Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

OPENRIG on YouTube: 18 months running a few hundred AI agents, and the coordination failures they fall into.

Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

OpenRig 소개 코딩 에이전트를 여러 개 함께 쓰다 보면 에이전트보다 세션 관리가 먼저 문제가 됩니다. 터미널 탭을 하나씩 늘려 가며 오케스트레이터, 리뷰어, 리서치 에이전트를 띄우고 tmux로 서로 연결해 두면, 며칠 뒤에는 어느 세션이 어떤 역할이었는지, 작업 디렉토리가 어디였는지를 기억하기 어려워집니다. OpenRig 개발자 Mike Schwarz는 이 상태에서 컴퓨터를 재부팅하면 에이전트 구성을 되살리는 데 30분 정도가 걸렸고, 존재를 잊어버린 에이전트는 그대로 잃었다고 설명합니다. 그래서 재부팅을 미루게 되는 상황이 OpenRig를 만든 출발점입니다. 이번에 소개하는 **OpenRig는 Claude Code와 [Codex](https://github....

Multi-agent harness that runs Claude Code and Codex together as one system - mvschwarz/openrig

YouTube video by OPENRIG

この3日でスターと手間が集まったのは、エージェントに書かせた後の3つの層だった。記憶のHindsight、束ねるopenrig、実機やファザーで確かめるGoogleとSonnet 5.5の実測。どれも伸びてるけど、人の目なしでは回ってないんだよね。 エージェントにコードを書かせるのはもう当たり前になったよね。で、この3日（9/27〜29）スターと手間が集まってたのは、その「後」の部分だった。書かせた後に何を覚えさせるか、何体をどう束ねるか、出てきた物をどう確かめるか。ハルはリポジトリ、公式ブログ、実測記事を片っ端から読み込んでみた。そしたら、3つともまだ放っておけば回る状態にはなってなかったんだよね。覚える係：Hindsightが2日連続トレンド1位、45日でスター倍まずは記憶の話から。エージェントって、会話が終わると基本ぜんぶ忘れちゃうじゃない？ 昨日教えたプロジェクトの決まりごとも、次の日にはまた一から説明しなおし。この穴を埋めるエージェント用の記憶システムが、vectorize-io の Hindsight だよ。agents-radar の日次集計だと、9/28と9/29の2日続けて GitHub の AI トレンド1位。1日のスター増は9/28が+4,520、9/29が+4,413だった。毎日4千人以上がブックマークしてる計算だね 👀伸び方がすごく急なんだよね。開...

_This is a submission for Weekend Challenge: Generosity Edition_ ## What I Built I grew up teaching myself to code in Pakistan on hardware that struggled to run modern dev tools. GPU hours on Lambda Labs cost $0.50–$3/hr. For a student trying to fine-tune a model or run a training job, that's not accessible. OpenRig is a peer-to-peer compute sharing platform. Donors register their idle hardware and run a lightweight agent binary. Recipients describe what they need in plain English, get matched to the best available rig via Google Gemini AI, and receive SSH access — time-limited, isolated, and without ever touching the donor's actual machine. No money changes hands. You're donating the thing that actually matters: compute. ## Demo Here is the video link Full flow: 1. Donor registers their rig → copies agent token → runs the agent binary 2. Recipient browses available hardware → submits a request explaining why they need it 3. Gemini scores the request 1–10 for social/educational impact 4. Donor sees the impact score and reason → clicks Approve 5. Agent spins up an isolated Docker container with SSH, starts a bore tunnel, reports the public connection command back to the backend 6. Recipient gets an SSH command and password in their dashboard → connects 7. Donor can revoke access at any time → container and tunnel torn down automatically ## Code Here is the code: OpenRig ## How I Built It **The stack** * Backend: Go + Fiber v3, SQLite (GORM) * Frontend: SvelteKit + Tailwind CSS * Auth: Supabase (JWT verified on the backend via JWKS) * AI matching: Google Gemini 3.6 Flash via Google Gen AI Golang sdk * Tunneling: bore (open source TCP tunnel) * Isolation: Docker with hardened security flags **The hardest problem: giving strangers SSH access without exposing the donor's IP** The naive approach — direct SSH — exposes the donor's home IP to whoever connects. That's a non-starter. The solution is a relay architecture. The donor's agent binary runs on their machine and maintains a persistent poll loop against the backend. When a request gets approved, the agent: 1. Spins up an Ubuntu container with SSH on a random local port 2. Waits for SSH to actually be ready (TCP retry loop, not a fixed sleep) 3. Starts a bore tunnel pointing at that local port 4. Parses bore's stdout for the public port 5. Reports `ssh openrig@bore.pub -p XXXXX` and the session password back to the backend The recipient connects to bore.pub. bore forwards TCP to the donor's machine. Docker maps it into the container. The donor's real IP never appears anywhere. **Why bore and not WireGuard** WireGuard is the right production answer but would have taken the whole weekend to set up correctly. bore is one binary, one command, and the public port comes out of stdout. For a 48-hour hackathon, that's the right tradeoff. I already run wg-easy for my homelab — WireGuard is the v2 plan. **Container security** Recipients land inside a resource-capped Docker container, never the host: docker run -d \ --cpus="2" --memory="4g" \ --security-opt no-new-privileges \ --cap-drop ALL \ --cap-add NET_BIND_SERVICE \ ubuntu:24.04 Known limitation: bore and the container are separate processes. If the agent exits unexpectedly, bore dies but the container keeps running until the donor manually revokes. Production fix would be Kata Containers for VM-level isolation — which is actually my existing side project, so v2 writes itself. **Google Gemini for matching** When a recipient submits a request, Gemini 3.6 Flash reads their reason and scores it 1–10 for social or educational impact. The donor sees the score and a one-sentence explanation before deciding. This means donors aren't just picking at random — they can prioritize a student training an NLP model for an underrepresented language over someone running benchmarks. **The agent architecture** The agent is a standalone Go binary donors run in a terminal. No daemon, no system service. It polls the backend every 3 seconds: GET /api/v1/agent/poll?token=xxx → { to_provision [...], to_revoke: [...], to_restart: [...] } `to_restart` handles agent restarts mid-session — if the agent exits and relaunches, it detects running containers via `docker port`, restarts bore against the same local port, and reports the new public port to the backend. Recipients get a fresh SSH command on their next poll. **What I'd do differently with more time** * Replace Docker with Firecracker for proper VM-level isolation * Self-host the bore server for more control * Add session expiry enforcement (the DB field exists, the background worker doesn't yet) * Proper bore process monitoring instead of letting it die silently * Make the agent a proper system service with socket activation ## Prize Categories **Best Use of Google AI** Gemini 1.5 Flash is the matching layer between recipients and donors. Rather than a simple filter by specs, recipients describe their work in plain English and Gemini determines which available rig fits best and how impactful the use case is. Donors make informed decisions based on AI-generated impact scores rather than just first-come-first-served. _Built in ~48 hours on a fresh linux install because previous one broke.Technical debt acknowledged, demo works, shipping it._ And the GPU sharing part is not functional yet because I ran out of time and energy. Another limitation is that the docker container is that the host can see inside the container. So having personal data is not recommended.