Agenta Agent Workspace
Build conversational agents, share them with a team, and run recurring work in the background.
- Source repo
- Agenta-AI/agenta
- Stars
- ★ 4.9k
- Last updated
- today
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 58/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude Code · OpenAI API · Claude API
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A knowledge-work team that regularly organizes research or maintains a shared wiki can use a background agent in the shared workspace.
- Not a fit if
- Teams that need fully offline operation with no data leaving their machines
- Users who want drag-and-drop workflow builders for fixed, predefined steps
- Teams unwilling to deploy and maintain the service themselves
- Source review
- 58/100 · Major gaps
What does this agent do, and when should you use it?
Agenta is an open-source workspace for building and running specialized agents. Users describe work in chat, connect the applications an agent needs, and improve the agent through feedback. It combines shared file workspaces, role-based team access, scheduled background work, and triggers from connected-app events. The platform traces model and tool calls and retains agent-configuration history for failure analysis, change comparison, usage tracking, and estimated-cost review. Its self-hosted path is described as usable with existing Claude or ChatGPT subscriptions, and the currently supported harnesses are Claude Code and Pi.
A user describes a task in chat, connects applications, and works with an agent on files, documents, research, or a wiki in a shared workspace. Agenta can run the agent interactively in chat, on a schedule, or after an event in a connected app. It connects applications through MCP servers and, through Composio, lists access to more than 1,000 apps including Gmail, Slack, Notion, and GitHub. It traces every model and tool call, tracks token usage and estimated costs, and retains version history for each agent configuration. Agent definitions may use AGENTS.md, skills, and MCP servers; the README also lists Local runtime, Daytona sandboxes, and Docker sandboxes.
- A knowledge-work team that regularly organizes research or maintains a shared wiki can use a background agent in the shared workspace.
- An individual or team with an existing Claude or ChatGPT subscription that wants to self-host local agent work can use Agenta as the operating workspace.
- A team that needs work to begin after events in connected Gmail, Slack, Notion, or GitHub can use event-triggered background agents.
- An engineering or operations team investigating agent failures, model requests, tool calls, token usage, or estimated cost can inspect traces and configuration history.
- A team that wants one agent to collaborate in chat and also perform recurring work can combine the interactive and scheduled execution modes.
How do you install or deploy this agent?
The supplied README does not provide a standalone, reproducible self-host installation command, runtime version, environment variables, ports, or credential list. Its documented bootstrap route is npx skills add Agenta-AI/agenta-skills, followed by asking a skills-capable agent to self-host and test Agenta; the supplied material does not specify the configuration needed to complete deployment.
How do you use this agent?
Describe the work to automate or augment in chat, connect the applications the agent needs, and improve it through feedback. Work with the agent directly in the shared workspace, schedule recurring work, or have it run when an event occurs in a connected application. Set tool permissions to distinguish actions that may run automatically, require human approval, or are blocked; then use traces, usage, costs, and configuration history to review results.
What are this agent's strengths and limitations?
- Combines interactive chat, scheduled execution, and connected-app event triggers in one workspace rather than only predefined workflow steps.
- Traces every model and tool call while preserving configuration history, enabling failure investigation, change comparison, and token and estimated-cost review.
- Supports Claude Code and Pi harnesses and lists model paths including OpenAI, Anthropic, OpenRouter, and Ollama.
- Provides shared file workspaces, role-based team access, and per-tool human approval or blocking controls.
- The supplied material lacks complete self-host installation commands, environment variables, ports, credentials, and runtime prerequisites, so deployment still requires further documentation.
- Only Claude Code and Pi are marked as completed harnesses; Codex, Gemini, and OpenCode remain roadmap items.
- Several runtimes and trigger capabilities remain planned, including E2B, Vercel, Cloudflare, generic webhook triggers, and OAuth MCP transport.
- Repository metadata says NOASSERTION for the license while the README says MIT; adopters should verify the governing license files and terms.
How does this agent compare with similar options?
