Open-Source AI Agent Framework on GitHub Actions

I run an agent that writes me a morning brief before I'm out of bed. It tracks the repos I care about and stays quiet on slow days. It pays for the web searches it needs out of its own wallet. It does all of this on a schedule, in GitHub Actions, on my own copy of the aeon public repo.

No server. No always-on laptop. No SaaS dashboard charging me by the seat.

This post kicked off a series on why that's possible, and where aeon does things other AI agent frameworks can't - or won't. The series is complete now, and the framework has moved fast since June, so this is the refreshed version with current numbers - including the one change that removes the last excuse: you can now start an agent from the browser, no terminal involved.

aeon is an open-source autonomous AI agent framework by @aaronjmars. There are two ways in. The fast one: sign in with GitHub at Aeon Connect and launch an agent from the browser. The classic one: click Use this template on the repo (or clone it and run ./aeon init), then onboard from the local dashboard. Either way you end up in the same place - an agent running on GitHub Actions runners, in a repo you own.

Out of the box it ships with 85 skills across six packs - small units of work the agent can perform. digest. github-monitor. narrative-tracker. vuln-scanner. create-skill. Each one is a single SKILL.md: a small frontmatter block (name, category, required keys, a per-run var), then a plain-language prompt. The agent reads the file at run time. No DSL. No plugin SDK. No compile step. If you can write a prompt, you can write a skill.

When this article first went out, the catalog held 61 skills and Claude Code was the only engine. Today aeon drives nine harnesses - Claude Code, Grok, Codex, Pi, Vibe, Kimi, fx, Cursor, Hermes - and you pick one when you connect a model, not when you create the repo. The skill files don't care which one runs them.

The default delivery channel is Telegram. You DM a bot once, your chat ID lands in a secret, and from then on every skill posts its output to your phone. You can also reply to the bot and route messages back into aeon as triggers. Discord, Slack, and email work the same way.

Three things stand out. The series worked through each in turn.

1. It runs in the background by design - no server, no orchestrator

LangGraph, CrewAI, AutoGen, AutoGPT - they assume a process is already running somewhere. You decide where to host it. You monitor it. You restart it when it dies. Infrastructure first, agent second.

aeon starts from the opposite end. The unit of deployment is a cron line in YAML. GitHub Actions does the running. The result is a scheduled and reactive flow - you fire skills by cron, by DMing the bot, or by pushing a commit. Failed runs self-heal: the next scheduled tick fires regardless, and since the summer the healing got teeth. skill-health watches for skills that keep failing, skill-repair patches them, aeon-doctor checks the installation itself, and every run gets rated so drift shows up before you notice it in your inbox.

That sounds small until you have five skills firing every day for a month and realize nothing has paged you, nothing has gone offline, and the only thing you ever maintain is the YAML.

The agent that runs while you sleep covers this architecture in detail - what changes when "the agent runs while you sleep" stops being a slogan.

2. It can spend money - and now it can earn some back

Most AI agents are read-only. They fetch, they summarize, they post - but they don't act in any economic sense. They can't, because acting in the world usually means an API key, a billing account, a credit card.

aeon plugs cleanly into the Base MCP and the x402 protocol. The agent has its own Base Account wallet. It can search the x402 bazaar, find a paid API, pay for one call, and use the result - no API key, no bank account, no prefunded billing setup. Every payment clears through Base Account's approval step, so nothing settles until you tap approve.

What this unlocks in practice:

• Prediction-market odds for a narrative-tracker run.
• Live web search when free APIs rate-limit you.
• Premium feeds for a digest.
• A reliability check before any of it: cortx-reliability verifies an x402 endpoint is actually delivering value before the agent spends a cent on it.

And the flow now runs in both directions. compute-resell lists your spare provider compute on an open market and prices it reactively inside guardrails you set. taskmarket-delegate hands work to other agents for pay. The agent decides what's worth a few cents, pays for it, and increasingly earns a few cents of its own.

Letting your agent spend money covers the x402 side in depth.

3. Skills are Markdown files, and they compose

A skill in aeon is a single Markdown file. That's the whole interface. You can:

• Edit any of the 85 catalog skills in place in your copy.
• Write a new one in twenty minutes by copying a similar one and rewriting the spec.
• Chain skills together in aeon.yml - the output of one skill becomes the input of the next.
• Ask aeon to write a new skill for you. create-skill and self-improve are real skills in the catalog, and community skill packs now exist for the ones the catalog doesn't cover.

