Build your own models.

Orchestrate Claude Code, Codex, and your own agents, then turn their work into datasets, evals, and models that continually improve with reinforcement learning.

Open-source Download desktop app

Open web app

iOS app· Android soonSee it work

OpenPond Desktop

Everything stays with the work.

Keep the conversation, tools, files, changes, and result together. Work beside local projects or use an isolated cloud sandbox when you need more room.

Local-firstCloud sandboxesSubagentsHuman review

OpenPond workspace with a conversation beside review, files, browser, and terminal tools

Better the next time around.

Refiner reviews completed tasks and looks for improvements that can help with future work. Useful changes become recorded, testable Harness updates. If no change is justified, nothing changes.

01
Capture the work
Work with local projects or cloud sandboxes. The conversation, tools, files, changes, and result stay together.
02
Learn from the outcome
Refiner reviews completed tasks and looks for reusable improvements that can help with future work.
03
Improve what matters
Update the Harness first. Create Tasksets, run evaluations, and train a model only when the evidence supports it.

OpenPond showing model training metrics and evaluation results

Managed reinforcement learning

Train when the task earns it.

Compare a baseline, inspect reward variation, run bounded GRPO or RFT, and preserve the resulting adapter with its complete recipe. A failed hypothesis is still a useful result.

Explore Tasksets →

One repo. The whole harness.

OpenPond Desktop, CLI, TUI, local runtime, Tasksets, and the training workbench live together. Run locally without an OpenPond account, inspect the contracts, and bring in the web app when the work needs cloud execution or a team.

Read the documentation →

github.com/openpond/openpondDesktop · CLI · TUI · local runtime · training workbench

View repository →

Start from the terminal

npx openpond@latest