Agenta

Overview

Agenta is an open-source workspace for building, testing and sharing AI agents. You start from a working template for engineering, support, sales, knowledge or operations, then customize it for your own use case. Build agents through chat with your apps and files, then automate them with one more message — schedule agents to run on a recurring basis.

Agenta is built on open standards: your agent is defined by AGENTS.md, skills and MCP tools, so you can swap the harness, model or runtime without touching a line. It supports Claude Code, Codex, pi.dev, OpenCode and Hermes harnesses, and providers including Anthropic, OpenAI, Ollama, xAI and OpenRouter. Self-host on your own infrastructure with the MIT-licensed code.

Reliability features include human-in-the-loop approval for consequential actions, versioned prompts/skills/tools like code, and per-run traces to find exact failure points. Feedback becomes test cases so every change is evaluated before it ships.

Core capabilities

Chat-first agent building

Build custom agents that share your context and integrations

Automation

Schedule agents with one more message, like 'every Monday at 9:00'

Open standards

Agent is AGENTS.md, skills and MCP tools — swap harness, model or runtime freely

Template library

Start from engineering, support, sales, knowledge or operations templates

Human in the loop

Consequential actions wait for approval, edit or reject before shipping

Versioning

Prompts, skills and tools versioned like code, roll back anytime

Run tracing

See each step, its cost and outcome, and find exact failure points

Self-hostable

MIT-licensed, run on your own infrastructure with Docker Compose or Helm

Multi-harness support

Claude Code, Codex, pi.dev, OpenCode and Hermes

Use cases

Code review agents

Customer support agents

Sales automation

Company knowledge agents

Operations workflows