Integrations

Editors, harnesses and connectors that actually exist today: not the ones a README hoped for.

ValorBrain talks to any MCP client and any HTTP client. The pages below are the combinations we run ourselves or ship installers for. The connection steps have one source, connect.md, repeated in the Quickstart.

Connect

The canonical path to a working vbm_, in at most three commands.

Claude Code

claude mcp add --transport http with a vbm_ bearer, or stdio.

Codex

codex mcp add with --bearer-token-env-var VALORBRAIN_MCP_TOKEN.

Gemini CLI

httpUrl in ~/.gemini/settings.json.

Cursor / Windsurf

~/.cursor/mcp.json with a vbm_ bearer, or OAuth DCR.

VS Code

.vscode/mcp.json with type http.

Devin

Custom MCP over HTTP with OAuth: each person signs in as themselves.

Grok

Harness templates + stdio through the CLI.

Hermes

Native MemoryProvider plugin.

OpenClaw

memory slot plugin.

ZCode

Marketplace plugin with hooks and a skill.

Connectors

Notion, Drive, Slack, GitHub, Linear, HubSpot, Jira and Granatum: BYOK, paid plans. And the claude.ai custom connector.

Hosted vs on-premise

Hosted (start here): no engine on your laptop. The harness talks to MCP at https://mcpbrain.valor.digital/mcp with a vbm_ token (the signup's mcp_token, or one created under Settings → MCP tokens), or to REST at https://valorbrain-api.valor.digital, with the same vbm_ or the legacy vb_agent_ key (the signup's api_key). The MCP takes only vbm_: a vb_ key on MCP gets 401 wrong_credential_type.

Any stdio-only client uses npx -y @valorbrain/cli mcp as the server command. The optional full install with rules and hooks is npx -y @valorbrain/connect@latest --token <vbm_…>.

On-premise (Enterprise): the engine on your network is part of the Enterprise plan — there is no community or self-service edition and no public docker compose up; deployment is accompanied by our team. What Enterprise includes.

The quality loop (every harness)

Consult before answering — memory is authoritative about the operation; training data is not.
Declare what you used — memory_used with the docids you relied on. When the human confirms or corrects, pass verdict.

If a README tells you to run valorbrain setup harness <name>, that subcommand is not on the published CLI. Use the page for that harness instead.