← Back to archive

Repo of the Day

Parcha-ai/agentrun

Published: Nov 16, 2026

Open repository ↗

The Agentrun Workflow DSL

Summary

AgentRun is an open-source JavaScript platform that adds learning, durable execution, and a control plane around agents running on Claude Code, Codex, OpenCode, or pi. It retains notes, skills, and workflows between runs, stores run state for recovery, and provides approvals, spend caps, triggers, history, and cost tracking in one console or through an API.

What it is useful for

  • Repeatable agent work: With "learning": true, a completed run can propose notes, skills, or a workflow. AgentRun checks and grades proposals, commits accepted material to the agent’s Git-backed learning folder, and makes it available to later runs. This suits work such as code review or merchant classification, where an agent might otherwise rediscover the same process each time.
  • Operational control: Every agent, revision, run, and step is stored. If the server stops mid-run, the run can continue after a restart; with several runners, another can take over after the failed runner’s 30-second lease expires. Tool permissions, approvals, spend caps, and human questions can pause work for review.
  • Structured automation: JSON workflows support typed decisions, tool calls, model steps, parallel mapping, bounded loops, and agent fallback. The console shows each node and decision and allows workflows to be paused and resumed.
  • Existing-agent visibility: The Connect agent flow can observe agents started elsewhere through the CLI and OpenTelemetry Collector. This shows activity and token usage without changing where the application starts its jobs.

How engineers can use it

The documented local setup requires Node 24.22.3 or newer on macOS or Linux:

npx @parcha/agentrun

This starts a server and runner on port 7340 or the next available port, stores data in ~/.agentrun, and opens the console. Set a password, then either choose Continue with ChatGPT for Codex and pi agents or add an OpenRouter, Anthropic, or OpenAI key as appropriate. A first trial can apply the KYB due diligence template and provide a company name and website; afterward, inspect its cited report and the Learning tab.

To use an existing agent, open Connect agent in Mission Control or Connected agents. Observation requires no model key and does not automatically enable learning or import transcripts. For future jobs, engineers can retain their current project, instructions, skills, MCP tools, and sign-in, then optionally enable self-improvement and review proposed learning in Learning Dojo.

Important limitations include the added model usage from reflection and grading. Docker-based sandboxes build their image on first use; without Docker, agents can still read and write files but receive no shell. Remote connections currently require the server and Python connector from the same reviewed checkout, while the published CLI is observation-only. OpenCode also requires its opencode-ai package beside the server.