Agenta ships an agent workspace under the MIT license
The repo bundles background agents, per-tool approval rules and tracing of model calls — run it yourself or use the hosted Agenta Cloud.
Symbolic image: at a team desk, someone confirms a tool approval while monitors show abstract flow diagrams and a small rack beside the table blinks green and amber.
Agenta is an MIT-licensed open-source workspace where teams build agents, run them in the background, and approve or block each tool an agent may call.
At a glance
- License: MIT; the repository calls Agenta fully open source.
- GitHub counters on September 6, 2026: roughly 4,700 stars, 661 forks, 86 contributors and 27,293 commits on the main branch.
- Open on GitHub: 189 issues and 127 pull requests.
- Model providers per the README include OpenAI, Anthropic, Google Gemini, Mistral AI, Groq and AWS Bedrock, plus local models via Ollama.
- Tools connect through MCP servers and Composio, reaching 1,000+ applications per the README.
Agenta is an MIT-licensed open-source workspace where teams build agents, run them in the background, and approve or block each tool an agent may call.
The repository now bills itself as the open-source workspace for building and running agents. An earlier line — a workspace where you and your team build agents and automations — is what surfaced this story. The current wording is tighter, and it points the same way.
Permissions are the interesting part
Plenty of projects will connect an agent to a tool. Agenta leads with the opposite question: what should an unattended agent be allowed to touch? Teams set permissions per tool and sort every action into one of three buckets — run automatically, wait for a human, or blocked outright.
What the README covers
Agents can be developed and tested interactively, then shared across a team and moved into background runs. Those runs fire on a schedule or when an event lands in a connected app. Model and tool calls are traced, and configurations carry version history.
Models and hosting
Listed providers include OpenAI, Anthropic, Google Gemini, Mistral AI, Cohere, Groq, Azure and AWS Bedrock, plus local models through Ollama. Tools arrive over MCP servers, with Composio named as the route to 1,000+ applications. A hosted option, Agenta Cloud, sits alongside self-hosting — and the self-hosting path in the README is an installable agent skill rather than a docker command.
How big is the project
On September 6, 2026 the project page showed roughly 4,700 stars, 661 forks and 86 contributors across 27,293 commits on the main branch, with 189 issues and 127 pull requests open. Those are activity signals, not adoption figures. Nothing on the page indicates how many teams run Agenta in production.
What we could not verify
Exactly one source was available for this story: the project's own page. That falls short of our two-domain rule, and no outside reporting corroborates any of it. We also could not confirm the current release number and its date, Agenta Cloud pricing or free tiers, or whether every component really is MIT-licensed rather than split against a separate enterprise edition. Until someone outside checks the claims, the feature list is a vendor's account of its own software.
FAQ
Is Agenta free to use?
The code is MIT-licensed according to the project page, so self-hosting carries no license fee. The page states no pricing for the hosted Agenta Cloud, so that part is unconfirmed.
Which models does Agenta work with?
The README names OpenAI, Anthropic, Google Gemini, Mistral AI, Cohere, Groq, Azure and AWS Bedrock among others, and supports local models through Ollama.
How does Agenta keep an agent from doing something risky?
Permissions are set per tool. Each action either runs automatically, waits for human approval, or is blocked, and background agents inherit those limits.
