Skip to content
right agent
You usually won't need these docs. The agent guides you right in the chat, and simple setup, maintenance, and use is the whole point. They're a work in progress and actively being expanded.

Concepts

One bot per agent

Section titled “One bot per agent”

You talk to an agent in Telegram. Each surface, a dm, a group, or a forum topic inside a group, is its own Claude Code session, keyed by chat and thread. All sessions share one chat-tagged memory, so separate working contexts still remember the same things about you. In groups the agent stays quiet until you @mention or reply to it.

Every agent in its own sandbox

Section titled “Every agent in its own sandbox”

Each agent gets a persistent OpenShell sandbox with its own filesystem (landlock), its own scoped network, and a tls-terminating proxy. A misbehaving agent cannot reach the host, the other agents, or arbitrary networks. Sandboxes persist: they live as long as the agent and survive restarts.

Credentials stay outside the box

Section titled “Credentials stay outside the box”

mcp tokens and provider keys live on the host. The sandbox sees opaque placeholders that the outbound proxy substitutes on each request. The secret bytes never enter the sandbox, and the proxy never resolves them onto the open internet. A compromised agent can misuse a tool while it runs, but it cannot read the credential.

Memory it can trust

Section titled “Memory it can trust”

Memory persists across sessions and chats. It runs on Hindsight by default, with an agent-managed MEMORY.md file mode as fallback. Recall shows when each memory formed and lets the model judge relevance, with no hidden staleness thresholds. Memory is treated as untrusted data: incoming memories are sanitized, and recalled memory is framed so the model weighs it as information, not commands.

Skills it learns on its own

Section titled “Skills it learns on its own”

When the agent works something out during real use, an api quirk or a multi-step workflow, a per-turn pipeline captures it into a reusable skill and loads it in later sessions. The platform records what each skill costs and how often it runs, a curator prunes the ones that do not earn their keep, and you pin or unpin skills from the dashboard.

A dashboard inside Telegram

Section titled “A dashboard inside Telegram”

/mcp and /providers open a Mini App dashboard with views for health, activity, identity, learned skills, and usage with cost. All management is proxied to the bot’s control plane, and credentials you enter are accepted but never displayed. You never hand-edit config or credential files.