> ## Documentation Index
> Fetch the complete documentation index at: https://starforge.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP Integration

> AI agents like Cursor / Claude read job data over MCP to assist with tuning

MCP (Model Context Protocol) exposes job configurations, hyperparameters, metrics, logs, and validation samples to your AI coding assistant — ask directly in Cursor / Claude: "analyze the reward curve of job xxx and suggest how to tune lr and KL".

## Setup steps

<Steps>
  <Step title="Open the MCP page">
    Console → **MCP** (`/mcp`).
  </Step>

  <Step title="Copy the client configuration">
    Pick your client (Cursor / Claude / VS Code, etc.) and copy the generated JSON snippet into the client's MCP configuration.
  </Step>

  <Step title="OAuth authorization">
    When the client connects, it automatically opens the browser to complete OAuth 2.1 + PKCE authorization — no manual token entry.
  </Step>

  <Step title="Start asking">
    <br />

    ```text theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
    Analyze the reward curve and KL of job lab-grpo-20260816,
    compare against the previous run of the same experiment,
    and suggest the next tuning step.
    ```
  </Step>
</Steps>

## Tools and permissions

The MCP page lets you toggle tools individually. **Read tools** are enabled by default; **write tools** are disabled by default and must be explicitly enabled.

| Category               | Tools                                                                                                                                                                                      |
| ---------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Job observability      | `list_jobs` `get_job` `get_metrics` `get_metrics_summary` `get_logs` `get_hardware` `get_environment` `get_validation_samples`                                                             |
| Config & comparison    | `get_job_config` `get_hyperparams` `compare_jobs` `get_job_diagnosis` `get_project_memory` `list_projects`                                                                                 |
| Platform surface       | `list_methods` (method catalog with hyperparameter declarations), `get_benchmarks` (benchmark score matrix), `get_sweep` (sweep variant comparison), `list_servings` (Playground services) |
| Write (off by default) | `stop_job`, `submit_job` (tweak & resubmit), `create_submit_plan` (queue for approval)                                                                                                     |

<Warning>
  Enabling write tools means your AI agent can actually submit jobs (subject to full quota and validation, and the platform generates a submission plan for confirmation first). Be cautious about enabling them in shared team environments.
</Warning>

## Endpoint and protocol

* Transport: Streamable HTTP MCP ([protocol 2026-07-28](https://modelcontextprotocol.io/specification/2026-07-28))
* Dual-era: the same endpoint serves modern clients (no handshake, `server/discover`) and legacy `initialize` clients on 2025-11-25 and earlier
* Auth: OAuth 2.1 + PKCE. Prefers [Client ID Metadata Documents](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/client-registration); Dynamic Client Registration remains as a compatibility fallback. Authorization-code redirects include RFC 9207 `iss` (skipped in `FORGE_NO_AUTH=1` local debugging mode)

Console path: `/mcp`