A free, open-source personal AI agent that lives on your machine, remembers everything, and grows in capability as the mesh grows. Apache 2.0 licensed.
Status: v1 mostly built, not yet released. The daemon (Tauri + Rust) compiles end-to-end with HTTP + MCP API surfaces, library-driven Mem0 retrieval, llama.cpp / vLLM inference adapters, a SvelteKit chat UI, settings panel, and curated model registry. The release workflow is wired but no v0.1.0 tag has been cut yet. See "What's built" below for the per-crate state.
Website (planned): ocm.shortcircuit.bot — under the shortcircuit.bot ecosystem.
A peer-to-peer mesh that gives every participant access to AI agent compute by sharing the idle GPU/NPU cycles of their own machine. SETI@home-style volunteer compute, bidirectional, AI-agent-focused.
The user pitch: "Your personal AI agent — lives on your machine, persistent memory, OpenAI-compatible API, free, private. Give compute when you're idle, draw compute when you need it."
The contributor pitch: "Run OCM on your Mac Mini or gaming PC. Other users access AI through your idle GPU; you get cloud-tier agent capability through theirs."
OCM is built as a stack of borrowed open-source frameworks plus a thin original orchestration layer — vLLM/SGLang for inference, llama.cpp for Apple Silicon, Mem0 for persistent memory, libp2p/iroh for mesh transport, MCP + OpenAI-compat for the client API. The novel work is the orchestration and reciprocity ledger.
| Version | What | Effort | Status |
|---|---|---|---|
| v1 | Single-node personal agent — works standalone, no mesh required | 8-12 wks | Mostly built; pre-release |
| v2 | Two-node mesh via iroh / libp2p | 6-8 wks | Trait scaffold landed; impl pending |
| v3 | Reciprocity ledger (give-to-get accounting) | 4-6 wks | After v2 |
| v4 | Public bootnet + codesigned daemon | 4-6 wks | After v3 |
| v5 | Sandboxing + Sybil resistance | 6-10 wks | After v4 |
| v6 | Sharded inference for frontier models (Exo / Prima.cpp patterns) | 8-12 wks | After v5 |
v1 is a useful product on its own — a free, local, persistent AI agent with an OpenAI-compatible API and an MCP server. The mesh comes later.
- Open foundation models (Llama 3.1, Qwen3, DeepSeek V3, Mistral) have closed the quality gap with closed cloud APIs
- The frameworks for distributed inference, efficient serving, mesh transport, and persistent agent memory all exist as mature OSS — OCM is mostly orchestration
- MCP and OpenAI-compatible APIs let any client (Claude Code, Cursor, Cline, Continue.dev, ChatGPT Connectors, Gemini, etc.) plug in immediately
- Consumer hardware — especially Apple Silicon Mac Mini class — is now genuinely capable of useful agent workloads
- Stacking efficiency multipliers (Q4 quantization × continuous batching × RadixAttention prefix-caching × EAGLE-3 speculative decoding × Qwen3-30B-A3B MoE) gets ~20-30× aggregate throughput per node, ~80× quality-adjusted, vs naive single-stream FP16 serving. Fewer machines deliver more aggregate compute than the naive math suggests.
OCM is a hard project, and we want to be honest about that.
- Petals — the closest prior art for "BitTorrent-style LLM inference" — is functionally dead as of 2026. Working tech, never crossed escape velocity.
- Folding@home only mobilized millions during a galvanizing event (COVID) and lost 99% of capacity post-event.
- 60% of solo OSS maintainers are unpaid; 60% are considering quitting; 44% cite burnout (Tidelift 2024 survey).
- Median GitHub Sponsors monthly income: $50. Caleb Porzio's $100K/yr is exceptional.
- The realistic best-case OSS exit is the llama.cpp pattern — acquired by Hugging Face (Feb 2026) with team retained and full autonomy. Not IPO.
OCM-as-described is a 24-month bet on mission + learning + optionality. The realistic financial outcome distribution:
| Outcome | Probability |
|---|---|
| Burnout / archive at month 14 | 50-60% |
| Niche subsistence ($5-10K/mo blended income, real users) | 25-30% |
| Acquired/folded with team retained (llama.cpp pattern) | 10-15% |
| Foundation / open-core breakout ($200K-$2M ARR, small team) | 5-10% |
If you're considering contributing or running a node, do so because you believe in the mission. Not because you expect financial return.
