Agenta
Agenta is an open-source workspace where technical and non-technical teams build AI agents by chatting instead of coding, then version, evaluate and trace every run, either on Agenta Cloud or self-hosted on their own infrastructure.
What is Agenta?
Agenta calls itself the open-source workspace for your agents. It is published by Agentatech UG (haftungsbeschränkt), a Berlin company, and released under the MIT licence, so the code you self-host is the code that runs Agenta Cloud.
The premise, set out when Agenta 2.0 was announced on 22 July 2026 by co-founder Mahmoud Mabrouk, is that you no longer define every step of a workflow in advance: you describe the work and the agent reasons its way through it. Version 1 was a classic LLMOps platform — prompt management, evaluation, observability — and remains separately documented.
An Agenta agent is a standard, inspectable package: an AGENTS.md instruction file, reusable skills, tools and MCP servers. Three layers stay interchangeable — the harness (Claude Code, Codex, pi.dev, with OpenCode and Hermes announced), the model (Anthropic, OpenAI, Ollama, xAI, OpenRouter or Agenta's own gateway) and the runtime (a local machine through Docker, Agenta Cloud, Daytona, with E2B and Vercel announced). Run locally, an agent can even use an existing Claude or ChatGPT subscription.
Every agent keeps a shared working folder — documents, research, a wiki, files — so context survives from one session to the next, and the workspace is multiplayer from the start: agents, conversations and files are shared with the team. The same agent serves two modes: interactive chat, then background execution on a schedule or on an event in a connected application.
Reliability is the other half of the pitch. Consequential actions can require human approval, every change is versioned, every run is traced with its estimated cost and failure point, and user feedback becomes a test case. Permissions are granular, with per-tool rules separating what runs automatically, what needs approval and what is blocked. The documentation claims more than 1,000 integrations, through native connectors and MCP.
Self-hosting is documented for Docker Compose and for a Kubernetes Helm chart. On the cloud side, Agenta is SOC 2 Type II certified and offers two data regions, the EU (AWS Frankfurt) and the United States (AWS Ohio). Model costs are never included: you bring your own provider credentials.
What it does
- Describe a job in plain language and get a working agent, with no code written
- Start from a business template — code review, customer support, sales, company knowledge, operations — and adapt it
- Give the agent a persistent working folder and connect your team's applications through native integrations or MCP servers
- Version prompts, skills and tools like code, and roll back to any revision
- Evaluate a change on a test set before shipping, with LLM-as-a-judge or your own evaluators
- Trace every run: steps, model and tool calls, latency, tokens, estimated cost and failure point
- Schedule an agent or trigger it on an event, with per-tool permissions and human approval before any consequential action
When to use Agenta / When not to
A quick filter to help you decide if Agenta is the right fit.
When to use Agenta
- Teams already running LLM agents in production that need every change versioned, evaluated on a test set and traced run by run
- Organisations under data residency or sovereignty constraints, who can pick the EU (Frankfurt) or US (Ohio) region or self-host the MIT-licensed platform on their own infrastructure
- Non-technical people who must own their prompts — product, support, sales or operations staff who cannot wait on an engineering ticket to change how an agent behaves
- Companies determined to avoid vendor lock-in, since the harness, the model and the runtime can be swapped without rewriting the agent
- Teams automating recurring work on a schedule or on an event in a connected application, where a consequential action must still pass through human approval
When not to use Agenta
- Anyone making a single model call with no versioning, evaluation or tracing: the publisher itself says to call the model provider directly
- Teams looking for somewhere to host or serve a model — Agenta orchestrates and observes agents, it is not an inference provider
- Buyers expecting model usage to be bundled into the subscription: you bring your own provider credentials and pay the provider separately
- Consumers and individuals after a personal assistant, since the service is stated to be primarily intended for business use
- Users under 18, who are barred by the terms of service, and anyone who needs a mobile app: Agenta is web and API only
How to use Agenta
A typical end-to-end flow, from setup to results.
