A library of ready-to-use AI agents. Create a new agent in three steps. In the cloud, with fine-grained access control.
Until a connection is bound to an agent, its tools don't exist for that agent. Bindings are visible from both sides: on the agent and on the connection.
You can deny an agent a single tool, limit the values of its parameters or close a connection entirely. If a rule blocks the call, the request never reaches the external service, and the denial is written to the log.
The log shows which agent called a tool, with what parameters, and what the service answered. Errors and policy denials can be filtered out on their own.
On its own, a model only writes text. To send an email or file a task it needs tools, keys to your services, rules and a log. We call all of that the harness. It is shared by every agent and configured for each one separately.
Tools come from connections: an MCP server, your own API behind a key, Telegram, your phone or computer. An agent sees a tool only while that connection is bound to it. For example, sending a Telegram message or reading a spreadsheet row.
A skill says in what order to call tools, what not to do, and which connections it needs. One skill can go on several agents; ready-made ones are in the catalogue.
In plain words: A skill gives the agent the steps: read the receipt, add the expense to the sheet, and ask you if the totals don't match.
Connection keys and tokens are stored on the platform. The model sees only tool names and descriptions, and the key is added after it has chosen a call. The exception is the LLM provider key: the model can't be reached without it, so it goes to the platform worker.
You can allow or deny an individual tool, limit its parameters, or close a connection entirely.
In plain words: The agent may read your mail but not delete it. The check runs before the request reaches the mail server.
For every run you can see the agent's steps, each call with its parameters and the service's answer, and how many tokens each model used.
The agent runs on the platform: in our cloud, or on your own server if you host AgiMate yourself. Your phone and computer connect as connectors and carry out commands, but the agent itself never runs on them.
In plain words: So every call passes through the platform, where the rules get a chance to work. An agent running directly on your machine talks to the system with nothing in between to check it.
How it works
A trigger, say a new email, arrives from a connection or a device. The platform checks it against the rules and hands it to one agent. When the agent needs to act, it calls a tool through the core, and that call is checked the same way. The core adds the keys; the agent never gets them. Allowed and denied events are written to the trigger log and the tool call log.
Diagram: triggers from connections and devices (Mail, Calendar, Laptop, Android) pass a policy check in the core and go to one agent. Tool calls pass the same check, and the agent's answer goes to the user in the web chat, Telegram, the phone app or an IDE. Below the diagram are two logs, triggers and tool calls, with each row marked allowed or denied. Packets on the diagram carry Greek letters, which match them to rows in the log.
A model can break an instruction if it's pushed hard enough. A tool from a connection that isn't bound to the agent can't be called at all: for that agent it doesn't exist. So an agent's role is set first of all by what it is connected to.
The provider and a priority list of models are set per agent: for chat, for reading images and for generating them. An analyst can get a strong model, while a reminder agent does fine on a cheap one.
Each agent has its own instruction text, from a ready-made role or written from scratch. One answers customers briefly; another goes through a report and cites the numbers.
Connections are bound to a specific agent, and inside a binding there are rules for individual tools and their parameters.
In plain words: Like staff passes: the accountant gets the spreadsheets, the secretary gets mail and the calendar.
Skills are added to a specific agent, and tools that other agents' skills need don't appear on it. The less an agent has lying around, the less often it reaches for the wrong tool.
| Agent | Model | Instructions | Access | Skills |
|---|---|---|---|---|
| Analyst | strong | “go deep, show your working” | data and reports — read only | summaries, charts |
| Assistant | cheap, fast | “remember what matters, remind me early” | calendar over MCP, push to the phone | fact notes, reminders |
| Designer | vision and image generation | “show three options, then refine one” | files: read and write | image editing, generation |
| Language tutor | mid-range | “correct gently, praise a streak” | Telegram, a vocabulary file | vocabulary, error review, motivation |
| Home accountant | mid-range, careful with numbers | “count from receipts, ask when they don't add up” | the expenses sheet in files | expense tracking, monthly summary |
A ready-made role comes with instructions and matching skills. You can change any of it while creating the agent or later.
Every role in the gallery already has a description, instructions and a set of skills. If none fits, start with a blank agent.
The role's skills are already ticked. Untick the ones you don't need, add your own, or skip the step and come back to it later.
The agent answers in the dashboard chat straight away. Telegram, your own model and connections to services can be added later, when you need them.
Each scenario is built from parts the dashboard already has. Under the description you can see which ones.
Connect Telegram as a channel and the agent starts answering there according to your instructions. You can read the conversations in the dashboard afterwards.
On a schedule, the agent checks an external service and sends you the result. For example, it checks every hour whether your site is up, or collects new leads each morning. A job can be paused or run outside the schedule.
An external system sends a webhook, say about a new payment or a failed build, and the agent decides what to do next. Deliveries and tool calls are visible in the trigger log.
The Android and desktop apps connect your device to the agent. A notification arrives, the phone rings or a file changes, and the agent receives it as an event. In reply it can send a push, read text aloud or run a command on the device.
Over the ACP protocol the agent connects to your code editor. It keeps the same skills, access rules and log as in Telegram or the web chat.
Connect your API as an app with a key or as an MCP server, and its methods become the agent's tools. Rules decide which of them the agent may call.
Several agents work from one task board: each task has a status, an assignee and a comment thread. Changes on the board show up in the dashboard immediately.
Tokens are billed by your provider at your rate
Connect a provider with your own key, and tokens are billed by that provider at your rate. The platform takes the model list from the provider and marks the ones that are currently unavailable.
Chat, image recognition and image generation each get their own priority list of models. If the first model on the list is unavailable, the platform uses the next one.
The dashboard shows how many tokens each agent spent and on which model. If you don't have a key yet, you can start on the platform's free allowance.
Talk to your agents wherever suits you
MCP servers and your own API behind a key connect today. The catalog of ready-made integrations keeps growing.
Connectors for the phone and the computer: events on the device reach the agent as triggers, and the agent calls tools on the device.
An Android app. A call, a Wi-Fi change, a low battery or entering a zone reaches the agent as a trigger. In return, the agent can show a notification, ask a question, look into the calendar and contacts or read text aloud.
Learn moreAn app for Windows, macOS and Linux. A new file in a folder or a quick-button press reaches the agent as a trigger. In return, the agent can show a notification on the computer or read text aloud.
Learn moreThe source is open in full: the core, the dashboard, the chat app, the connectors and the n8n nodes. Running it yourself is free, company use included; a licence is only needed to sell AgiMate as a service to others. Your data, model keys and logs stay with you.
Sign in with GitHub, Google or Yandex — the account is created on the first login.
Self-hosted · Your own model key · Every call logged