Works with every major agent

You picked your agent — Claude Code, or Cursor, or Vibe, or whatever you reach for at 2am. brainclaw doesn’t ask you to switch. It runs as an MCP server (Model Context Protocol — the standard the major agents speak), and any of them connects to it natively. No adapter code. No per-agent SDK. No proprietary protocol you have to babysit.

That’s what makes “shared memory across agents” actually possible: Monday’s Claude Code session captures a trap, Wednesday’s Codex session reads it without translation, Friday’s Vibe run picks up the half-done plan.

# Same MCP server, every agent
bclaw_work intent=execute scope="src/auth/"
bclaw_get  entity=trap id=trp_abc123
bclaw_create entity=decision data={text: "...", author: "..."}

66+ tools sitting on 17 entities — six canonical verbs cover the whole CRUD surface, plus facades like bclaw_work and bclaw_coordinate that bundle the daily loop into one call. Schemas are versioned so you can pin and upgrade on your terms.

Here’s what that looks like from the agent’s side — the same MCP surface, across onboarding, code navigation, parallel dispatch and review loops:

Claude Code· acme-checkout brainclaw MCP

Reads shared memory in one call — bclaw_work.

Every plan, decision, trap and sequence lives in .brainclaw/, versioned in git — exposed as MCP tools to any agent (Claude, Codex, Copilot…), or as a CLI when you want one. The agent picks up the project; you don't re-explain it.

The same server also exposes a structural Code Map of your repo (bclaw_code_find, bclaw_code_brief) — so the agent navigates your code, not just your memory.

See the integrations page for the full per-agent capability matrix.

Analytics only if you say yes.