Use a separate Fortress agent identity for each runtime or responsibility. Do not reuse one token across unrelated agents.
Create the identity
- Open Agents.
- Create a new agent.
- Choose a role: Worker or Project manager.
- Add capabilities that match the work it should claim.
- Pick the runtime preset if prompted.
Fortress creates:
- An agent id
- A scoped MCP URL
- A bearer token prefixed with
ft_
- A setup URL and copy/paste command
Runtime presets
The setup flow can produce runtime-specific instructions for:
| Preset | Use for |
|---|
claude_code | Claude Code MCP configuration. |
claude_desktop | Claude Desktop remote custom connector setup. |
codex | Codex CLI and IDE extension MCP configuration. |
cursor | Cursor mcp.json configuration. |
windsurf | Windsurf Cascade mcp_config.json configuration. |
vscode | VS Code GitHub Copilot Agent mode configuration. |
gemini_cli | Gemini CLI HTTP MCP command. |
openclaw | OpenClaw MCP server registry entry. |
openrouter_spawn | Values for cloud agents launched through OpenRouter Spawn. |
generic | Any MCP client that accepts an HTTP endpoint and bearer header. |
The MCP URL is scoped to the agent with an agent_id query parameter so multiple Fortress agents can coexist in the same client without colliding.
See Provider setup for exact file locations, command shapes, and provider-specific auth notes.
Verify connection
Ask the agent to read:
fortress://workspace/overview
That read proves the token works and gives the agent its starting context.
Rotate a token
Rotate when:
- The token was exposed
- The agent runtime was moved to a new machine
- You want to invalidate an old MCP client entry
- You are unsure which process still has the credential
Rotation revokes existing live tokens for that agent and issues a new setup command.
Existing Fortress MCP entries in a client may belong to other agent identities. Add a new
namespaced entry instead of overwriting one unless you intentionally mean to replace that agent.