Extensions

MCP serversCopy Anchor Link

MCP servers add tools to the agent. A server named studio exposing a timeline tool appears as studio.timeline. Server names can't reuse a built-in family name such as local, web, media, agent, google, or seo.

Manage servers with /mcp or rolit mcp, or list them in .rolit/config.json. Connections go directly from the CLI to the server, over stdio or HTTP. Tool definitions are deferred and loaded on demand with tool_search.

{
  "tools": {
    "mcpServers": [
      {
        "name": "example",
        "enabled": true,
        "transport": { "kind": "http", "url": "https://mcp.example.com/mcp" },
        "auth": "rolit",
        "timeoutSecs": 300,
        "riskOverrides": { "send_campaign": "send" }
      },
      {
        "name": "local-files",
        "transport": { "kind": "stdio", "command": "my-mcp-server", "args": ["--stdio"] }
      }
    ]
  }
}

ApprovalCopy Anchor Link

Every MCP call is the External MCP tier. It asks unless session Auto is on, and it is denied in plan-only and non-interactive sessions.

Risk can only be raised, never lowered. A tool is raised by riskOverrides, by labels in the built-in ESP, CMS, CRM, and ads connector maps, and by name: a tool called like send_email, publish_post, create_ad, or update_budget is treated as Send, Publish, or Spend, and the prompt says (risk inferred from tool name). A raised tool asks in every mode, becomes a pending action in scripts, and is never available to a subagent when it is Send or Spend.

A server that fails to connect stays listed in /tools as MCP server · connection failed with the error. Image results are passed to the model as images and not as base64 text.

Language servers (lsp.*)Copy Anchor Link

Give the agent code intelligence from language servers. Servers start on first use and stay alive for the session. Results use workspace-relative paths and are redacted.

/lsp                        # list the catalog and current state
/lsp install rust           # rust-analyzer, via rustup
/lsp install typescript     # typescript-language-server, via npm
/lsp install python         # python-lsp-server, via pip
/lsp install go             # gopls, via the Go toolchain
/lsp rust off               # disable an installed server

Installs run in the background with a ten-minute limit, then restart the session with history preserved. The required ecosystem tool (rustup, npm, python3, or go) must be installed. To use another server, add it under tools.lspServers:

{
  "tools": {
    "lspServers": [
      { "name": "rust", "command": "rust-analyzer", "extensions": ["rs"], "languageId": "rust" }
    ]
  }
}

All are Read tier.

DuckDB data (duckdb.*)Copy Anchor Link

Load files, query them with SQL, and chart the results in the terminal. Interactive chat only. Install the engine first:

/setup duckdb

See Setup.

The model supplies a table name, SQL, and column names, never a path, command, or server.

Open the charts with /data (←/→ switches charts, t shows the table, s stops a live chart). /data stop [id] stops live charts.

Subagents (agent.*)Copy Anchor Link

A subagent is a second run with its own context, tool budget, and journal, working on the same workspace in the background. Use it to hand off a longer, independent goal.

• Run /spawn <goal> yourself, or let the agent spawn one.
• Press Ctrl+G to open the subagent panel: j/k select, x cancels, Esc closes. The composer footer shows how many are running or waiting.
• A subagent can't spawn other subagents and can't ask you questions. It asks its parent with agent.ask_parent.

Capability levelsCopy Anchor Link

The level bounds what a subagent may ever do, whatever the parent asks for.

A subagent runs under its parent's approval mode and the same workspace. Send and Spend tools are never available to it.

SettingsCopy Anchor Link

Change them in the /agents panel, with /agents model <id|clear>, level <c0-c3>, pane <n>, process <n>, or timeout <secs>, or in .rolit/config.json. Changes apply to the next session.

{
  "tools": {
    "subagents": {
      "capabilityLevel": "c3",
      "maxConcurrentPerPane": 4,
      "maxConcurrentPerProcess": 12,
      "defaultTimeoutSecs": 600,
      "maxTimeoutSecs": 1800,
      "defaultEscalationTimeoutSecs": 300
    }
  }
}

Set defaultModel to a faster or cheaper model for routine delegated work. Without it, a subagent uses its parent's model.

ToolsCopy Anchor Link

A subagent's final answer comes back to the parent as ordinary user-level content, so it gains no extra authority. A subagent that is killed mid-run is recovered as orphaned on /resume, not restarted silently.