AgentRun: open-source managed agents that learn from each run
From AgentRun: README · Open at AgentRun: README
Parcha Labs, Inc. has published AgentRun, an open-source platform for running AI agents. It sits on top of an existing agent harness. It adds self-improvement, durable runs and a control plane. The code is on GitHub, and the AgentRun README on GitHub describes it. The README shows no release date. The project was posted as a Show HN on the Hacker News front page on 2026-11-12.
What the original source confirms
| Item | What the source says | Where |
|---|---|---|
| Maker | Parcha Labs, Inc. The README says it was "Built by Grep.ai". | AgentRun README |
| Licence | Apache-2.0. Each bundled component keeps its own licence. | AgentRun README |
| Tagline | "Run smarter agents. Use fewer tokens." | AgentRun README |
| Learning | After a run, the agent can keep notes, skills or workflows. Kept items go into a git folder. | AgentRun README |
| Durable runs | Runs are stored in SQLite by default, or Postgres. Runs resume after a server restart. | AgentRun README |
| Harnesses | pi (default), Claude Code, Codex and OpenCode. | AgentRun README |
| Model keys | OpenRouter, Anthropic and OpenAI keys are named. | AgentRun README |
| Requirements | Node 24.22.3 or newer, on macOS or Linux. Docker is optional. | AgentRun README |
What changed for people who use it
AgentRun starts with one command, npx @parcha/agentrun. This starts a server and a runner on port 7340, or the next free port. Data is kept in ~/.agentrun. You set a password in the console and add a model key.
Learning is turned on with "learning": true in an agent's definition. A Learning tab lets you review or retire what the agent kept.
The README lists control features. These are per-tool approvals, spend caps that pause a run instead of failing it, and agent questions that wait for a reply. With several runners, another runner takes over a dead runner's runs after a 30-second lease.
Without Docker, agents can read and write files but have no shell. With Docker, agents get a sandbox with a shell.
Other ways to run it are a Render one-click deployment and Docker Compose with a server and two runners. The README also names a /v2 API, a CLI, a TypeScript client and a Python client.
What the source leaves out
| Item | What the source says | Where |
|---|---|---|
| Release date | Not stated. | AgentRun README |
| Version number | Not stated. A CHANGELOG is referenced. | AgentRun README |
| Windows as a host for the server | Not stated. The requirements name macOS and Linux. | AgentRun README |
| Independent tests of the cost case studies | Not stated. The cost figures in the README are the vendor's own. | AgentRun README |
Running open models in your own data centre is what Command is for. Get Command early access