- Choose your entry point: sign up free on Agenta Cloud with no credit card, or self-host the MIT-licensed platform
- On the cloud, pick your data region — EU or US — at sign-up; the choice is final and cannot be changed later
- Start from a template (Engineering, Customer support, Sales, Company knowledge, Operations) or simply describe what you need in chat
- Give the agent its instructions (AGENTS.md), its skills, its files and access to the tools it needs
- Connect your applications through native integrations or by adding an MCP server
- Pick a harness and a model, or plug in your own provider credentials
- Iterate on real work: read the traces, give feedback, and let the agent propose configuration changes
- Evaluate before shipping: build a test set, run LLM-as-a-judge or your own evaluators, compare side by side in the playground
- Set per-tool permissions and human approval points, then ship and automate — "Do this every Monday at 9:00" is enough to schedule a run, and event triggers are available too
- For the technical route, use the REST API (Authorization: ApiKey), the OpenAPI 3.1 schema or the Python SDK (pip install agenta); self-hosting runs on Docker Compose or a Kubernetes Helm chart
Pros & Cons
Pros
- MIT licence and self-hosting of the same code that runs the cloud: no lock-in, and data can stay on your own infrastructure
- Portability by design — change harness, model or runtime without rewriting the agent
- A contractual no-training commitment written into section 24 of the terms of service, not just into a FAQ answer
- An unusually complete and public compliance file: DPA, named sub-processor list, Standard Contractual Clauses, UK IDTA, SOC 2 Type II and a Trust Center
- Choice of an EU or US data region at sign-up, from a German publisher to whom the GDPR applies directly
- Run-based rather than per-seat pricing: unlimited users, projects, agents and workflows from the first paid tier
- Open to non-technical staff, while keeping native tracing, versioning, estimated cost per run and serious guardrails
Cons
- Model costs are not included: you need your own provider credentials and pay the provider separately, and that bill can exceed the subscription
- The data region is chosen at sign-up and cannot be changed afterwards
- Short trace retention on the lower tiers: one week on Hobby, one month on Pro, three months on Business
- The free Hobby plan is capped at 2 team members and 5,000 runs per month, and simply stops once the quota is reached
- On the cloud, SSO, RBAC and the SOC 2 Type II report start at Business (299 USD per month); audit logs and custom domains are Enterprise-only, self-hosted included
- Support is community-based through GitHub Issues on Hobby, Pro and the open-source edition; priority support and a private Slack Connect channel begin at Business
- No mobile app and no browser extension, no published list of interface languages, HIPAA still in progress, and several announced building blocks marked "Soon" (OpenCode, Hermes, E2B, Vercel)
Pricing & Plans
A permanent free plan is available. The Hobby tier of Agenta Cloud costs 0 USD for ever — a free plan rather than a time-limited trial — and the self-hosted open-source edition is free for ever as well. The lowest paid entry point is the Pro plan at 29.00 USD per month, followed by Business at 299 USD per month and Enterprise on quotation. Billing is based on agent runs rather than seats, so users are unlimited from the first paid tier, and overage is charged at 5 USD per additional 10,000 runs on Pro and Business. Model costs are not included and are paid directly to the provider. Sign-up requires no credit card, and plans can be changed or cancelled at any time from Settings > Billing.
