Subscription users
You already pay for Claude Pro/Max, ChatGPT or similar. Rinne extracts multi-model orchestration from capacity you already bought, without metering you again.
v0.1.8 · open source · local only
One terminal.Every coding agent.
A CLI harness you talk to directly. Plans a graph, runs the AI tools and model APIs already on your machine and verifies until the goal is met.
No accounts · No telemetry · No hosted component
Conductor · DAG · verifying loop
See how a run works ↓Rinne supports all the world class agents, harnesses, api & beyond
Why Rinne
Orchestration
without the bill
You already pay for coding agents or hold raw API keys, or both. Rinne turns that spend into a multi-model team you actually direct, from one prompt in one terminal.
API-key users
A clean local orchestrator over raw model access. Any OpenAI-compatible provider. Keys live in your OS keychain, never in config files.
Mixed pools
Use a cheap API model as the evaluator and a subscription harness as the generator. One plan, two worker families, one contract.
Local only
Runs on your machine. No hosted component, ever.
Open source
No pricing, no tiers, no accounts.
No telemetry
Network calls are worker calls you configured. Nothing else phones home.
Terminal-first
A CLI that controls other CLIs. Routing is always narrated.
How a run works
Conductor + loop.
Filesystem as substrate.
You never open Claude Code, Codex, Grok or OpenCode yourself. You live in Rinne. It reaches down to those tools as workers. The conductor composes a per-task team; the durable loop drives verification; the filesystem is the shared blackboard.
- 01
Plan
The conductor turns your goal, blackboard digest and worker registry into a JSON DAG: roles, capability requirements, optional preferred workers. Granularity scales with difficulty.
- 02
Schedule
The loop engine resolves a concrete worker for each ready node from live availability and quota. Independent nodes run in parallel. Rate limits cascade instead of killing the run.
- 03
Execute
Harness workers get paths and act autonomously. API workers get assembled context inline. Streams land in the live viewport while the transcript stays in normal scrollback.
- 04
Evaluate & loop
AI, tool or human evaluators gate each result. On failure: loop-back with critique, hand to a fixer or replan the DAG. Stuck loops escalate to you.
Workers
Two families. One contract.
Harness CLIs honor the login you already have. API workers use keys from your keychain. Mix them freely inside a single plan: generator on a subscription harness, evaluator on a cheap API model.
Features
Built for long runs,
not demos
Status: v0.1.8, actively built. Single-machine, single-user. The surface is a harness, not a SaaS, not an IDE.
Conductor planning
Prompt → JSON DAG. One node for easy tasks; multi-node graphs with evaluator loops for hard ones. Re-plans on failure, escalation or new information.
Verifying loop
Generator → evaluator → loop-back with critique until the goal is met or the budget runs out. Evaluators can be AI, a tool (tests) or you.
Pool-aware routing
Tiers and escalation computed from workers actually present. A rate-limited preferred worker never silently kills a node.
Dynamic API support
Any OpenAI-compatible provider, any base URL, rotated keys across rate limits. Connect-time verification surfaces bad keys immediately.
MCP tools & Skills
Attach Model Context Protocol servers and SKILL.md packs. The conductor wires tools and instructions only to the nodes that need them.
Inline TUI
Transcript in native scrollback. Live viewport for the plan tree, active worker stream, conductor narration and prompt, with @-file mentions and tab completion.
Secrets in the keychain
Rinne never writes API keys to config. OS keychain on macOS, Linux and Windows. Env vars still win for one-offs.
Doctor
Detects installed workers, auth mode (subscription / api-key / free) and warns about metered-billing footguns before they cost you.
Install
Get Rinne on your machine
One install script, a Mac app or build from source. Then point Rinne at the workers you already have and describe what you want.
$ curl -fsSL https://raw.githubusercontent.com/GIKSN-RESEARCH/Rinne/main/install.sh | shCLI (recommended)
Prebuilt binary for macOS, Linux and Windows. Checksum verified. Installs to ~/.local/bin.
macOS app
Native Apple Silicon app. Drag to Applications, optional CLI on PATH, then rinne . Shipping soon.
Coming soon
Build from source
Rust 1.85+. cargo build --release from the repository root, or cargo install --path crates/rinne-cli.
First run
rinne doctor: see workers and auth modes already availablerinne connect claude-code: or connect any API provider with a key- Optionally point the conductor at a cheap planner (Groq, Cloudflare, Ollama, …)
- Run
rinneand describe the goal
› @src/api.ts add Redis rate limiting and prove it works. · planning… · Plan: ○ n1 generator ○ n2 evaluator · routed n1 → claude-code · routed n2 → deepseek
Parent lab
GIKSN Research
An independent research lab working on AI and memory. Papers, tools and reasoning shipped in the open across AI, Deeptech, Hardware and Distributed Systems.
Rinne is the first tool the lab has shipped. It is not the memory research; it came out of the memory research. While building the substrate that would give a small model real context, we needed a way to run agents on our own machines without a subscription wall, a hosted account or a single model family. Rinne was the tool we built for ourselves. Then we released it.
The bet
AI agents keep almost working, then fall apart when the task gets real. Two reasons at once: memory that does not give the right context at the right time and small-model ceilings that retrieval alone cannot fix. Working both sides shapes the tools we ship.
How the lab operates
- 01Own research first: pick a problem and work it until something ships.
- 02Ship out of the work: tools land with the paper that argues for their design.
- 03Community alongside: readers and vetted contributors around the research, not instead of it.