What it does
Grimaste runs AI agents that belong to you. Each agent has its own Linux container with real files, a shell and a web browser, and it keeps working when the app is closed. You talk to it in threads from the web app, or from Telegram, Discord or email once you connect them.
Six dials set how it behaves: how much it may reach, how often it asks first, whether it acts on a schedule, how it talks, how much it remembers, and which model thinks. A change applies in about two seconds. The model can be one of the Hoq models or your own key from OpenAI, Anthropic, OpenRouter, MiniMax or any OpenAI-compatible endpoint.
You decide what it can touch. Other Hoq apps such as Repos, Books, Todo and Board are switched on per agent as off, read only, or read and write, and each grant is a separate key that is revoked the moment you turn it off. Secrets go in a vault and reach the agent as files, never in a conversation. When the agent wants to run something risky, you approve it once, for the session, or always, straight from the notification.
In numbers
6dials that set how an agent behaves
8Hoq apps an agent can be granted access to
600ktokens of conversation context per thread, at the largest setting
3outside channels: Telegram, Discord and email
Specifications
| Workspace | Each agent runs in its own Linux container with files, a shell and a browser you can watch and take over. |
|---|---|
| Always on | Agents keep working with the app closed; with the initiative dial at scheduled, an agent acts on its own schedule. |
| Models | HOQ 5.x, HOQ 5.x Flash or MiniMax M3, or your own key from OpenAI, Anthropic, OpenRouter, MiniMax or any OpenAI-compatible endpoint. On your own key, chat bills to you directly. |
| Channels | Reach an agent through Telegram, Discord or email with your own bot token or mailbox password, sealed to that agent. |
| App access | 8 Hoq apps can be granted per agent as off, read only, or read and write. Each grant is its own key for that agent alone. |
| Approvals | Risky actions wait for you: approve once, for this session, always, or deny, including from a push notification. |
| Secrets | Vault secrets are pasted once and given to chosen agents as files. Their contents are never returned to the app. |
| Context | Each thread keeps 100k, 250k, 400k or 600k tokens before summarizing older turns, sized automatically by default. |
| Organizations | An agent can belong to an organization; everyone there can open it, and each person keeps their own threads. Turns from people other than the owner get a restricted tool list. |
| Private agents | Agents can be locked behind a PIN, and their notifications then do not show what they ask. |
Features
27 features, all shipped. Listed from the product's code and documentation; each row shows the day it was last checked.
| Own container | Each agent runs in its own Linux container with files, a shell and a browser. | |
| Watch and drive the browser | Watch the agent's browser live, take over with mouse and keyboard, and hand it back. | |
| Browser pop-out | Move the agent's browser into its own always-on-top window, or fill the whole window. | |
| Shared bookmarks | You and the agent keep one bookmark list, marked with who saved each one. | |
| Terminal | Open several shells into the agent's container in tabs; each one persists between visits. |
| Threads | Talk to an agent in separate threads, each with its own history. | |
| Context per thread | Each thread keeps 100k to 600k tokens before summarizing older turns, sized automatically or fixed by you. | |
| Turn trail | Each step of a turn shows what it is about and is ticked as it finishes. | |
| Channels | Reach an agent through Telegram, Discord or email with your own bot token or mailbox. | |
| Pictures | Pictures the agent sends open full screen. |
| Approvals | Risky commands wait for you, showing the command and why it was flagged; approve once, for the session, always, or deny. | |
| Approve from a notification | Answer an approval from the browser notification without opening the app. | |
| Dials | Six settings for reach, asking first, initiative, voice, memory and model, applied in about two seconds. | |
| Private agents | Lock an agent behind a PIN; its notifications then do not show what it asks. | |
| Running or paused | Pause an agent's machine and start it again. |
| Data sources | Grant Hoq apps per agent as off, read only, or read and write; each grant is its own key, revoked when you turn it off. | |
| Hoq Repos | The agent reads and writes repositories with a capped agent key, every write labelled as the agent's. | |
| Hoq Board | The agent lists, reads and edits boards, and in ask-first areas its changes wait for your approval. | |
| Hoq Commerce and Logistics | The agent reads sales reports, orders, stock and shipments, and can make changes if you allow it. | |
| Vault | Paste a secret once and give it to chosen agents as a file; its contents are never shown back. |
| Hoq models | Run on Hoq Intelligence models billed to your Hoq balance. | |
| Bring your own key | Use your own OpenAI, Anthropic, OpenRouter, MiniMax or OpenAI-compatible key, billed to you directly. |
| Agents in organizations | An agent can belong to an organization; members open it and each keeps their own threads. | |
| Inbox | One inbox holds notifications and tickets from every agent, with read marks that follow you across devices. | |
| Tickets | Agents file tickets you can mark built, answer or decline; tickets about a Hoq app go to that app's builders. | |
| Browser notifications | Replies, approvals and reminders arrive as browser notifications with the app closed, one per thread. | |
| Phone layout | The web app works on a phone, and can be added to the home screen. |
Works with
- Hoq IntelligenceAgents think through Hoq Intelligence and their AI usage bills to the owner's prepaid balance.
- Hoq ReposAn agent with Repos access gets 12 repository tools and every write it makes is labelled as the agent's.
- Hoq BoardBoard lets you set per area whether an agent may look, ask first or change, and lists its pending asks.
- Hoq TodoAn agent with Todo access can work with your lists.
- Hoq BooksAn agent can read or write your books when you grant it.
Your data
| Where it is kept | Each agent's files, memory and conversations stay in its own container and belong to the agent's maker, including for agents shared with an organization. |
|---|---|
| Getting it out | Download files from the agent's Files tab. A full export of memory and threads is not published. |
| If you leave | Not published. |
APIs and standards
- Web app at grimaste.hoq.app, installable with notifications
- Telegram, Discord and email channels
- MCP tools from other Hoq apps (Repos, Board) granted per agent
- Standard: MCP
- Standard: Web Push
- Standard: OpenAI-compatible API (bring your own model)
Compared
| Checked 2026-11-01 | Grimaste | ChatGPT Business | Claude Team |
|---|---|---|---|
| Price | not published | $25 per user per month billed monthly, $20 billed annually (Standard seat)1 | $25 per seat per month billed monthly, $20 billed annually (standard seat)2 |
| Minimum team size | None; an agent can be just yours | 2 seats1 | 2 seats2 |
| Model choice | Hoq models or your own key from any OpenAI-compatible endpoint | OpenAI models1 | Claude models2 |
| Own computer per agent | Linux container with files, a shell and a browser | not published1 | not published2 |
| Reachable from chat apps | Telegram, Discord and email | not published1 | not published2 |
| Scheduled work | Yes, initiative dial set to scheduled | not published1 | Hand off and schedule tasks listed for Pro, Max and Ultra, not Team2 |
| Training on your data | not published | Not by default3 | Not by default2 |
| SSO | not published | SAML SSO included1 | Included2 |