MCP server

The Sazabi MCP server exposes the Sazabi API to any Model Context Protocol client — a coding agent, an AI IDE, or a custom client. It lets that client query your logs, investigate issues, and run the Sazabi agent through the same operations the API offers.

This page is the protocol reference. For copy-paste setup for a specific client (Claude Code, Cursor, Codex, ChatGPT, and others), see Connect your coding agent, which owns the per-client configuration.

Server URL

Connect over Streamable HTTP:

https://mcp.sazabi.com/mcp

Authentication

Authenticate with a secret key. Create one in the dashboard under Settings > Secret keys or with sazabi secret-keys create <name>. Send it either as a bearer token or in the X-Sazabi-Secret-Key header:

{
  "mcpServers": {
    "sazabi": {
      "type": "http",
      "url": "https://mcp.sazabi.com/mcp",
      "headers": {
        "Authorization": "Bearer sazabi_secret_your_secret_key"
      }
    }
  }
}

The server also supports OAuth: a client that implements the MCP OAuth flow can connect with a browser sign-in and no key at all. Use OAuth where your client supports it; a secret key works in every client.

Secret keys are organization-scoped by default and grant read/write access. Keep them out of source control and rotate them if exposed.

The two tools

The server exposes exactly two tools — search and execute — rather than one tool per feature. New API operations become reachable through search automatically, so the tool surface stays stable as the API grows.

search

Discovers the Sazabi API operations and connected MCP tools the caller can run. Results are compact by default — each hit carries a copy-paste name (the value to pass to execute) and a description.

• query — free-text query matched against operation ids and tool descriptions. Omit or leave empty to list connected MCP connectors.
• namespace — a connector key from a previous search, or sazabi for only the Sazabi API catalog.
• includeInputSchema — set to true only after selecting a hit; pass its namespace and copy its full name into query. The exact lookup returns at most one hit with its input JSON schema and never substitutes a fuzzy match.
• limit / offset — pagination over compact hits (capped at 20 per page).

execute

Invokes one search hit. Copy the name from search and pass an arguments object matching that hit's input schema:

• name — a search hit's name, either sazabi/<operationId> (a Sazabi API operation) or <connectionKey>/<toolName> (a connected MCP connector's tool).
• arguments — an object matching the hit's input schema.
• projectId — optional; targets a specific project's connectors when the credential does not already select one.

The typical flow is: call search to find the tool you need, repeat the search with its namespace, copy its full name into query, and set includeInputSchema: true to load that one schema, then call execute with the returned name and matching arguments.

Deferred operations

Some operations start long-running Sazabi agent work — for example creating a thread or appending a message. For those, execute waits a short time for the run to finish. If it does not complete in that window, the response reports the run as incomplete and returns a run identifier. Call execute again with the runs.get operation and that run identifier to poll for the result.

What you can do

Because every API operation is reachable through execute, the MCP server can do everything the API can — query logs, search and manage issues, create and continue agent threads, and inspect projects. Run search with an empty query to list your connected MCP connectors, or with namespace: "sazabi" to browse the Sazabi operation catalog.

For the full list of operations and their fields, see API reference.

Further reading