Introduction
egin is a filesystem-first framework for durable backend agents. An agent is a folder, and one Rust binary runs, serves, evaluates and benchmarks it.
Most agent frameworks are libraries: you write a program around them and the agent lives inside that program. egin turns this around. The agent is a folder of plain files, and egin is the binary that runs it.
my-agent/
agent.toml model and switches
instructions.md system prompt
tools/ one script per tool, any language
evals/ prompts and checks
workspace/ mounted in the VM
That folder is the whole agent. You can commit it, review a change to its instructions like any other diff, and run the same folder from a laptop, a server or a CI job.
What egin does for you
- Runs the loop. Models come from the Claude API, OpenAI, Gemini, Amazon Bedrock, the Vercel AI Gateway, the You.com Inference Router or any OpenAI-compatible endpoint; egin picks one from the key you set. Web search comes from You.com, Parallel, Firecrawl, context.dev, Brave, SearXNG or the model provider’s own.
- Isolates every session. Tool calls run in a sandbox created for that session (a microVM where the host supports one, or E2B, Vercel, Daytona or Docker), with network egress you can limit to a list of hosts.
- Survives crashes. Every message and tool result is stored in SQLite before the loop moves on. A killed run resumes from the last stored message and never repeats a model call.
- Serves it.
egin serveexposes an HTTP API with a server-sent event stream, Slack and webhook channels, and cron schedules. - Measures it. Evals, config variants, per-model profiles and full benchmark suites run through the same binary, on the same VMs.
The name
egin is Basque for “to do, to make”. It is short, it types easily, and it says what an agent is for.
Where to go next
- Quick start: a working agent in four commands.
- The agent folder: every file egin reads and what it does.
- Durability: what happens when a run is killed halfway.