- 0 USD for ever
- 2 team members
- unlimited projects
- unlimited agents and workflows
- 5
- 000 agent runs per month
- 20 evaluations per month
- 1 week of trace retention
- 29 USD per month
- unlimited members
- unlimited schedules and triggers
- 10
- 000 agent runs per month included then 5 USD per additional 10
- 000
- unlimited evaluations
- 1 month of trace retention
- 299 USD per month
- everything in Pro plus team roles and RBAC
- SSO
- the SOC 2 Type II report
- 3 months of trace retention
- priority support and a private Slack Connect channel
- pricing on quotation
- everything in Business plus custom usage and retention
- audit logs
- custom domains
- tailored security and legal terms
- deployment support
- dedicated support and an SLA
- free for ever
- unlimited users and projects
- unlimited agents and workflows
- unlimited schedules and triggers
- RBAC
- SSO
- tracing
- evaluations
- pricing on quotation
- everything in the open-source edition plus audit logs
- custom domains
- deployment and upgrade support
- security and architecture reviews
- priority and dedicated support
- a private Slack Connect channel and an SLA
Data, GDPR & hosting
A consolidated view of how Agenta handles your data.
GDPR overview
GDPR implementation is documented and concrete. Agentatech UG is established in Berlin, so the GDPR applies directly and no Article 27 representative is required or appointed. The privacy policy and terms of service are effective 5 February 2026; the DPA and the named sub-processor list were published 4 February 2026. The policy sets out GDPR and UK GDPR legal bases and grants access, rectification, erasure, restriction, portability and objection rights to EEA, UK, Swiss and Canadian residents, with complaints open to the relevant supervisory authority, the ICO or the PFPDT. Requests go through a hosted Termly DSAR form; a DPO is reachable at team@agenta.ai, with privacy@agenta.ai for privacy matters. The public DPA incorporates the EU Standard Contractual Clauses and the UK IDTA, gives 30 days' notice on sub-processor changes and allows one audit per year. SOC 2 Type II is certified; HIPAA is in progress.
Who owns the data?
Under section 20 of the terms of service, you retain all rights to the Customer Data you and your authorised users submit, and Agentatech UG processes it on your behalf solely to provide the service. The DPA formalises the split: the customer is the controller (or processor), Agenta the processor (or sub-processor). Section 8 adds that Agenta asserts no ownership over user Contributions. Throughout the subscription, data can be exported through the product's own features, the REST API or a support request. Service data — usage analytics, performance metrics, logs, device and billing information — is a separate category, processed by Agenta on its own account under the privacy policy.
Reuse rights
Because you keep the rights to your Customer Data, you may reuse, export and redistribute it freely, without asking Agenta for permission. On its side, Agenta processes account data (email, optional name, organisation name, hashed authentication credentials) and product data (prompts and configurations, test sets and evaluation results, traces, spans and logs, encrypted provider API keys) to deliver the service, authenticate users, fix bugs and develop features, plus anonymised analytics on feature usage, performance and error rates. Three exclusions are explicit: Agenta does not sell your data, does not use it to train, fine-tune or otherwise improve any machine-learning or large-language model — a contractual commitment in section 24 of the terms, not merely a FAQ answer — and does not share it without consent unless legally compelled. By default your content does not travel through the LLM providers on Agenta's behalf: you bring your own keys and requests go from your browser or application straight to the provider. Protection rests on AES-256 encryption at rest, TLS 1.2+ in transit, PGP symmetric encryption for secrets, least-privilege internal access that is logged and audited, and organisation- and project-level isolation with no cross-tenant access.
Data retention & training
Hosting summary
Agenta Cloud runs in two regions, chosen at sign-up: the EU on AWS eu-central-1 (Frankfurt, Germany) through eu.cloud.agenta.ai, and the US on AWS us-east-2 (Ohio) through us.cloud.agenta.ai. The choice is final. Everything is announced as staying in the selected region — prompts and configurations, test sets and evaluation results, traces, spans and logs, account data and encrypted secrets — with encrypted per-region backups for disaster recovery. The named sub-processor list qualifies that residency: the EU stack relies on AWS (Frankfurt), SuperTokens, Daytona and New Relic in the EU, Crisp (EU, global), and PostHog, Stripe, Loops and SendGrid based in the US; the US stack uses AWS (Ohio), SuperTokens, Daytona and PostHog in the US, New Relic and Crisp in the EU, and Stripe, Loops and SendGrid in the US. International transfers, including to the United States, are covered by the Standard Contractual Clauses and the UK IDTA. This list applies to Agenta Cloud only: self-hosted deployments run on your infrastructure, with your own sub-processors. Dedicated instances, VPC deployment and a custom region are available on Enterprise.
