Self-hosted · Open source
Your personal AI headquarters
MyAgens is a self-hosted fleet of AI agents powered by your choice of backend: Claude Code, Google Antigravity, OpenAI Codex, Cursor CLI, local Ollama, or Grok. Message your crew from Telegram or Slack, watch them work in the panel, and approve anything risky before it runs.
Chain of command
Runs like a small ops team
You sit at the top and make the calls. Atlas holds it together and delegates by name; leads and their assistants do the work. Everything they file, ask, suggest, and report moves up and down one chain.
You
Set direction from Telegram, Slack, or the panel and approve anything risky. The fleet pauses for your call; it never assumes it.
Atlas
claude codeHandles what you send, remembers everything, knows every portfolio, delegates by name, and reports back when it is done.
Finance
claude codeOwns the money: budgets, invoices, renewals.
DevOps
codexOwns the machines: deploys, certs, backups.
Research
ollamaOwns the questions: digs deep, reports back.
The roster has no fixed shape. Spin up a persona for the work you do every day, and it slots into the same chain. Plenty of people run Atlas on its own and never add a lead at all.
the names above are examples, not a menu: you write the persona, the memory, and the working directory
Atlas hands a task card to a lead; it streams progress and closes the card itself when done.
Task board
A board where the cards do the work
Every piece of work is a card. Agents file their own, split the big ones into subtasks, and take them by name. A delegated card streams its run log while it works, then moves itself to done. The columns are yours to name.
Cards that run themselves
Hand a card to a lead by name, or let Atlas route it. The board moves it to In Progress when the run starts and to Done when it finishes. Nothing to drag.
The whole run, live and afterwards
Every tool call and every line of output streams into the card while it works. When it is over the transcript stays on the run, so you can read what actually happened.
Repeats and schedules
A card can repeat daily, weekly, or monthly: the template stays put and drops a fresh copy into the backlog on time. Scheduled prompts file their results the same way.
Queues, blockers, and retries
Three runs at a time by default and the rest queue. A card can wait on other cards before it starts. A run that fails resumes its own session on retry instead of starting from scratch.
Any engine
Swap the engine, keep the agent
Memory, skills, connectors, approvals, and cards live in MyAgens, not in the model. Point an agent at a different backend and it keeps everything it knows.
point an agent at another engine, or let a usage limit move the turn for you
What makes it a fleet
Not just a bot
Real AI agents across 6 backends with file, shell, and service access, wrapped in the structure that makes a crew useful day after day.
Memory that accumulates
Agents save durable facts and recall them automatically. Hot, warm, and cold tiers decay and promote on their own, with optional semantic search that runs fully offline.
Skills that compound
When Atlas works out a procedure worth reusing, he distils it into a skills library and pulls it next time. The library grows with use.
Council votes
Put a proposal to the whole crew. Every lead evaluates it from their domain and votes with a reason and a concern; the tally lands in Telegram or Slack.
Approval before anything risky
Every non-read-only tool call pauses and shows exactly what is about to run. Approve, deny, or always-allow, from Telegram, Slack, or the panel.
Secrets, encrypted
Agents read API keys and tokens from an encrypted store, scoped per agent, never from plaintext files. Nothing sensitive lands in logs or in a message you did not approve.
Proactive heartbeat
Background monitoring of CPU, memory, disk, and stalled tasks. It can ping you, or run an autonomous turn to investigate and act first. Quiet hours included.
Scheduled runs
Any agent can run a prompt on a timer: check disk space at 9am, summarize logs every 2 hours, pull a report every Monday. Results push back to Telegram or Slack.
Task board with delegation
A Kanban board where Atlas hands cards to agents by name. Each card streams its run log live and moves itself to done, so you watch the work flow without chasing it.
Connectors & Browser
Multi-account Socials (Bluesky, Mastodon, Discord, Reddit, X, YouTube, Facebook Pages), Notion, Gmail, Drive, Calendar, GitHub, Jira, Linear, and a headless Playwright Browser Sketchpad for visual web verification.
Six AI backends, one set of memories
Run any agent on Claude Code, Google Antigravity (agy), OpenAI Codex, Cursor CLI, local Ollama, or Grok. Your tools, persona, memory, and approvals come along, and a usage limit fails the turn over to a backend that still has room.
Watch or take over, live
Put an agent in a persistent tmux terminal with cinema view, watch live from the panel viewer, or connect via claude.ai and the Claude mobile app with Remote Control.
Yours to brand
Rename the product and agents, drop in your own logo, favicon, and accent, or describe a look and let a one-shot pass draft a full custom theme. GPLv3, no seat limits, no telemetry.
MyAgens Panel
The whole fleet on one screen
An optional web dashboard served in the same process. Chat with any agent, watch live system health, run and schedule work, and manage memory, skills, and secrets.
Telegram & Slack
Your machine, in your pocket & team chat
The old loop for touching a server was open a terminal, SSH in, run something, close it. Now it is a chat in Telegram or Slack with an AI agent living on your host.
- Replies stream back live as they are written, then land as structured messages.
- Tool calls pause with approve, deny, and always-allow buttons inline in Telegram & Slack Block Kit.
- Leads get their own bots or direct-message surfaces: message your devops lead directly.
- Send photos and files straight into a chat; answer clarifying questions by typing or clicking.
Install
Up and running in one command
On a fresh machine the wizard installs everything it needs (Node 20+, git, the Claude CLI), clones the repo, walks you through setup in your browser or right in the terminal, and can register a background service. macOS, Linux, and Windows.
Browser setup
One pasted command; the terminal does the boring part, then a guided setup page opens in your browser and verifies every value live. No IDs to look up, no typos to make.
Get the command →Terminal only
Headless box, server, or over SSH: the same installer runs the classic wizard right in the terminal instead, no browser needed.
Get the command →The setup page walks you through creating your bot in BotFather's Mini App and recognizes you when you press START in its chat, so there are no IDs to look up, on macOS, Linux, and Windows alike. Only the terminal-only wizard (headless boxes, SSH) asks for the token and your numeric user id inline. Prefer to read before you run? The scripts live in the repo.
A sleeping Mac takes the whole crew offline
Keepresso is a native menu-bar app that keeps your always-on Mac awake the honest way: real power assertions, smart triggers, closed-lid mode, and a headless setup toolkit. Free and open source, from the same desk as MyAgens.
Get KeepressoWhat would your whole crew cost?
tokenstat prices every local session at list rates, so you can see exactly what your crew would cost if you paid the APIs yourself, and how much value your subscription gives you. Local by default, with a free plan and source code available.
Claim your handle now