Open infrastructure for cooperative LLM inference on consumer hardware.
cocompute lets you share an idle GPU sitting on your home network, and use other people's GPUs through an OpenAI-compatible API. Anything that runs Ollama works (NVIDIA, AMD, Apple Silicon, CPU). NAT traversal is handled by iroh, so no port forwarding, no router config, no public IP required.
The protocol is open and AGPL-licensed. The same code that runs at cocompute.ai is the code in this repo. cocompute.ai is the hosted version for people who don't want to operate their own orchestrator.
┌──────────────┐ ┌────────────────────┐ ┌──────────────┐
│ host │ iroh / QUIC │ orchestrator │ HTTP / JSON │ consumer │
│ (your GPU) │◄──────────────►│ (cocompute.ai │◄──────────────►│ (curl, SDK, │
│ runs Ollama │ │ or self-hosted) │ │ openwebui) │
└──────────────┘ └────────────────────┘ └──────────────┘
Pre-launch alpha. Expect breaking changes between minor versions until 1.0. The protocol is stabilizing but not frozen.
The badges at the top of this README show real-time status and uptime for cocompute.ai. They're hosted off-domain (shields.io + UptimeRobot) so they stay reachable when cocompute.ai itself is down. Full history at the public status page.
You can be a host (share your GPU), a consumer (use the pool), or both.
Share your GPU: sign up, click "Add Host" in the dashboard for a one-line install command, run it on any machine that runs Ollama. Service installs via systemd (Linux) or launchd (macOS) and stays running in the background.
curl -sSf https://cocompute.ai/install.sh | COCOMPUTE_URL=https://cocompute.ai bash -s -- --token YOUR_TOKEN(Future) Reciprocity is tracked by pool-credit accounting. No tokens, no crypto.
Use the pool (no public pools yet): sign up at cocompute.ai/quickstart, create an API key, point any OpenAI-compatible client at https://cocompute.ai/v1/. Two minutes if you have curl handy.
curl https://cocompute.ai/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"llama3.2","messages":[{"role":"user","content":"hello"}]}'You can run your own orchestrator and host fleet. The orchestrator is a single Rust binary backed by SQLite. See ARCHITECTURE.md for the moving parts and docs/self-hosting.md for setup steps. Build from source for now (cargo build --release -p cocompute_orchestrator); container images are on the roadmap.
Inference infrastructure is becoming the substrate for everything. A handful of vendors deciding what models you can run, and at what price, is not the future we want. The protocol that lets a 3090 in someone's apartment serve LLM inference to a laptop in a coffee shop should be open, forkable, and operable by anyone. cocompute.ai is the convenient default; the code is the guarantee.
cocompute is licensed AGPLv3. Contributors sign a CLA so the project can offer a commercial license to the hosted operator (cocompute.ai) without forcing all derivatives to inherit AGPL semantics. See CONTRIBUTING.md for the full story.
Issues, PRs, and discussions welcome. Please read CONTRIBUTING.md before opening a PR. All contributions require signing the CLA via cla-assistant.
cocompute stands on the shoulders of:
- iroh for peer-to-peer transport and NAT traversal
- Ollama for the local model runtime
- Axum, Leptos, Sea-ORM, Tokio for the orchestrator stack
AGPLv3. Source code is available; modifications running as a network service must also be available to users of that service. Contributors agree via CLA that the project can dual-license to the hosted operator.