MCP Guide, Neuron

The three tools

1. query_company_brain

Ask a question, get a verified answer with sources.

query_company_brain(question: string)
  → {
      answer: string,
      confidence: number,    // 0.0, 1.0
      sources: Source[],
      verified: boolean,
      conflict?: ConflictRef
    }

2. get_company_context

Dump a scoped portion of the verified brain as markdown. Useful at the start of a conversation to load context. Honors workspace permissions.

get_company_context(scope?: string)
  → string  // markdown bundle

3. save_decision

Let an agent record a decision back into the brain. The decision goes through the same extraction pipeline, owner, date, source, and is marked pending until verified.

save_decision(
  rule: string,
  owner: string,
  source: string
) → KnowledgeItem

Installing in Claude Desktop

Add the following to your ~/.claude/mcp-config.json:

{
  "mcpServers": {
    "neuron": {
      "command": "npx",
      "args": ["-y", "@neuron/mcp"],
      "env": {
        "NEURON_API_KEY": "<your key>",
        "NEURON_WORKSPACE_ID": "<your workspace>"
      }
    }
  }
}

Restart Claude Desktop. The three tools appear in your tool palette.

Installing in Cursor

Open Cursor settings → MCP → Add server:

Name: neuron
Command: npx -y @neuron/mcp
Env: NEURON_API_KEY and NEURON_WORKSPACE_ID

Auth

API keys are workspace-scoped and use timing-safe comparison server-side. Generate keys in Settings → API. Each key can be scoped read-only or read-write.

Verification semantics

Calls to query_company_brain only return items where verified === true and frozen === false. If the best match is a frozen item, the response carries a conflict reference instead of an answer, your agent should ask the user for clarification rather than guess.

Rate limits

Read tools: 600 calls / minute / key (sufficient for streaming agent loops).
Write tools: 60 calls / minute / key.
Burst allowance: 2× for 10s windows.

Best practices

Always surface the source URL when you cite Neuron, it's the trust contract.
Pass conflict references through to the user as-is. Don't suppress them.
Use get_company_context sparingly, it returns up to 500 items.
When an agent writes back via save_decision, encourage the user to verify in the web app.

StatusMCP server is GA. SDK and reference clients in the docs.

Your agents, our verified brain.