Prerequisites
- Node.js 22.8.0 or newer
- npm 11.10 or newer
- pnpm 11 or newer (only needed for the web app)
- Python 3.10 or newer (for the IPython kernel used by every session)
- An API key or OAuth account for at least one provider (see Providers and models)
1. Install
Clone the repo and install the agent workspace..npmrc in the repo pins a 7-day min-release-age on dependencies. The first npm ci may take a minute longer than usual because of that.
For the full install matrix — web app, kernel Python, IDE integration — see Install.
2. Sign in a provider
Start the agent CLI once and use the/login slash command to add credentials for a provider.
~/.prime/agent/auth.json with mode 0600.
You can also pass a key inline for a single run without persisting it:
3. Run your first turn
Ask the agent to do something that exercises the IPython kernel:- A thinking card while the model plans.
- An IPython tool card that lists directories and runs a
Pathglob to count files. - A shell tool card if the model reaches for
findinstead. - A final assistant message with the result.
ChatStreamEvent — see Streaming protocol for the wire format.
4. Try a subagent
Inside an IPython cell the model can callrlm(...) to spawn a child agent. Ask:
5. Save what worked with /refine
If the run went well, refine the agent’s supplemental state so it does better next time:
/refine reads the trajectory, extracts evidence, and updates supplemental harness state. It never rewrites the immutable base prompt, and it takes a snapshot so you can roll back. See Refinement.
Next steps
Learn the concepts
Architecture, sessions, subagents, streaming events, refinement.
Try more features
Guides for tool cards, providers, session branching, slash commands, MCP, OAuth.
Pick your interface
CLI, TUI, daemon, or the Qredence web app.
Configure and secure
Config paths, environment variables, security model.