Agenta: Self-Hosted Workspace for Building and Running AI Agents

The Problem Agenta Addresses

Building AI agents that are reliable enough for recurring work typically involves assembling separate tools for prompt management, scheduling, tool connections, debugging, and access control. Agenta presents a single workspace that covers all of those concerns. According to the README, the central use case is building agents that automate repetitive work by chatting with them: a user describes the task, connects the applications the agent will need, and refines the agent through feedback rather than writing orchestration code. The result can be shared with teammates and run in the background on a schedule or triggered by an event in a connected application. A key design choice is that when the workspace is self-hosted, it can run agents against an existing Claude or ChatGPT subscription, meaning the marginal cost of running agents does not automatically increase with API calls the way it would with a metered key.

Harnesses, Models, and Runtimes

Agenta separates the agent harness from the underlying model. The README lists Claude Code, Pi, and Codex as currently supported harnesses, with Gemini and OpenCode on the roadmap. For models, the README lists OpenAI, Anthropic, OpenRouter, Mistral AI, Cohere, Anyscale, Perplexity AI, DeepInfra, Together AI, Groq, Google Gemini, Azure, AWS Bedrock, MiniMax, OpenAI-compatible models, and self-hosted models via Ollama. This separation means changing which model an agent calls does not require rebuilding the agent from scratch. For runtimes, the README lists local runtime, Daytona sandboxes, and Docker sandboxes as available, with E2B sandboxes, AgentComputer, Vercel, Cloudflare, Modal, and BoxLite on the roadmap. The definition of an agent uses AGENTS.md, skills, and MCP servers, which the README notes can be brought in from the broader agent ecosystem.

Installing Agenta via the Self-Hosting Skill

The README describes a two-step setup that runs inside an existing agent session rather than through a standalone CLI installer. Paste the following instructions into your agent and it will walk through the process:

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

This approach uses the skills mechanism to fetch the latest self-hosting instructions directly from the Agenta repository, which the README notes keeps the documentation current with the software. The README also points to a self-hosting documentation page at agenta.ai/docs/self-host/quick-start for users who prefer to follow written steps. For teams who want to try Agenta without infrastructure work, the cloud version at cloud.agenta.ai is the alternative.

Background Agents, Scheduling, and Integrations

Agenta distinguishes between interactive agents that the user talks to directly and background agents that run without a live conversation. Background agents start either on a schedule or when an event fires in a connected application. The README lists schedules and events from connected apps as current features, with generic webhook triggers and additional channel support for Slack, Telegram, Discord, and Teams on the roadmap. Tool connections work through MCP servers or through Composio, which the README describes as covering more than 1,000 applications including Gmail, Slack, Notion, and GitHub. Human approval is configurable per tool: the README describes choosing which tool actions a background agent can run automatically, which need approval, and which are blocked. The workspace also maintains a shared file area where the user and agent can work together on documents, research notes, or a wiki.

Tracing, Version History, and What They Enable

Agenta records every model call and tool call in a trace for each agent run. The README states that the workspace keeps a version history of each agent configuration. Together, these two features give a concrete basis for understanding why an agent failed on a specific task, comparing the behaviour of two configurations side by side, and recovering a previous configuration that worked better. The README also tracks token usage and estimated cost per agent, which provides some visibility into how much a given background task is consuming. This kind of observability is absent from a workflow that connects an LLM API directly, where failed runs typically produce only an error response without context about which step failed or what the model received.

Team Access and Where It Falls Short

The README describes role-based access for sharing agents with a team in the open-source version. This is a meaningful feature for organisations that want to give non-engineering colleagues access to specific agents without granting them the ability to edit configurations. The README does not detail what roles exist or what permissions each role carries, so teams with fine-grained access requirements should review the documentation before committing. The roadmap also shows several planned features that are not yet available: OAuth for MCP transport, mobile version, E2B sandboxes, and most deployment targets beyond local runtime and Daytona and Docker sandboxes. Teams whose agents need any of those features will need to wait for them or contribute to the project. The repository's top-level entries reveal the breadth of the codebase: api/, clients/, services/, web/, and sdks/ directories cover the backend API, frontend web application, and multiple language SDKs. The examples/ directory includes JavaScript, Python, TypeScript, Node, and Jupyter notebooks, giving a practical starting point for connecting custom tooling to an Agenta instance. The .agents/ directory at the project root is itself an AGENTS.md-defined configuration, showing that the project follows its own standards.

Cadence, Licence, and Maintenance

Agenta releases frequently. The three most recent releases at the time of writing were v0.121.5 on 2026-09-28, v0.121.4 on 2026-09-27, and v0.121.3 on 2026-09-26. The last push to the repository was on 2026-09-27. The repository is licensed under NOASSERTION according to the GitHub metadata, which is unusual and worth checking against the LICENSE file in the repository before building a commercial product on top of it. The project accepts contributions and publishes a CONTRIBUTING.md and CODE_OF_CONDUCT.md. The repository structure includes an api/ directory, sdks/, services/, web/, and a set of example files in JavaScript, Python, TypeScript, Node, and Jupyter, which collectively give a reasonable starting point for programmatic integration. The hooks/ directory in the repository suggests the project uses lifecycle hooks in its runtime, and the openspec/ directory points to an OpenAPI specification for the backend API. Teams building integrations against the Agenta API surface can use the openspec/ definitions to generate typed clients rather than writing HTTP calls by hand.

Editorial conclusion

Agenta fits teams who already pay for Claude or ChatGPT and want to run agents against that subscription rather than per-token API billing, and who need background scheduling, version history, and role-based access in one place. The open-source version covers the full feature set; the cloud offering at cloud.agenta.ai removes the infrastructure work. Before adopting, verify which harnesses your team uses against the current roadmap, since Gemini and OpenCode support are listed as planned but not yet shipped as of the last push on 2026-09-27.

Frequently asked questions

Official sources

1. Agenta-AI/agenta on GitHub
2. Issues
3. Project website
4. README
5. Releases