The README distinguishes Agenta from n8n, Activepieces, and Zapier by positioning those products around predictable, predefined workflows and Agenta around work where an agent plans, uses tools, and adapts. Against Claude Cowork, it emphasizes open source, selectable harnesses and models, inspectable agent components, and both interactive and background execution. Against Claude Code, Codex, Pi, and OpenCode, Agenta is positioned as the shared workspace around the execution layer, adding files, team access, triggers, versions, and traces.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agenta Agent Workspace This agent | 58 · Major gaps | Self-hosted serviceFreemium | ★ 4.9k | today | TypeScript | Claude Code · OpenAI API · Claude API |
| Okou | 54 · Major gaps | Hosted serviceFreemium | ★ 1.2k | today | TypeScript | OpenAI API · Claude API |
| CowAgent | 52 · Major gaps | CLIFree + model costs | ★ 48k | today | Python | OpenAI API · Claude API |
| holaOS | 42 · Major gaps | Desktop appFreemium | ★ 11k | 2d ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The entrypoint shows request credentials being passed through and explicitly holds service-process model and AWS keys away from requests; the README also claims sandbox isolation and configurable tool permissions. Least privilege and sensitive-data handling therefore have some code and documentation support, but the sandbox and permission claims lack implementation detail in this packet. There is no evidence of general confirmation or rollback for external actions. The security reporting route and split license terms are clear, while dependency-security evidence is limited and publisher identity and maintenance/update ownership are not verified.
The service entrypoint routes invoke and inspect exceptions to handlers and exposes a health check, but the provided code does not show how those handlers present errors. The announcement skill has an inconsistency between its file-format description and array example, and some tutorials rely on external services. Evidence for dependency availability and end-to-end consistency is limited, so scores remain thin.
The README addresses technical and non-technical team members and names chat, scheduled or event-triggered automation, self-hosting, model choice, and integrations. Skills state their triggers and scope. The README offers cloud and self-hosted entry points, but delegates concrete installation steps to external skills and documentation, so environment guidance is useful but incomplete.
The repository map and layered guidance are clear, and skills include names, trigger descriptions, and examples. The security policy describes disclosure, while the license explains MIT coverage and different licensing for EE directories. The README gives a self-hosting entry point, and the release workflow shows coordinated versioning across components. Version/update evidence is limited, and contributor guidance does not fully establish product maintenance responsibility.
The product purpose and scenarios are specific, and the README describes team sharing, tool integrations, and background automation. Skill examples and the entrypoint provide actionable information. The packet lacks execution evidence, detailed costs, and limitations; many capability claims remain product statements, so actual marginal value and cost-benefit cannot be established strongly.
Some security behavior is traceable to the entrypoint, and installation, licensing, and security disclosure have corresponding files. However, key README claims about sandboxing, auditability, and permissions are not corroborated by the code excerpts supplied, and independent cross-source support is limited. Product claims can be identified, but the distinction between fact and inference is not consistently explicit.
- The README says agents can use tools autonomously, work with websites and files, and run on event triggers. Before deployment, verify sandboxing, permissions, credential isolation, and confirmation and recovery paths for external writes.
- Concrete self-hosting steps are delegated to external skills and documentation. This assessment uses only the supplied static material and does not execute or independently verify product safety or reliability.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/agenta)<a href="https://followagents.com/en/agents/agenta"><img src="https://followagents.com/badge/agenta.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * feat(frontend): edit a queue's name and description in /m by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8206 * feat(frontend): show each step's trace and metrics in the /m scenario drawer by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8207 * feat(frontend): compare evaluation runs on the /m run page by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8208 …View on GitHub ↗ -
Release notes excerpt
What's Changed * feat(frontend): queue All Annotations in /m shows trace inputs and outputs by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8174 * feat(frontend): View Results in the /m annotation queue row menu by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8177 * fix(frontend): playground results in a labelled card, compact rows and primary Run all in /m by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8182 * fix(frontend): playground header and toolbars like /w, no overlap in compare mode in /m by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8188 …View on GitHub ↗ -
Release notes excerpt
What's Changed * fix(frontend): send /w app and developer links to the matching /m page under beta by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8133 * feat(frontend): Observability parity in /m (rail entry, readable Pretty JSON, test set queues, auto-refresh) by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8134 * feat(frontend): beta developer mode is the one gate for /m developer views by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8141 * feat(frontend): /m developer mode rail like /w by @mmabrouk in https://github.com/Agenta-AI/agenta/pull/8140 …View on GitHub ↗
FAQ
Can it use an existing subscription in a self-hosted setup?
How do background agents start?
How are risky actions controlled?
What can be inspected when a run fails or costs rise?
Thanks — we use this to improve the page.