Skip to content

Why Rhizome

A coding harness that doesn’t forget. Rhizome runs its own agent loop on any model, free ones included, and keeps what matters as plain Markdown in a vault you own. You can check what the agent did to your files after the fact.

Most agent harnesses either forget everything when a session ends, or remember things in a store you can’t easily read: a hidden memory file, a database, a cloud account. You can’t tell what the agent “knows”, fix it when it’s wrong, or take it with you.

Rhizome keeps memory in a vault: an ordinary folder of Markdown files on your disk. You can start from a template with starter folders and note types, so you don’t have to design a structure first. You can open the same files in any editor, search them with grep, and back them up however you like.

When a chat produces something worth keeping, you click Save to vault. Rhizome writes a note. Nothing is saved without a click unless you turn automatic saving on. How memory works

If the vault is a Git repository, every change to it can be committed, including the agent’s changes. You can see what a note looked like before, see what changed, and revert it. Git is optional, but it’s what makes agent memory safe to trust. Git

Mycelium replays a Prime session as a map. The folder the agent worked in becomes a city of files, and each file the agent searched, read or edited lights up. When something breaks, you can see which files a run touched and where it went in circles. Mycelium uses the open-source Mindwalk project, which you install separately. Mycelium

Taproot, Rhizome’s free automatic router, sends a chat to free hosted models (Groq, Mistral, LLM7, OpenRouter’s free models, NVIDIA NIM) in a fixed order. When one fails or runs out of free use, Rhizome moves to the next. Models whose free tier trains on your prompts are left out. You can add your own endpoint as the last stop. It builds on two open-source projects, OmniRoute and freellmapi. Free models and routing

An agent loop is the cycle that sends your message to a model, runs the tools the model asks for, and repeats until the answer is done. Rhizome has its own, and that is the only chat 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 memory stay with Rhizome.

Rhizome’s engine can read, search, edit and create notes in the open vault. In Power User it can run one approved shell command in the vault folder. Agents, engines and models

If you’ve used Claude Code, Hermes Agent, Cursor or Prime, most of the basics carry over:

You knowIn Rhizome
Sessions you can resumeSaved native chats with automatic titles
Tool approvalPermission modes: Limited tools and Power User
Bring your own modelHosted providers, local models (Ollama, LM Studio), any OpenAI-compatible endpoint; keys in the system keychain
CLAUDE.md / AGENTS.mdA vault AGENTS.md, with CLAUDE.md kept as a compatibility file
MCPRhizome runs an MCP server, so other apps can read and write your vault
Past chat historyImport Claude Code chats into the vault
ExtensionsRhizome’s own plugin system is planned
An editorA rich Markdown editor with a raw mode, plus PDF and image previews, sheets, diagrams and whiteboards
A graphThe Wiki Graph: notes as nodes, links as edges

Coming from another harness maps each one in detail.

Rhizome is a macOS developer preview. You build it from source; there is no signed download or in-app updater. Chat runs on Rhizome’s own engine and does not need Prime.