Hey Product Hunt 👋

This started as a weekend hack to poke my home server from my phone. I kept opening a terminal and SSH-ing in for tiny things. Restart a service, check a log, pull a repo. So I put a coding agent on the VPS and gave it a chat bot. Then I gave it a memory. Then a team. Then it stopped being about the phone at all, and the thing I actually use every day is the web panel sitting on my own machine.

MyAgens is what it grew into: a self-hosted fleet of AI agents that lives on your hardware and reports to you. There is no MyAgens service in the loop. The only network calls it makes are the ones you configure, your model provider and any connector you explicitly turn on.

- No chat app required. The web panel is the whole product: chat, crew, board, memory, skills, connectors, schedules, logs, settings. Telegram and Slack are optional front doors you can add later without reinstalling anything, and the setup wizard has a Skip button on both. It started life as a Telegram bot, and getting it to the point where Telegram is genuinely optional took most of the last two releases.

- Six agent backends, one fleet. Claude Code, OpenAI's Codex CLI, Google's Antigravity, Cursor's CLI, xAI's Grok CLI, and local models through Ollama. The CLI ones wrap that vendor's own agentic CLI rather than reimplementing a tool loop, and Codex, Antigravity and Cursor go further: they get your MyAgens tools, your persona and playbook, and the same Approve/Deny gate, so they answer as your agent rather than as a stock CLI. Agents on different backends sit in the same fleet sharing one memory, one skills library, one task board.

- Failover that actually fails over. Every agent carries its own fallback. Hit a rate limit and the turn retries somewhere else mid-flight and tells you. Background work moves before you hit the wall, and it won't switch into a fallback whose own numbers say it is already spent.

- It runs like a small ops team. Atlas, the chief coordinator, handles day-to-day work and commands specialized Leads (Finance, DevOps, Research, or whatever you define). Each Lead has its own memory, model, backend, and persona.

- Nothing risky runs without you. Every non-read-only tool call stops for an inline Approve, Deny, or Always-allow. Approvals are shared across surfaces, so a pending call shows up wherever you are and answering once settles it everywhere. Autonomy is a dial, not a switch: supervised, standard, or full.

- Memory that accumulates, skills that compound. Facts decay from hot to warm to cold storage on their own, and expensive work gets distilled into reusable skills your fleet keeps improving.

- A council, because one opinion is cheap. `/council <proposal>` puts an idea to a vote across Atlas and every enabled Lead: simple majority, supermajority, or unanimous. With embeddings on, each voter is weighted by how relevant the proposal is to their domain.

- Heartbeat. It watches CPU, memory, swap, disk, and stalled tasks continuously, with quiet hours if you want them, and the alert reaches whichever surfaces you actually run. When a service falls over at 2 am you can fix it from your phone, no SSH client required.

- Honest security. Secrets live in an AES-256-GCM vault, keyed to the macOS Keychain or a locked-down `0600` file on Linux. Full searchable audit log with anomaly detection, a hard allow-list on every surface, and no telemetry, ever.

It is free and open source under GPLv3, written in TypeScript, runs on macOS, Linux, and Windows and installs with one command.

Where it is still rough, so you hear it from me first: Grok and Ollama are thinner integrations than the other four backends (Ollama gets a small hand-built prompt and a single approval-gated Bash tool, which is the price of running fully local), Antigravity reports no token or cost accounting because its CLI emits none, and three connectors that run as their own process (Unreal, Unity, Browser Sketchpad) are still Claude-only. Cursor is the one that has to reach into your project, since it reads its hook and MCP config only from the folder you are working in, so MyAgens writes into <project>/.cursor/ for the duration of a turn and restores it afterwards. There is a checkbox to switch that off if you would rather it never touched the folder.

I would love your feedback, and especially the first workflow you would hand to a fleet. I will be here all day answering everything.

(Bonus: if your fleet lives on a Mac, MyAgens pairs with Keepresso. Keepresso keeps the lights on, MyAgens works the night shift.)