What is AgentRun?
AgentRun, from Parcha Labs (the team behind Grep.ai), sits around the coding-agent harnesses people already use, Claude Code, Codex, OpenCode or pi, and adds what they lack for repeatable work: a store that survives crashes, per-tool approvals, spend caps, triggers and one console for every run. Not to be confused with Alibaba Cloud's AgentRun, a separate commercial product.
Its core is a workflow language. A workflow is a JSON document of typed steps: tool calls, model calls, fast typed decisions (yes or no, pick one, a score, each with a confidence) from a decision model such as Jev, branches, bounded loops, parallel maps and a fallback to a full agent when a case is unclear. The console shows each run node by node, with every decision and its reason, and it can be paused and resumed. A learning mode lets an agent keep notes, skills and eventually a learned workflow in its own git folder after each run.
The workflow package runs on its own: npx @parcha/agentrun-dsl@latest demo plays a support-triage example with scripted judgments, needing no model SDK, API key or network. The full server starts with npx @parcha/agentrun on port 7340 and needs one model key for real runs. It is Apache-2.0, first published on 23 September 2026, and still in beta (v0.1.0-beta.37 on 14 November 2026).
Is AgentRun worth it?
AgentRun's argument is that a lot of agents should not be agents. A job an agent has done a hundred times, such as screening a company or reviewing a pull request, can be written as a typed workflow in which cheap typed decisions route the easy cases and a full agent is called only for the unclear ones. The console then shows every step and every decision with its confidence, which makes a run auditable in a way a long agent transcript is not.
The operational layer is more complete than most projects this young. Runs live in SQLite or Postgres and resume after a crash; with several runners, a dead runner's runs are picked up after a 30-second lease, and a write cut off mid-call is marked interrupted rather than retried. Each tool is allowed, asks first or is denied; a run with a cost cap pauses at the cap instead of failing; secrets are sealed under one key and never enter the agent's computer. Five templates (code review, KYB due diligence, market monitoring, lead scoring, release notes) come with their own eval cases.
The headline numbers are self-reported. The README says its tuned code-review template costs 18 cents a review against a Greptile bill that rose from $180 to $6,451 a month, and caught 1.6 times the real bugs on 30 unseen pull requests; a merchant-classification job fell from $0.89 to $0.12 per merchant once learned. No independent replication exists yet. On Hacker News (51 points) the main objection was that Temporal, Windmill or Airflow already run workflows, to which the answer is the agent-specific layer: harness integration, typed model decisions and learning.
Risks: the project went public on 23 September 2026 and ships near-daily betas, with 70 open issues and pull requests. The learning mode lets runs write notes and workflows that later runs execute; the code filters secrets, out-of-folder files and rules lifted from web pages or emails, and a grader model judges the rest, but that is still model-written behaviour persisting between runs. "Never ask" mode lets an agent run unattended, and the Docker Compose deploy binds to loopback only and needs a TLS proxy before exposure.
Where it earns its place
A compliance team at a 50-person fintech
Run KYB checks on new business customers with an audit trail
Apply the KYB template: registry and certificate checks run in parallel, a typed decision approves, rejects or sends to review below a 0.75 confidence bar, and every claim in the report is cited. Each run can be capped in dollars.
An engineering lead at a startup paying per review for AI code review
Bring code review cost down without losing real bug catches
Tune the code-review template on your own pull requests, as the makers did; they report 18 cents a review. Trigger it from GitHub pull request events on a self-hosted server, about $27.50 a month on Render.
A developer evaluating agent workflow engines
See how a typed workflow behaves before giving it a model key
Run npx @parcha/agentrun-dsl@latest demo, then dry-run your own workflow with synthetic answers; neither calls a model, so the control flow can be checked for free.
Before you commit
- Setup
- npx @parcha/agentrun starts a server and runner on port 7340, stores data in ~/.agentrun and opens the console; set a password, add one model key (OpenRouter, Anthropic or OpenAI) and apply a template. Docker gives agents a sandboxed shell.
- Integrations
- Harnesses: pi, Claude Code, Codex, OpenCode. Computers: Docker, Daytona, Modal, E2B sandboxes, Linux desktops, cloud Android phones, Browserbase or Kernel browsers. Remote MCP servers, Composio toolkits and triggers, schedules, signed webhooks, GitHub events; TypeScript and Python clients.
- Data and privacy
- Self-hosted: runs, files and journals stay in your SQLite or Postgres store. Secrets are sealed under AGENTRUN_SEAL_KEY and shown only as set or missing; model calls go to whichever provider key you add.
- Cost at scale
- The software is free; spend is model tokens plus any remote sandbox or browser vendor, each billed to your own account. Per-run caps stop a loop from running up a bill.
- Limits
- Beta software with 70 open issues and pull requests; runners on Postgres still share one data volume in this release; iPhone, macOS and Windows computers are not yet supported.
What’s good
- Typed workflows with confidence-scored decisions and an agent fallback
- Runs survive crashes; interrupted writes are flagged, not retried
- Per-tool approvals, cost caps that pause rather than fail, sealed secrets
- Works over Claude Code, Codex, OpenCode or pi
- Workflow demo runs with no API key or network
What’s not
- Beta since September 2026, with near-daily releases
- Cost and quality claims are the makers' own
- Learning mode persists model-written notes and workflows between runs
- Needs Node 24.22.3 or newer on macOS or Linux
Best for
Teams running the same agent job many times a week who want each step visible, costs capped and the routine cases moved off the expensive agent loop.
AgentRun alternatives
AgentRun: common questions
- Is AgentRun worth it?
- AgentRun makes a persuasive case that most agent jobs should become checkable workflows, and it ships the approvals and caps to back it, but it is a 52-day-old beta whose cost claims come only from its makers. We rate it 3.8 out of 5.
- How much does AgentRun cost?
- Free under Apache-2.0; you pay your model provider and any remote sandbox vendor. The one-click Render deploy costs about $25/month for the instance plus $2.50 for the disk, per the README.
- What are the downsides of AgentRun?
- Beta since September 2026, with near-daily releases; cost and quality claims are the makers' own; learning mode persists model-written notes and workflows between runs; needs Node 24.22.3 or newer on macOS or Linux.
- Who is AgentRun best for?
- Teams running the same agent job many times a week who want each step visible, costs capped and the routine cases moved off the expensive agent loop.
- How much does AgentRun cost for a team or in production?
- The software is free; spend is model tokens plus any remote sandbox or browser vendor, each billed to your own account. Per-run caps stop a loop from running up a bill.
- Is AgentRun safe for company data?
- Self-hosted: runs, files and journals stay in your SQLite or Postgres store. Secrets are sealed under AGENTRUN_SEAL_KEY and shown only as set or missing; model calls go to whichever provider key you add.
- What are the best AgentRun alternatives?
- The closest alternatives we have reviewed are Temporal, LangGraph and Mastra.
Answers checked Nov 14, 2026.