User repository on GitHub
SO-101 robot-arm agent workbench: Bun/Effect coordinator, React workbench, Python LeRobot motor owner
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Cookbooks
Robo Harness
An SO-101 workbench where an evaluation model or rules baseline picks bounded joint steps from typed candidate actions: decision runs are spend-capped via the Vercel AI Gateway and drive a real LeRobot arm or mock hardware.
- Stars
- 4
- Language
- TypeScript
- Reading time
- 3 min
- Published by
- Community
- Author
- grmkris
- Last verified
- Sep 2026
About this project
A standalone SO-101 playground: a custom agent loop, Rerun as the main visual workspace, explicit robot tools, and a LeRobot hardware boundary. The lab deployment is live at using the real SO-101 and both Pi cameras, with no operator login. See deployment and measured acceptance. A fresh checkout defaults to mock hardware with synthetic cameras. This machine has a local .env pointing at the Pi; normal development commands here use that configured endpoint. See validation evidence and remaining configuration.
Quoted from the project's README. Read the full README.
Highlights
- The decision runner calls Jev through the AI SDK experimental_evaluate API on Vercel AI Gateway; cumulative spend is capped by ROBO_JEV_BUDGET_USD, default 10.
- Free Gateway credits exclude Jev, and zero data retention is opt-in via ROBO_JEV_ZDR=1 on Pro or Enterprise plans.
- Chat supervision and decision runs never move the arm at the same time; moves are bounded and supervised.
- Rerun is the main workspace, showing robot geometry, frames, camera streams, proposed paths, joint plots, and a capture timeline.
- A fresh checkout defaults to mock hardware; the lab deployment runs the real SO-101 with both Pi cameras behind Tailscale access.
Quickstart
bash
bun install
uv sync --extra dev
bun run build
bun run devWatch out
No license file, so reuse terms are unclear; needs Bun, Python 3.12, uv, and a WebGL2 or WebGPU browser, and real Jev calls require paid Vercel AI Gateway credits.
Questions about this entry
Short answers drawn from the linked project or page.
Supervised throughout. Chat and a decision run never move the arm at the same time, a control lease lasts three seconds and expiry cancels unfinished motion, and a real-arm run has to be started with --execute --supervised with an operator present. The chat path is documented as unable to take over human control.
No. A fresh checkout defaults to mock hardware with synthetic cameras; you only reach the real service by pointing ROBO_IO_URL at the Pi endpoint. Binding is deliberately restricted too: public internet binding is refused and Rerun's viewer and stream bind to loopback, proxied through the application.
Jev is an evaluation model, not a chat provider, and it is reached only through TanStack decide() on Vercel AI Gateway, so you need AI_GATEWAY_API_KEY with paid credits. Free credits exclude it, zero data retention is opt-in via ROBO_JEV_ZDR=1 on Pro or Enterprise plans, and cumulative spend is capped by ROBO_JEV_BUDGET_USD, default 10.
Not really. The arm display uses a mesh-free URDF as a kinematic skeleton, and the README states plainly that it is neither a physics simulator nor a full collision model. Estimated depth stays labeled relative unless the backend explicitly supplies calibrated metric depth, so treat the visualization as inspection rather than planning.
Bun, Python 3.12, uv and a browser with WebGL2 or WebGPU. From a fresh checkout: bun install, uv sync --extra dev, bun run build, bun run dev, then open http://127.0.0.1:5178 or the built workbench on port 8940.
A lease lasts three seconds and must be renewed deliberately while controlling; expiry cancels unfinished motion. Degrees apply to arm joints, percent to the gripper and meters to Cartesian positions, and an accepted move is queued, not reached, so poll the operation for measured completion.
With LeRobot 0.6.0 installed, robo-export takes a recording id and output path and uses LeRobotDataset.create, add_frame, save_episode and finalize. It uploads nothing, resamples the recorded time axis to the declared frame rate, and does not label the recordings as expert demonstrations.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.
Show HN: Verifying robot picks with a decision modelRead the thread on Hacker NewsX: Jev in front of the HeyGen MCP
X: Playwright CLI + Jev vs Playwright MCP
X: A local MCP server that cuts computer use from 23s to 0.8s. Cartwright built Cortex, a local MCP server where…
More Jev resources like this
A Python CLI that connects multiple robots under one contract with an optional Jev stepper: Jev picks among permitted discrete calls while an LLM writes poses and prose and the executor enforces safety gates.
Jev Tools & CLIs#community#python#robotics
Community
rokbenko/quackdPython
rokbenko/quackdPythonA chat-shaped command bar where Jev picks the tool, arguments, confirmation, and reply type each turn, and code calls the MCP servers, so no LLM ever writes the reply text.
Jev Cookbooks#community#typescript#mcp
Community
w3cj/jev-chatTypeScript
w3cj/jev-chatTypeScriptA camera-only MuJoCo quadrotor that uses Jev for tactical maneuvers at 2.5 Hz while classical perception and a 50 Hz reflex layer keep safety outside Jev, clearing a five-station course.
Jev Cookbooks#community#python#robotics
Community
RomanSlack/jev-dronePython
RomanSlack/jev-dronePythonA fine-grained robot control study loop for LIBERO tasks: Jev layers an intent, a contact or motion family, and one of 27 control inputs, while reversible MuJoCo previews evaluate each candidate before execution.
Jev Cookbooks#community#python#robotics
Community
Dimweaker/jev-liberoPython
Dimweaker/jev-liberoPythonA self-hosted AI trading platform that can put Jev in front of strategy and Quick Trade entry orders as an auditable PASS/REJECT gate, with LLM fallback and deterministic bypasses for exits and protective orders.
Jev Cookbooks#community#python#trading
Community
OpenByteInc/QuantDingerPython
OpenByteInc/QuantDingerPythonAn MCP server and agent skill that verifies coding-agent work by driving the real running app and returning pass, fail, or couldn't-tell verdicts with the file:line to fix; its listing frames the approach as Jev-style machine-native runtime perception.
Jev Tools & CLIs#community#typescript#mcp
Community
reticlehq/reticleTypeScript
reticlehq/reticleTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.