The MCP server.

How a coding agent works with Perfloop, and the tools a member can call.

Updated 28 September 2026.

Connect

The server is at https://app.perfloop.ai/mcp. It requires a login: the agent signs in with your Perfloop account, and your role in the workspace decides what it can read or change. There is no anonymous access. Add Perfloop to your agent, then sign in — every agent Setup offers, with the same steps Setup gives:

claude mcp add --transport http perfloop https://app.perfloop.ai/mcp

{
  "mcpServers": {
    "perfloop": {
      "url": "https://app.perfloop.ai/mcp"
    }
  }
}

codex mcp add perfloop --url https://app.perfloop.ai/mcp --oauth-resource https://app.perfloop.ai/mcp
codex mcp login perfloop

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "perfloop": {
      "type": "remote",
      "url": "https://app.perfloop.ai/mcp"
    }
  }
}

{
  "mcpServers": {
    "perfloop": {
      "url": "https://app.perfloop.ai/mcp"
    }
  }
}

amp mcp add perfloop https://app.perfloop.ai/mcp

For automation with nobody present, a member API key with read,contribute stands in for the login. Add admin to that key when the automation reads Usage or changes workspace setup.

What the agent can do

Reads return the model, cases, Initiatives, the Inbox, Usage, and submissions, in bounded pages. Writes start or stop Sessions, file feedback, decide Inbox items, and change an Initiative's direction. Every write is recorded under the authenticated user or API key. The server cannot record a result, and it cannot approve or merge a pull request.

Admins can allow a public HTTPS path or its exact host from the Inbox. Approved destinations have no count limit. Manage the complete list in Setup under Network access. Removing a destination blocks the next matching read unless another destination still covers it. The Inbox keeps at most 100 network requests per workspace. At capacity, new requests replace decided rows first, then the oldest pending request whose requesting Session has ended. A live Session can request a removed destination again. Removal grants no access.

Model building

Call model with section: "buildStatus", then call buildModel to start one Session. It accepts:

• modelSeq (required): the model sequence returned in that build status. The build records the live model sequence at start; a stale value does not reject the request.
• repoID (optional): the repo to build. Omit it to let Perfloop choose the next repo.
• focus (optional, requires repoID): an exact repo-relative path to a source file or directory. Add name for a case-sensitive source name as written under that path, such as Server.serve. The build stops with a named gap if the name is not written there.
• patternSlugs (optional, requires repoID): active pattern slugs from the catalog. Omit it for a broad build. An empty list, repeated slug, or unavailable pattern is rejected.

For example, use your model sequence and replace repo_… with your repo ID:

{
  "modelSeq": 7,
  "repoID": "repo_…",
  "patternSlugs": [
    "identity_collision",
    "resource_leak"
  ]
}

With selected patterns, the Session reuses fresh mapping or maps first, then hunts for each pattern over the related paths. Completed work is reused; unfinished work stays pending. Normal Session limits apply. The result reports this bounded attempt, not coverage of the whole repo. Autonomous model building does not need to be enabled.

Tools

What each tool is for, grouped by what you do with it; your agent reads each tool's inputs from the server itself. These can start a paid Session: approve, submitIdea, workCase, buildModel, and startResearch; the rest start none. Tools only a workspace admin can call are not listed here; the catalog marks them. Every docs page, this one included, is also a resource on the server, so an agent can read it without leaving the session.

To read a Case's timeline or attempts, pass its caseID and the section. Add sessionID for one Session, or candidateID for one attempt. A redundant first on an exact read is ignored. To page the attempts, pass the previous candidateIndex.pageInfo.endCursor as after; a row number is not a cursor.

Once a Session ends, its row in the timeline carries sessionResult and sessionEndedAt. A Session that stopped early also carries sessionStopReason, a fixed code such as case_session_stalled or controller_replaced that says why.

A Case summary and its hypothesis section return the hypothesis's forecast. The model's hypotheses section returns it too: size (S, M, L, XL, or unknown), reason, sourceChanged, and evidence artifact IDs. The forecast is null when no assessment is available. sourceChanged: true means the source differs from the assessment's saved source basis. A false value can be inconclusive when that basis is missing. Private guidance references are not returned.

Also in this section: The loop, Case states, Steering and learning, Initiatives, and Autonomy and limits.

Questions: hello@perfloop.ai