Everything is in git. Every change is a commit. Every run is a workflow log. You can see exactly what your agent did last Tuesday at 04:00 UTC, why, and what it produced - no observability stack to spin up.

Skills that build skills covers the composition story and Receipts, not vibes covers the audit trail.

aeon is not a hosted agent. Nobody runs your agent for you - it lives in your GitHub repo and executes on your Actions minutes, with your keys in that repo's secrets, even when you set it up through Aeon Connect. Connect hosts the dashboard and a remote MCP server; the execution layer stays yours. Stop using it and the agent keeps running. That used to come with a catch - the bottom rung of the ladder was "clone the repo and read the YAML". Connect removed the catch, not the ownership.

It doesn't have a flashy chat UI either. You onboard in the browser at Connect or in the local dashboard at localhost:5555 - connect a model, add a channel, toggle skills - and after that the interface is your inbox, your Telegram, your repo log.

The fast way - no terminal. Go to www.aeon.fun/connect, sign in with GitHub, and create your aeon: public is free on Actions minutes, private keeps memory and run logs private. Install the Aeon Connect GitHub App on the one repo it creates (it never sees the rest of your account), paste a model credential - Claude token, API key, ChatGPT login, or OpenRouter - pick skills, and run. Five steps, nothing installed. The Aeon Connect article walks through each one.

Using a coding agent? Paste this into Claude Code, Codex, Hermes, or OpenClaw: read https://www.aeon.fun/skills/aeon.md and follow the instructions to set up your aeon agent.

The classic way. You need Node.js 20+ and the GitHub CLI authenticated (gh auth login). Then:

git clone https://github.com/aeonfun/aeon && cd aeon
./aeon init

./aeon init creates your repo and connects GitHub, a model, and Telegram; ./aeon opens the same dashboard locally on localhost:5555, where you authenticate with any of the nine harnesses, add a channel (Telegram, Discord, Slack, or email), pick skills and their cron schedule, and run. The schedule starts on the next cron tick. Everything the dashboard does is also an ./aeon CLI command or a /aeon chat command, installable as a Claude Code or Codex plugin. If you want a stable point to pin to instead of main, releases exist now, starting with v0.1.0.

Both paths end at the same repo. Start on Connect, drop to the CLI later when you want to edit skills by hand - nothing to migrate.

Is aeon free to run?

The framework itself is free and open-source (MIT), and so is Aeon Connect. You'll pay for GitHub Actions minutes beyond the free tier (public repos get minutes free; private repos include 2,000 a month) and for the provider behind whichever harness you pick. Optional paid APIs pay themselves through the agent's x402 wallet.

Do I need to install anything?

Not anymore. Aeon Connect runs the whole setup from the browser - sign in with GitHub, connect a model, pick skills. (The one exception: a Claude subscription token comes from claude setup-token; an API key or OpenRouter needs no terminal at all.) The terminal path still exists for people who want it, and you can switch between the two freely.

How does aeon compare to LangGraph or CrewAI?

LangGraph and CrewAI are libraries you assemble into a running process. aeon is a deployment pattern - skills as Markdown files, scheduled in YAML, executed by GitHub Actions. You can think of it as "the agent framework where the orchestrator is git + GitHub".

Do I need to know Python?

No. Skills are Markdown. The engine is whichever harness you authenticate - Claude Code, Grok, Codex, and six more. You'll touch YAML for scheduling and Markdown for skill prompts. Bash and Python show up only inside specific skills that need them.

Can I run aeon without Telegram?

Yes. Telegram, Discord, Slack, and email are all supported - pick any, all, or none. With no notification channel configured, skills still run; output lands in the repo as commit artifacts.

Can I run more than one agent?

Yes. spawn-instance stands up a second aeon from your first one, and fork-fleet plus fleet-control coordinate a set of them. One repo per agent, git as the shared memory.

Is my data private?

Your copy is yours. Secrets live in GitHub Actions secrets. Skill outputs commit to your repo. The framework doesn't phone home. On Connect, a public agent's memory and logs are public like any public repo - pick private at creation if that matters to you.

Article 2: The agent that runs while you sleep - the architecture behind background-first agents.

Article 3: Letting your agent spend money - x402 in practice.

Article 4: Skills that build skills - the composition story.

Article 5: Receipts, not vibes - the audit trail.

New: Aeon Connect: start your agent from the browser - the zero-terminal on-ramp.