Identity
An agent can only do what you've authorised it to do. Never more than the person who started it.
An agent that only gives advice saves you a few minutes. You get far more time back from one that completes the task in your systems on its own. With Agentic you can let it work with money, clients and sensitive data, because limits, approvals and a record of every step are part of each run.
In a short meeting we'll show you how an agent would handle one of your processes and what it would save you.
Once an agent is changing system state, spending money or touching sensitive data, you need answers to five questions. Click a row to see the answer you'll have with Agentic.
WITH AGENTICEvery risky step has a named approver on your side. Who decided, when and why is in the same record as the agent's work.
WITH AGENTICYou know what every run and every step cost. You set the budget for a team or an agent yourself, and the platform enforces it in real time. Once it runs out, further calls stop.
WITH AGENTICYour keys to AI services are kept in a vault where nobody can see them. Every agent has its own identity and temporary access for a single call, never with more rights than the person who started it.
WITH AGENTICAgents work inside your infrastructure, so they can reach internal systems too. Sensitive details are replaced with tokens before the model sees them, and the data never leaves your network.
WITH AGENTICYou don't lose work in progress. The state is saved after every step, and after an outage or days of waiting for approval the run picks up exactly where it stopped.
The LLM Gateway gives you control over every access to AI models and tools. On top of it, the Platform runs the agents' actual work. Pick a layer.
The LLM Gateway is always part of the Platform. So your agents never run without control over access, budget and data.
All your AI models sit behind a single entry point. For every call you know who made it, what it cost and whether it passed your rules.
An agent can only do what you've authorised it to do. Never more than the person who started it.
You set a budget for every key and every step, and it's checked on every call.
Sensitive details are replaced with tokens before they reach the model. Data stays inside your infrastructure.
Models from the major providers and local models through a single interface. Switch providers without rewriting integrations.
Pseudonymisation, which the GDPR recognises as a safeguard, is applied automatically to every call. The detectors are tuned for Czech and European data. They recognise a Czech birth number (the local equivalent of a social security number) by its check digit and an IBAN by its checksum. The original can only be retrieved with authorisation and an entry in the audit trail.
…pay to account [IBAN_TOKEN_7f3a], birth no. [RC_TOKEN_91cd]…
Token → original only with authorisation, a stated purpose and an entry in the audit trail.
The model only makes recommendations. Control is enforced outside it, around every input, output, tool call and action.
Step limits, cost limits and a circuit breaker apply for the whole run.
Every guardrail can reject a request, send it to a person for approval, warn, just log it, or mask the sensitive part. You can test rules in a simulator before switching them on.
Detectors for Czech birth numbers (mod 11), IBAN (mod 97), payment cards (Luhn), email, Czech phone numbers and custom patterns. If masking fails, the request isn't sent.
On the Platform your agents can work for long periods, stop for your people and work with each other. And you always know exactly what they're doing.
In Agent Studio you describe in your own words what the agent should do. You get a plan built from the real tools and models available to you, with every choice explained. You run the agent once for real and keep exactly the version that passed.
Business users work in a simple mode with no technical terms. Developers get scripts, containers and the full API in advanced mode.
You see the cost and the record of every step. You can give a step a goal ("done when…"), a number of attempts and a spending cap that the LLM Gateway won't let it exceed.
The approver sees exactly what the agent wants to do and with which data. They decide in the app or with a button in Slack. For a tool call, they can even adjust its parameters before approving.
A checkpoint is saved after every successful step. The run continues exactly where it stopped, and you can restart it from any checkpoint with adjusted inputs.
Up to 20 sub-agents working in parallel over a list of items, plus nested calls to other agents. You see everything on one timeline, so the audit stays complete.
The record is created during the run, in the same stream as the agent's work. You can show an auditor or regulator the whole case, including steps, decisions, approvers and costs. Security-relevant records are kept for up to 365 days.
Run history and state survive a process crash and can be fully replayed. The approval gate returns the waiting state directly in the response (HTTP 449), not through a webhook or a notification alongside the running process.
Give your people time back for the work that needs a human. Whatever process you hand to an agent, you keep the same control over it. A few examples out of many:
Don't see your process here? All the better for a first meeting.
Walk through one agent run. Steps on the left, what is happening at that moment on the right.
The agent reads the PDF, extracts the details and checks the supplier against the register. Every model call goes through the LLM Gateway and is logged for you.
Your approver sees the whole case in the app or directly in Slack and decides with one click. The history is stored in the platform for later review. The data in this example is fictional.
You run the whole platform in your own environment, even without an internet connection. Every component writes what it did into a single audit trail, so your security team can examine it in depth.
| Component | Technology | What it does |
|---|---|---|
| Agent Studio | React · Vite | Building agents: describe → plan → verify → keep. Business and developer modes, tests and versions. |
| API | FastAPI | Authentication and management of projects, agents and runs. OpenAPI for your integrations. |
| LLM Gateway | FastAPI | Budgets, guardrails, personal data protection and tool access via MCP. Always part of the Platform. |
| Worker | Temporal SDK | Runs agents step by step according to their definition. |
| Temporal | durable execution | Run history and state survive a process crash. Any run can be replayed. |
| Storage | MongoDB + GridFS | Agent definitions, runs, audit records and outputs. |
We verified all five directly in the code. On top of that, thousands of automated tests check every change, including simulated cross-tenant attacks, so we catch most bugs before they reach you.
SSO via OIDC/Keycloak. New users always start with no extra permissions.
Platform administration and project roles are kept separate. Only selected roles can approve.
Vault or Azure Key Vault behind a single interface, with no changes to the code.
Custom scripts run without network access, with memory and time limits.
Webhooks and tool connections are checked against access to the internal network.
Measured in the platform repository on 27 August 2026.
No generic presentation. We start with one of your processes, and you'll see straight away whether and how an agent pays off.
You pick one process and we go through together how an agent would handle it and what it would cost you. No commitment.
We build the first agent on your real data. We do the implementation together with your team.
Agents move into production, and the same rules apply to every further process, team and model provider. You're backed by a team that has delivered critical projects for 30 years.
Get control over the AI you already use. It goes on top of your current setup with no migration, and you see costs and data straight away.
Pricing upon requestYou pay for completed and approved work. Attempts and failed runs aren't charged. The LLM Gateway is always included.
Pricing upon requestNo. You keep the licences you already have, and with Agentic you gain control over them. The LLM Gateway also handles GitHub Copilot and ChatGPT subscriptions, and you see their usage in the same budget as paid APIs.
A proxy tells you what a single API call cost. With Agentic you know what a whole agent run cost, step by step, and your people can stop an action before it happens. You can keep your existing gateway, and Agentic connects to it.
They're great for connectors and simple automation. Agentic is for cases where an agent touches money, clients or code and you need budgets, personal data protection, approvals and runs inside your own network. The two can run side by side.
Yes. You can run the platform yourself, even in a fully isolated network. Without internet access you use local models, for example through Ollama or another model in your cluster.
OpenAI, Anthropic, Google Gemini, Azure OpenAI, GitHub Copilot, ChatGPT/Codex and local models through Ollama, plus any OpenAI-compatible endpoint. You choose the model for each agent separately and can switch it at any time.
No. No telemetry leaves the platform. Your data only goes to the model providers you set up, through your own keys.
Human oversight, logging, risk management and testing are built into the product. They take the form of approval gates, an audit trail, guardrails, budgets and test suites. In the intro meeting we'll show you how they cover Articles 9–15.
Show us one process and we'll show you how it would work with an agent. The intro meeting doesn't commit you to anything. You choose the time directly in the calendar.