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

# Connect an AI assistant over MCP

> Let Cursor, Claude or VS Code read your runs so you can ask about a curve instead of exporting one.

```json theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
{
  "mcpServers": {
    "tuneplane": {
      "url": "https://tuneplane.your-company.com/api/mcp/rpc"
    }
  }
}
```

Paste that into your client's MCP configuration. On first connect it opens a browser for
authorisation, and then you can ask about a job directly:

```text theme={"theme":{"light":"github-light","dark":"github-dark-dimmed"}}
Analyze the reward curve and KL of run-4f2a91, compare it against the previous run of the same
experiment, and suggest the next tuning step.
```

The console's [MCP page](https://docs.tuneplane.ai/en/console/mcp.md) generates the exact snippet for your client, which is
easier than assembling it by hand.

## Set it up

**1. Open the MCP page**

Console → **MCP**. It is a premium feature, so the entry may not be there.

**2. Copy the snippet for your client**

Cursor, Claude, VS Code and others each want it in a slightly different place. The page shows
the right shape for each.

**3. Authorise**

The client opens a browser for OAuth 2.1 with PKCE when it connects. There is no token to
paste.

**4. Pick which tools are on**

Read tools are on by default. Write tools are off until you turn them on.

## What the assistant can read

| Group | Tools |
| - | - |
| Jobs | `list_jobs` `get_job` `get_metrics` `get_metrics_summary` `get_logs` `get_hardware` `get_environment` `get_validation_samples` |
| Configuration and comparison | `get_job_config` `get_hyperparams` `compare_jobs` `get_job_diagnosis` `get_project_memory` `list_projects` |
| Platform surface | `list_methods` `get_benchmarks` `get_sweep` `list_servings` |

`list_methods` is the one worth knowing about: with a method name it returns the full
hyperparameter declaration including types and ranges, so an assistant proposing a change can
check it is legal before suggesting it.

## What it can write, once you allow it

| Tool | What it does |
| - | - |
| `stop_job` | Stops one of your running or queued jobs |
| `submit_job` | Clones a job's code and profile, applies overrides, and submits it. **Consumes quota** |
| `create_submit_plan` | Puts a set of changes in the approval queue. Nothing is submitted until somebody approves it in the console |

> [!WARNING]
>
> `submit_job` spends GPU hours. Quota and validation still apply, and the platform produces a
> submission plan for confirmation first, but an assistant with this tool on can queue work. Think
> twice on a shared deployment.

`create_submit_plan` is the safer shape for the same intent: the assistant proposes, a person
approves.

## Check it worked

Ask your client to list its tools. You should see the read tools above and nothing else, unless
you enabled a write tool.

Then ask for something concrete: `get_metrics_summary` on a finished run returns each metric's
latest value with its min and max and the step each happened at. If the client reports "not
authenticated: complete the Connect / login authorization in your MCP client", the OAuth round
trip did not finish.

## Endpoint and protocol

* Transport is Streamable HTTP MCP, protocol revision `2026-07-28`.
* The same endpoint serves modern clients, which discover with `server/discover` and no
  handshake, and older clients that send `initialize` on `2025-11-25` and earlier.
* Authorisation is OAuth 2.1 with PKCE. Client ID Metadata Documents are preferred, and Dynamic
  Client Registration stays as a fallback. Authorisation-code redirects carry RFC 9207 `iss`,
  which is skipped only in local `TUNEPLANE_NO_AUTH=1` debugging.

## Next

[The MCP page](https://docs.tuneplane.ai/en/console/mcp.md) · [Hugging Face](https://docs.tuneplane.ai/en/integrations/huggingface.md) ·
[Job detail](https://docs.tuneplane.ai/en/console/job-detail.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.