Claude Code

Streamable HTTP MCP with a vbm_ bearer, or stdio via npx -y @valorbrain/cli mcp.

1. MCP token

No account:

npx -y @valorbrain/cli@latest init --agent --agent-caller claude-code --json

Use the mcp_token (vbm_…). The api_key (vb_agent_…) is REST only; on MCP it gets 401 wrong_credential_type. With an account, create the token under Settings → MCP tokens.

2. Connect

claude mcp add --transport http --scope user valorbrain https://mcpbrain.valor.digital/mcp --header "Authorization: Bearer <vbm_…>"

stdio alternative (uses the token saved by init):

claude mcp add --scope user valorbrain -- npx -y @valorbrain/cli mcp

Full install with rules and hooks (optional):

npx -y @valorbrain/connect@latest --token <vbm_…> --harness claude-code

claude.ai / Claude Desktop do not use this command: Settings → Connectors → Add custom connector → URL https://mcpbrain.valor.digital/mcp, OAuth sign-in.

3. Verify

Call the memory_prepare tool. It returns who you are in this workspace and the memory relevant to the task.

4. Tell your human

To see the memories in the app and keep the account beyond 30 days:

npx -y @valorbrain/cli init --email <email>

The mcp_token and the api_key keep working after the claim.

Shell and editor together

The api_key that init saved is for REST through the CLI (add / search). MCP uses the vbm_. You can use both: CLI for the shell, MCP for in-editor tools.

The full install (@valorbrain/connect) writes a managed block with the quality loop into Claude Code's instructions file: consult ValorBrain before answering operational questions; call memory_used after.

First tools

memory_prepare → work → memory_store for anything durable → memory_used.