What is Rna Agent
A local agent that treats the project as a lasting subject. It keeps understanding the project and decides on its own when to act, when to speak and when to stay quiet.
Rna Agent runs on your own machine. Unlike a question-and-answer chat assistant, it works on a project: a real folder, plus its code, documents, conversations, the preferences you stated and the things it promised you.
The conversation is just where you and the project talk. When you leave, the project is still being understood: Rna investigates when there is new evidence, speaks when it needs your decision, and waits quietly when there is nothing to do.
How it differs from a chat assistant
| Chat assistant | Rna Agent | |
|---|---|---|
| Unit of work | One conversation | One project (folder + sessions + memory + capabilities) |
| Initiative | Responds when asked | Chooses to wait, work, share or learn from project evidence, and remembers what it promised |
| Learning | Memory is a piece of context | Corrections become rules, skill revisions or checks that run on writes |
| Judgment cost | Every decision calls the main model | Yes/no, choice and score go to JEV first; the main model only writes |
| When a change goes wrong | You undo it yourself | Subagents and background work change files in their own worktrees; every turn can be rolled back |
| Where it runs | The cloud | A local daemon shared by desktop, CLI and Web |
Parts
- Daemon: manages projects, sessions, scheduling, permissions and recovery. Closing a window does not stop it.
- Workbench: the React desktop UI, packaged with Electron as a macOS app, also usable in a browser.
- CLI: the
rnacommand, connected to the same daemon. - Rna SDK: the in-house execution kernel for model requests, session logs, tool calls, skills, MCP and the initiative loop. No third-party runtime dependencies.
- Memory: retrieval isolated per project, in one of two ways: the built-in local memory (keyword search, no external service) or OpenViking semantic memory (deployed with Docker).
Where to start
Install
Get the desktop app, or start the daemon from source.
Models
Connect an OpenAI-compatible or Anthropic model service.
First project
Pick a folder and see how Rna initializes it.
Initiative
How Rna decides between speaking and staying quiet.
About these docs
Every Rna release ships with an acceptance record, including what failed. When the docs say something is “not yet supported”, that is the real boundary today, not modesty.