Agenta is an open-source workspace for building and running agents that automate work. You build an agent by chatting with it: describe the job, connect the applications it needs, and refine it through feedback. Agents can then be shared with a team, worked with directly in chat, or run in the background on a schedule or when an event fires in a connected app.
The agent definition uses open standards — AGENTS.md, skills, and MCP servers — so skills and servers from the wider ecosystem carry over. The execution layer is swappable: harness and model can be changed without rebuilding the agent. Self-hosting also lets agents run locally against an existing Claude or ChatGPT subscription rather than metered API billing.
Agenta positions itself around the workspace rather than the execution layer. Coding agents supply planning and tool use; Agenta adds the files, team access, triggers, versions, and traces around them. You and the agent share a workspace of files — documents, research, a wiki — and every run is traced with model and tool calls, token usage, and estimated cost. Each agent configuration keeps a version history for comparing changes after a failure. Per-tool permissions decide which actions a background agent may run on its own, which need approval, and which are blocked.
Compared with workflow builders such as n8n, Activepieces, and Zapier, which run predefined steps, Agenta targets work where the agent has to plan, use tools, and adapt — and where the same agent should be usable both interactively and in the background. It is available as a hosted cloud service or self-hosted, with a skill that walks an agent through the setup.
Features
- Chat-built agents: describe the work, connect apps, and improve the agent through feedback rather than code
- Swappable harnesses: Claude Code, Pi, and Codex are supported today, with more on the roadmap
- Broad model support: hosted providers, gateways, and self-hosted models, switchable without rebuilding the agent
- Open agent standards: AGENTS.md, skills, and MCP servers define the agent and port in from the ecosystem
- Shared workspace: you and the agent work on the same files, documents, and research
- Background agents: run on a schedule or trigger on an event in a connected app
- Permissions and approval: per-tool policy for what runs automatically, what needs approval, and what is blocked
- Tracing and cost: every model and tool call is inspectable, with token usage and estimated cost per agent
- Version history: each agent configuration is versioned so failures can be compared against earlier states
- Team access: share agents and control access by role in the open-source version
- Integrations: connect through MCP, or reach more than 1,000 apps through Composio
- Agent runtimes: local, Daytona sandboxes, and Docker sandboxes, with more planned
