OpenAgentsInc/openagents

Monorepo & docs

What it solves

OpenAgents creates a collaborative ecosystem for developing and improving coding agents. It addresses the gap between agent development and real-world verification by introducing a "Gym" where users (trainers) can create, test, and verify plugins that measurably improve an agent's performance on specific tasks.

How it works

The project is built around a "loop" of continuous improvement:

  1. Chat & Create: Users interact with the agent (Coder) via a mobile or desktop app to draft plugins (containing skills, workflows, or knowledge) and associated test sets.
  2. Execution: Tests are run with and without the plugin on the user's connected computer or a hosted runner to determine if the plugin actually improves performance.
  3. Verification: Results are published to the Gym. Other users can verify these results by rerunning the tests, earning XP for confirmed improvements.
  4. Adoption: When a plugin is proven effective, Coder adopts it for all users.

The system uses the Nostr protocol for communication between phones, computers, and agents, and employs a Wasm-based plugin architecture for extensibility.

Who it’s for

  • Developers who want to use a high-performance coding agent that evolves based on community-verified improvements.
  • AI Trainers who want to contribute to agent capabilities through plugin development and evaluation.
  • Playtesters who want to participate in an open-source agent collective.

Highlights

  • Agent Collective: A network-effect driven approach to agent improvement where community contributions are tracked via XP.
  • Coder: A specialized coding agent capable of running code on local machines or in the cloud.
  • The Gym: A dedicated environment for running evals and benchmarks (like Terminal-Bench) to measure agent progress.
  • Cross-Platform Support: Native apps for iOS, Android, and Desktop (macOS, Linux, Windows).
  • Nostr Integration: Uses decentralized protocol primitives for agent trajectories and communication.

Related

  • Project
    ConardLi/easy-agent

    Easy Agent is a terminal‑based AI coding assistant that can read, edit, and run code under configurable permission rules and optional OS sandboxing. It supports major LLM providers, offers persistent sessions, extensibility via plugins/skills, and multiple safety modes (plan, auto, default). Install via npm, trust a workspace, then interact with the REPL to ask questions, make multi‑file changes, run tests, and undo actions.

  • Project
    Nano-Collective/nanocoder

    An open-source, local-first coding agent for the terminal that lets developers use their own AI models to write and refactor code while keeping data private.

  • Project
    shareAI-lab/Kode-CLI

    A terminal-based AI coding assistant that integrates codebase understanding, file editing, and command execution with support for multi-model collaboration and sub-agents.

  • Project
    opendev-to/opendev

    An open-source, Rust-based coding agent that uses a compound AI system to orchestrate parallel agents and mix multiple LLM providers per workflow.

  • Project
    Kuberwastaken/claurst

    Claurst is a Rust‑based terminal coding assistant that talks to multiple LLM providers, offers a rich TUI, supports the open Agent Client Protocol for editor integration, and includes experimental features like multi‑turn goals and an ultra‑think mode.