MCP tools
What the engine exposes to AI clients over MCP.
When an AI client connects over MCP, it discovers a set of tools. The authoritative list is dynamic — it depends on which connectors you've installed — and is discoverable two ways:
- the client's own
tools/list(refreshed live viatools/listChangedwhen a connector is added), and - the engine's capability manifest at
GET /.well-known/mcp(a JSON document oftools,resources, andprompts).
Built-in tool categories
| Category | What the AI can do |
|---|---|
| Query | Search the graph and get grounded, cited answers. |
| Compile | Ingest and compile sources into claims. |
| Branches | Create/diff/merge branches (and the automatic session branch). |
| Orient | Brief / investigate tools that summarise the graph around an entity. |
| Skills | Load a workspace skill, and skill_define to author a new one. |
| Capability acquisition | mcp_server_install to mount a new MCP server live, expanding the toolset on the fly. |
Connector tools
Any MCP server you install adds its own tools — GitHub's issue/PR tools, Postgres's query tool, and so on. They appear alongside the built-ins, namespaced by server.
Tools can be gated by approval
During an ask, a tool call that needs human sign-off pauses for approval
rather than acting silently. Auto-approve or require review per your setup.
Egress still applies
Tools that reach the network (a connector, or a Root Function's fetch) are
subject to your project's egress allowlist. A tool
can't call a host you haven't permitted.