Where Agenta works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Agenta.
- The data region, EU or US, is decided once and for all before sign-up and can never be changed: an architectural commitment disguised as a form field
- Model costs sit outside the subscription and can exceed it; Hobby simply stops agents at its quota, while Pro and Business bill overage automatically, so the budget surprise comes from opposite directions depending on the plan
- Trace retention of one week to three months is not the retention of conversations, project files or persistent agent context: reading one as the other creates false confidence about what can still be recovered
- LLM provider API keys live inside the product, encrypted but still a sensitive asset that needs ownership, rotation and access rules
- Self-hosting looks like a privacy win but moves infrastructure, retention, backups and the whole sub-processor chain onto you: Agenta's DPA and sub-processor list no longer apply
- Deletion after termination takes up to 30 days in production and up to 180 days in backups, so cancelled does not mean erased; contractual liability is capped at the fees paid over twelve months, or 1,000 euros if nothing was paid
- Agents that act on their own on a schedule invite trust by habit: without real approval points, per-tool permissions and someone actually reading the traces, teams stop questioning the reasoning and lose the ability to judge it — the same reflex that should make you notice the footer's 2020-2026 copyright against a Trust Center that dates the company to 2023
Setup & Integrations
Technical difficulty
Low on the cloud, higher if you self-host. Cloud sign-up takes seconds and needs no credit card, and the only structural decision at the door is the data region. Building an agent requires no development skill: it happens in chat, with business templates to start from. The real effort is giving the agent enough context (instructions, files, skills) and connecting the right tools. An optional technical route offers a REST API, OpenAPI 3.1 and a Python SDK. Self-hosting via Docker Compose or a Helm chart demands real operations skills, though a coding agent can drive the install.
Deployment
Integrations
Behind Agenta
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Agenta.
Frequently asked questions
Do I need to know how to code to use Agenta?
What counts as an agent run?
Are model costs included in the subscription?
Which model providers can I connect?
What happens when I reach my usage limit, and do unused runs roll over?
Is my data used to train models?
Where is my data hosted?
Is a DPA available, and is Agenta certified?
Can I self-host Agenta, and how does that differ from Agenta Cloud?
What happens to my data if I stop using the service?
Should you pick Agenta?
Agenta positions itself as the open alternative to closed AI workspaces and visual workflow builders, and the positioning holds up. Its strongest argument is structural rather than cosmetic: the MIT licence means the code you self-host is the code that runs Agenta Cloud, and the harness, model and runtime layers are all interchangeable, so an agent built here is not welded to one vendor's stack.
The second argument is the compliance file, unusually complete for a company of this size: a publicly available DPA carrying the Standard Contractual Clauses and the UK IDTA, a named sub-processor list per region, SOC 2 Type II certification, a contractual ban on training models with customer data, and a choice between an EU and a US data region.
The reservations are just as concrete. Model usage is billed by your provider, not by Agenta, and can easily outgrow the subscription. Trace retention is short outside Enterprise — one week, one month or three months by tier — and advanced governance such as SSO, RBAC, audit logs and custom domains sits on the upper cloud tiers, even though the self-hosted open-source edition already includes SSO and RBAC. The data region is chosen once and never again. Several announced components are still marked Soon, and HIPAA is described as in progress rather than achieved.
The natural audience is technical and business teams running agents in production who want to keep control of their data: the ones who will genuinely use versioning, evaluation and tracing, and who value being able to move the whole thing in-house. Teams that only make the occasional model call, or that expect an all-inclusive consumer product, will find the machinery heavier than the problem.
- Choosing a selection results in a full page refresh.
- Opens in a new window.