Apache 2.0 — see LICENSE. Choose your own license for derivative works as Apache permits.
| Crate | Role | State |
|---|---|---|
ocm-daemon |
Tauri desktop shell — system tray, settings, app paths, supervises subprocesses, hosts the API server. Tauri commands for settings (get/save) and model downloads (list/download). | Built |
ocm-api |
OpenAI-compat HTTP server (/v1/models, /v1/chat/completions with SSE streaming, /v1/registry/models). Library-driven Mem0 retrieval before every chat turn. Localhost-only auth middleware. |
Built |
ocm-inference |
InferenceBackend trait + llama.cpp + vLLM adapters + supervisor. Auto-selects backend by platform. |
Built |
ocm-memory |
Mem0 client (search before generation, persist after). | Built |
ocm-mcp |
MCP stdio JSON-RPC bridge. Lets Claude Code / Cursor / Cline / Continue.dev connect via standard MCP. | Built |
ocm-models |
Curated registry (3 GGUFs across tiny / default tiers, all SHA256-verified; canonical tier returns in v0.1.1) + streaming verified downloader. Refuses unverified hashes. | Built |
ocm-mesh |
Mesh transport trait + iroh / libp2p stubs. Real implementations land in v2. | Scaffolded |
| Route | What | State |
|---|---|---|
/ |
Chat UI with SSE streaming, model picker, abort-to-stop | Built |
/models/ |
Browse the curated registry, download with status feedback (Tauri-only) | Built |
/settings/ |
Edit all 7 settings.toml fields, save via Tauri command | Built |
Built with Tailwind v4, TypeScript strict mode, svelte-check 0 errors / 0 warnings.
Hypothesis-based sandbox runner. First isolation sandbox (vllm-q4-llama8b on RTX 4090) wired with expected.json. Multipass-fleet mesh-discovery sandbox stub committed for v2 activation.
.github/workflows/release.yml builds .dmg (mac aarch64 + x86_64), .msi (windows), .deb + .AppImage (linux) on tag push. Drafts a GitHub Release with auto-generated notes. Codesigning + auto-update deferred to v4 per spec. See docs/release-process.md.
Everything in docs/superpowers/ informed the architecture:
specs/2026-05-08-ocm-v1-design-spec.md— v0.5, 31 locked decisions across architecture, models, optimizations, privacy, deployment policyplans/2026-05-08-ocm-v1-implementation-plan.md— task-by-task implementation planresearch/— eight research notes covering per-node efficiency, mesh, memory, market viability, Memory Palace pattern, effective-context triad, encryption + compression, and remote-node deployment
- Real iroh/libp2p mesh transport (v2)
- Reciprocity ledger (v3)
- Codesigning + auto-update (v4)
- Sandboxing + Sybil resistance (v5)
- Sharded inference (v6)
- Qwen3 canonical-tier registry entries (dropped pending verified hashes; return in v0.1.1 — the shipping 3-model registry is fully SHA256-verified)
- Live deployment on ocm.shortcircuit.bot — domain reserved, site not yet up
Code contributions are welcome. Per-area:
- Daemon / API — Rust 1.78+,
cargo test --workspaceruns the suite locally. CI verifies fmt + clippy + tests across ubuntu / macos / windows. - Frontend — Node 20+,
cd frontend && npm install && npm run devfor HMR.npm run checkfor svelte-check + tsc. - Spec / research — critical reading is still a high-leverage contribution. Flag where the analysis is wrong, the assumptions are off, or the architecture has a hidden flaw.
File issues at github.com/OpenCircuitDev/opencircuitmodel/issues.
OCM stands on the shoulders of every project named in the research synthesis: vLLM, SGLang, llama.cpp, MLX-LM, MLC-LLM, Mem0, Smolagents, DSPy, libp2p, iroh, Exo, Prima.cpp, Hivemind, Letta, Outlines, Instructor, Tauri, the MCP working group, and the Folding@home / BitTorrent / IPFS / Storj / Filecoin / Tor projects whose patterns inform the mesh and reciprocity design.