MCP server

@unisonlabs/mcp exposes the brain as native tools in any MCP client. The tool set is deliberately curated - a small tool count keeps an agent's context lean.

Install

# Claude Code
claude mcp add unison-brain -e UNISON_TOKEN=usk_live_... -- npx -y @unisonlabs/mcp

# Codex
codex mcp add unison-brain --env UNISON_TOKEN=usk_live_... -- npx -y @unisonlabs/mcp

// Cursor / Claude Desktop
{
  "mcpServers": {
    "unison-brain": {
      "command": "npx",
      "args": ["-y", "@unisonlabs/mcp"],
      "env": { "UNISON_TOKEN": "usk_live_..." }
    }
  }
}

Brain tools

brain_context, brain_ingest, brain_remember, brain_search, brain_get, brain_list, brain_write, brain_edit, brain_resolve_entity, brain_facts_about, brain_record_fact, brain_status.

The loop agents should run: brain_context at task start, brain_ingest / brain_record_fact when they learn something, brain_remember at the end of a session to curate what's worth keeping, and brain_status to sanity-check the pipeline.

Per-user scoping

Set UNISON_ACTOR=<externalId> in the server env to scope every tool call to a shadow user - see actor delegation.