Skip to content

Agents, engines and models

Chat runs on Rhizome’s own engine (ADR-0187). Prime and Hermes are no longer chat engines (ADR-0189); leftover code is still in the repo until the follow-up PRs. The vault, permissions, and saved memory stay the same.

  • Agent loop: the cycle that sends your message to a model, runs the tools the model asks for, and repeats until the answer is done.
  • Engine: what runs that loop for Chat.
  • Model: the language model the engine talks to, hosted or running locally.
  • Tool: an action the agent can take, such as reading a note or running a command.

New chats use Rhizome’s engine: the loop built into the app. It does not need Prime. You pick catalog providers or Taproot, and the vault tools below run against the open vault.

  • Models: pick a catalog model, or Taproot. See Free models and routing. Anthropic models aren’t offered here yet.
  • Tools: see the table below. Read and search tools run without asking. Edits and shell commands ask first.
  • Saved chats: each chat is saved as it goes and can be reopened after you quit. Chats get a title automatically from the first exchange.
  • Errors: a problem that doesn’t stop the turn shows as a notice in the chat. A problem that does stop it, such as a missing key, ends the turn with a message that says what to fix.
ToolWhat it doesWhen it runs
read_file, list_dir, grep, globRead and search files in the open vault and workspaceAuto in both permission modes. Paths outside those roots are refused
edit_file, write_fileChange or create a file in those same rootsAsks every time unless you already allowed that tool and path for the session
create_noteCreate a new Markdown note without overwritingAsks in Limited tools. Auto in Power User
bashRun one shell command with the vault as the working directoryPower User only. Asks every time. There is no session grant
echoReturn its argumentsAuto

bash really runs the command (sh -c on macOS and Linux, cmd /S /C on Windows). Default timeout is 60 seconds, capped at 300. The working directory cannot be chosen by the model. stdin is closed. Output is capped. See the Chat Run command card for the full command, then stdout and stderr after it finishes.

Secret files (.env* except .env.example, .git/, key material) are blocked on the file tools. That denylist does not cover bash. Settings → AI Agents → Safety can turn the denylist off for reads only.

A permission mode decides what the agent may do without asking. It’s set per vault.

ModeRhizome’s engine
Limited toolsRead and search auto-run. create_note asks. Edits ask. No bash
Power UserRead, search, and create_note auto-run. Edits still ask unless a session grant covers that path. bash is offered and asks every time

Rhizome will have its own plugin system: add a tool, hook into a turn, or block a tool call. It’s planned, not shipped. Today, MCP lets other apps use your vault.

MCP (Model Context Protocol) lets other AI apps call Rhizome’s tools. With it, an MCP client such as Claude Code or Cursor can read and write your vault.

Run Set Up External AI Tools… from the command palette (⌘K). Rhizome shows which config files it will change (for example ~/.claude.json and ~/.cursor/mcp.json) and writes nothing until you confirm. It needs Node.js 18 or newer, or Bun, on your PATH. To paste the entry yourself, use Copy MCP config on Settings → AI Agents (ai.panel.copyMcpConfig).

Agent edits are file edits. In a Git-backed vault, review them in the Changes view, look at the diff, and commit or revert. Save chat results you want to keep with Save to vault. See Your vault is the agent’s memory.