Web
Chat with people and agents, open shared projects, and run the same hosted workflows from any browser.
The open-source agent harness for teams.
Build and ship agents, evaluate models, and prepare training runs from real conversations. Share the same agents, projects, sandboxes, and usage with your team across every surface.
Built for multiplayer from the start.
People and agents share one workspace without sharing Personal accounts. Team Chat, agents, projects, sandboxes, Tasksets, and cloud usage can all belong to the Team.
Team Chat
People and agents share channels, direct messages, project context, and the same execution history.
Shared workspace
Agents, projects, sandboxes, and Tasksets belong to the Team instead of one developer's Personal account.
Usage controls
One Team funds cloud execution while limits, runs, and audit remain attributable to each member.
One harness, every surface.
Web and Desktop are two views of the same runtime. The CLI, TUI, connected channels, and hosted agents use the same source-backed packages and execution model.
Desktop
Keep local files, source review, model settings, approvals, and terminal state close to the conversation.
CLI and TUI
Use the same runtime from scripts, terminals, CI, or an interactive terminal chat without inventing another agent contract.
Connected channels
Publish the same source-backed agents to Team Chat, Slack, Microsoft Teams, or another machine.
Agents are source packages, not saved prompts.
Use the Agent SDK when a workflow needs code, actions, setup requirements, or evals. Packages live in a git-backed profile and remain discoverable from Web, Desktop, CLI, TUI, and cloud execution.
Local control. Cloud execution.
Keep orchestration and review close to the developer while giving execution a clean environment built for long-running work and teammate handoff.
Local control
Keep chat, model settings, approvals, subagent placement, local files, and source review on your machine.
No OpenPond account is required for local work.
Cloud execution
Hosted sandboxes handle commands, dependencies, browser work, scheduled jobs, snapshots, and long-running processes.
The agent package and review workflow stay the same.
Web / Desktop / CLI / TUI ↔ git-backed agent source ↔ OpenPond Cloud sandbox
Use conversations as evidence, not raw training data.
Select the work worth repeating, turn it into an explicit Taskset, and inspect every verifier and grade before preparing a portable training bundle.
Select
Choose conversations that demonstrate a repeatable job.
Author
Turn selected evidence into explicit Taskset inputs and outcomes.
Verify
Add deterministic checks or calibrated model graders.
Baseline
Run attempts and inspect the grading evidence.
Prepare
Build a portable bundle for a configured destination.
Start from the implementation.
The app, runtime packages, clients, Agent SDK, and training workbench are public. Run locally, inspect the contracts, and keep the agents and Tasksets you create as ordinary source.
openpond/openpondWeb, Desktop, CLI, TUI, local server, runtime packages, and the training workbench.openpond/openpond-agent-sdkThe SDK for authoring agents, actions, workflows, evals, traces, and runtime artifacts.openpond/openpond-codeThe TypeScript client and command-line surface for projects, sandboxes, apps, and agents.