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

# Develop with the Lab MCP

> Give an agent the same source identities, contracts, documentation and relations that you explore in the Nika Lab.

The Lab MCP helps an agent prepare a Nika change from its sources. It can find a
language element, gather its related contracts and guides, read captured content,
and follow recorded links across repositories. The browser and the MCP use the
same graph identities and page definitions.

<Note>
  The Lab MCP reads a verified capture. It does not report the live state of
  your checkout, execute workflows or certify benchmark results. The
  [Nika MCP oracle](/reference/mcp-server) retains its separate role for the language.
</Note>

## Start in the Lab

Open **Projet & code → Accès agents**. The page explains the reading flow and
lets you try actual MCP calls with a language field, a builtin or a client guide.
A successful response links back to the same native fiche and graph you can read
manually. Each object fiche also offers **Contexte agent**.

```mermaid theme={"system"}
flowchart LR
  Human[Lab page] --> Graph[Shared identities and relations]
  Agent[Your agent] --> MCP[Lab MCP]
  MCP --> Graph
  Graph --> Contract[Spec contracts]
  Graph --> Docs[Documentation]
  Graph --> Code[Pinned source files]
  Graph --> Evidence[Evaluation evidence]
```

The connection indicator reports an actual response from the local server.
Publishing a static Lab page does not deploy an MCP service. If the local endpoint
is unavailable, the workbench explains how to start or restart it.

## Connect an agent

<Steps>
  <Step title="Build the agent pack">
    In a Nika Lab checkout, install the locked dependencies and run:

    ```bash theme={"system"}
    npm ci
    npm run mcp:pack
    ```

    The resulting `.artifacts/nika-lab` directory includes the server,
    integrity-checked captures and reusable skills. It is independent of the
    original checkout after packaging.
  </Step>

  <Step title="Choose MCP only or the complete plugin">
    For an MCP-only connection, run the command for your client from that same
    checkout:

    <CodeGroup>
      ```bash Codex theme={"system"}
      codex mcp add nika-lab -- node "$PWD/.artifacts/nika-lab/runtime/server.mjs"
      ```

      ```bash Claude Code theme={"system"}
      claude mcp add --transport stdio nika-lab -- node "$PWD/.artifacts/nika-lab/runtime/server.mjs"
      ```
    </CodeGroup>

    Other MCP clients can launch `node` with the absolute path to
    `runtime/server.mjs` as its argument. The built server requires Node 20.19
    or later; building from source uses Node 24 or later.

    To load the skills as well, install the built directory as a plugin. Codex
    uses its local marketplace mechanism. Claude Code can load the directory
    with `claude --plugin-dir /absolute/path/to/nika-lab`. Start a new agent
    session after installing or updating it.
  </Step>

  <Step title="Ask for a grounded change">
    “Using the Lab MCP, prepare the context for changing nika:assert. Find its
    contract, related documentation and exact sources to verify. Distinguish
    documented links from qualification evidence.”

    The included context skill guides this investigation. The source-audit
    skill checks identities, citations and versions without merging objects
    just because their titles match.
  </Step>
</Steps>

## Follow the evidence

| Task                        | MCP entry point           | What to keep                                       |
| --------------------------- | ------------------------- | -------------------------------------------------- |
| Establish the capture       | `lab_status`              | Date, fingerprint, owner revisions and limits      |
| Find a subject              | `lab_search`              | Canonical identity and its family                  |
| Understand its role         | `lab_get_node`            | Source definition and classification               |
| Prepare a change            | `lab_development_context` | Contracts, docs, code and evaluation reading paths |
| Inspect associations        | `lab_relations`           | Direction, relation type and recorded evidence     |
| Navigate between identities | `lab_trace_path`          | Original arrows and search limits                  |
| Read the source             | `lab_read`                | Captured content, digest and continuation          |

The server also exposes taxonomy, navigation and capture resources, plus resource
URIs for individual nodes and content. Results are bounded: follow the supplied
cursor or content continuation instead of loading the whole graph into a prompt.

A contract, documentation page, implementation file and published package may
share a subject while retaining different identities. Their relationship is
useful context; it does not make them interchangeable. Verify the captured
revision against the target checkout before changing current code.

## Keep the responsibilities clear

* **Lab MCP:** read captured knowledge, cross-repository relations and imported
  inventories. It does not mutate the graph or read private browser drafts.
* **Nika MCP:** use the language oracle and the capabilities documented by that
  server. A Lab response does not replace `nika check`.
* **Your repository tools:** inspect the actual checkout, make the authorized
  change and run its checks. Do not silently replace a missing pinned source
  with the current branch.

The pack's installation, optional exact-commit Git reads, protocol checks and
remote-hosting boundary are detailed in the Lab's
[implementation guide](https://github.com/supernovae-st/nika-lab/blob/main/docs/AGENTIC_LAB.md).
Continue with [navigating the Lab](/guides/explore-lab) or
[configuring an MCP client](/integrations/mcp-clients).

<Snippet file="_ecosystem.mdx" />