Agenta is an open-source workspace where a team builds, runs, and improves AI agents by chatting with them, intended for individuals and teams that want agent automation without moving every task onto metered API billing.

What it is

Agenta is a TypeScript workspace, published by Agenta-AI under a licence GitHub reports as NOASSERTION, in which agents are built by conversation rather than by writing configuration from scratch. A user describes the work, connects the applications the agent needs, and refines the agent through feedback. Agents can be used directly in chat or shared with a team, and recurring work can be handed to background agents that run on a schedule or when an event fires in a connected app. Agents are defined with open agent standards: AGENTS.md, skills, and MCP servers, so skills and MCP servers already present in the agent ecosystem can be brought into Agenta.

The concrete problem it solves is harness and model lock-in, plus the cost and opacity that come with hosted agent products. Agenta lets a team switch harnesses and models without rebuilding an agent, and it traces every run while keeping a version history of each agent configuration. Instead of rewriting an automation when a model changes or paying per-token for every routine task, the same agent definition is pointed at a different harness or model. It supports Claude Code, Pi, and Codex as harnesses today, with Gemini and OpenCode listed as unshipped on the roadmap.

Key capabilities

• Harness and model swapping: Claude Code, Pi, and Codex are supported as harnesses, alongside OpenAI, Anthropic, OpenRouter, Mistral AI, Cohere, Anyscale, Perplexity AI, DeepInfra, Together AI, Groq, Google Gemini, Azure, AWS Bedrock, MiniMax, OpenAI-compatible endpoints, and self-hosted models through Ollama.
• Open agent standards: agent definitions use AGENTS.md, skills, and MCP servers, with skills and MCP servers imported from the wider ecosystem.
• Agent runtimes: a local runtime plus Daytona sandboxes and Docker sandboxes, with E2B, AgentComputer, Vercel, Cloudflare, Modal, and BoxLite on the roadmap.
• Tracing, usage, and cost: every model and tool call can be inspected, with tracking of model requests, token usage, and estimated costs per agent.
• Human approval and permissions: permissions are set per tool, allowing a team to decide which actions background agents run automatically, which require approval, and which are blocked.
• Background agents: agents run on a schedule or start when an event occurs in a connected app.
• Integrations: connection to applications through MCP, or to more than 1,000 apps through Composio, including Gmail, Slack, Notion, and GitHub.
• Shared workspaces: agents and people work together on files in a shared workspace, and the open-source version allows agents to be shared with a team with access controlled by role.

Who uses it and how

• A team that already holds Claude or ChatGPT subscriptions and self-hosts Agenta so that agents run locally against those subscriptions instead of moving every task to metered API billing.
• A team that maintains several agents across different models and needs to compare configurations and diagnose failures from run traces and version history.
• A team running recurring automations, such as scheduled jobs or agents triggered by events in connected apps such as Gmail, Slack, Notion, or GitHub.
• A team that keeps some agent actions behind human approval, setting per-tool permissions so that some actions run automatically, some wait for approval, and some are blocked.
• A group using agents to write documents, organize research, or maintain a wiki inside a shared workspace.

Getting started

Agenta Cloud is described as the fastest way to try it. To self-host, the documented path is to run npx skills add Agenta-AI/agenta-skills and then have an agent walk through setup and testing using the repository, with further detail in the self-hosting documentation.

How it compares

No list of paid products that Agenta replaces is supplied here, and the facts name no directly comparable agent workspaces, so on this registry it stands alone as an entry. The adjacent tools the facts do name — Claude Code, Pi, and Codex — are harnesses Agenta wraps and switches between rather than projects it competes with.

When to use it — and when not to

A self-hoster must operate the repository and its runtime, choose between the local runtime, Daytona sandboxes, or Docker sandboxes, and manage MCP credentials and integrations themselves, so the setup is not a hosted turnkey experience. Teams that need Slack, Telegram, Discord, or Teams channels, generic webhook triggers, OAuth-based MCP transports, or a mobile client should wait, since each is listed as unshipped. Two further caveats are worth weighing: the licence is reported as NOASSERTION rather than a named open-source licence, and the repository carries 309 open issues.

1. Install the Agenta self-hosting skill: npx skills add Agenta-AI/agenta-skills
2. Help me self-host Agenta with its repository.