Agenta is an open-source, self-hostable workspace where a team builds specialized agents by chatting with them: you describe the work, connect the apps the agent needs, and improve it through feedback. Agents are shared, multi-user objects with roles and permissions, and the team works with them in the web app or from Slack, Telegram, or WhatsApp Business. Automations run an agent on a schedule or when an event occurs in a connected app, and every run is traced with its cost.
Agenta positions itself against three neighbors. It is not an agentic development environment for writing code like Claude Code or OpenCode: it targets the work around code (IT ops, go-to-market, HR, legal, finance, support). It is not a single-user personal assistant like OpenClaw or Hermes: agents are shared and run where the team works. And it is not a low-code workflow builder like n8n, Activepieces, or Zapier: you do not draw steps, the agent plans them. Coding agents such as Claude Code, Pi, and Codex are the execution layer underneath (supported today, with more harnesses on the roadmap), and Agenta adds files, team access, triggers, versions, and traces around them. Compared with Claude Cowork, it is open source and lets you choose harness and model.
Each session runs in its own sandbox, and model credentials never reach the agent. Agents keep memory, create and refine their own skills, and update their own instructions. Each agent has a browser and a file system of its own, connects to tools through MCP or to more than 1,000 apps through Composio (Gmail, Slack, Notion, GitHub), and can call other agents to hand off work. Models run on a Claude or ChatGPT subscription, a hosted API, or self-hosted models with Ollama, with per-agent cost visibility. A shared workspace holds documents, research, and a wiki with comments and previews, and agents can build small apps such as a board or a form for the team.
Features
- Build by chatting: describe the work, connect apps, and refine the agent through feedback, for technical and non-technical teammates
- Chat channels: agents answer in Slack, Telegram, and WhatsApp Business, and the web app works on a phone
- Multiplayer: shared agents and chat sessions, roles, and access for teammates or customers
- Automations: runs on a schedule or on an event in a connected app, with run history and the option to continue a run's session
- Permissions and sandboxing: per-agent tool permissions with read or write scope, approval steps you define, and an isolated sandbox per session that never sees model credentials
- Tracing and cost: every run traced and every cost tracked, with agent configuration, versions, and change history visible
- Models: a Claude or ChatGPT subscription, hosted models through an API, or self-hosted models with Ollama, chosen per agent
- Harnesses: Claude Code, Pi, and Codex as the execution layer, with more on the roadmap
- Integrations: MCP servers, or more than 1,000 apps through Composio including Gmail, Slack, Notion, and GitHub, with user-supplied secrets
- Agent computer: a browser and a file system per agent for websites and files without a dedicated integration
- Learning agents and teams: memory, self-authored skills and instructions, and agents that call other agents to coordinate specialists
- Shared workspace: documents, research, and a wiki with comments and previews, plus small agent-built apps like a board or a form
