the AI does the work your team does through the same actions — under the same permissions, into the same audit log.

Read the setup docs AI-native story →

What ships

One MCP server. Every module.

Auto-generated tool catalogue. Every action an agent may take becomes an MCP tool at server boot, narrowed to what your key allows. Zero hand-authored tool definitions; when we ship a new action, it shows up in MCP the same release.
Typed input + output schemas. Zod-authored per-action schemas surface as JSON Schema on the tool's input + output — so your LLM knows the exact shape, and malformed inputs get rejected at the boundary.
Per-actor authorization. AI tokens are separate from user cookies. Each token is scoped (can be restricted below the delegating user's permissions), short-lived, revocable. Never a "god mode" for the AI.
Dangerous actions stay with people. Actions the platform marks dangerous — deleting, voiding, refunding, sending something outside your workspace — are never offered over MCP, and a call that names one is refused with a pointer to the screen that does it. The built-in assistant refuses them the same way.
Query actions vs mutations. Read-only queries (crm.deal.list, hrm.employee.get) return paginated + typed projections. Every call an agent makes is recorded in the audit log, reads included, and so is every refused attempt.
Scoped agent keys. Give an agent its own API key (iam.user.apiKeys.create), narrowed to a subset of your permissions and set to expire when you choose. Revoke it any time.

Clients we test with

MCP is a protocol, not a lock-in.

We test our server against Claude (Anthropic's flagship MCP client), Cursor, Zed, Sourcegraph Cody, and our own in-app agent. Any MCP-speaking client works; if it doesn't, that's a bug on our side, not yours.

Claude Desktop / Claude Code
Reference MCP client; our default target.
Cursor
MCP tool support in the sidebar; tested against every release.
Zed
Native MCP support since Zed 0.148.
Custom agents
Any MCP-speaking runtime (Python mcp package, Node @modelcontextprotocol/sdk, self-authored).

Setup

Three steps.

1. Create an API key. Settings → Security → API keys: name it and create it. Copy the hak_… key — it is shown once.
2. Add the server to your MCP client. Add https://app.odexy.app/api/mcp as a remote (HTTP) MCP server and send the key in the Authorization: Bearer hak_… header. The key decides which workspace the agent works in.
3. Ask the AI to do something. "Show me the deals that closed this week." "Draft a follow-up to Acme." "Advance Maria's application to interview." The AI picks the right tool, calls it, and the result flows back into the conversation.

MCP

Wire your agent up.

No plugin. No custom SDK. Standard MCP over your existing token.

Standard MCP over HTTP

Any MCP client — Claude Desktop, Cursor, Zed, your own — talks the same protocol. No shim, no wrapper.

Same permission model

The agent inherits your permissions and can only be further restricted, never expanded. Every action is a policy call away.

Full audit

Every AI tool call, reads included, lands in the same audit log the UI writes to: who, the input, and the result or the error, with passwords, keys and tokens masked.

Read the setup docs How AI operates the UI