Handing an AI agent a GitHub issue and hoping for a solid pull request back is still a gamble — it might misread the ask, skip tests, or ship something nobody actually reviewed. Two products approach that gamble from opposite directions. AgentMachinist is a free, open-source, local-first CLI that gates every step behind an approval tied to a specific commit. Devin is Cognition's autonomous cloud engineer, built to run further ahead on its own — planning, coding, testing, and opening the PR — with enterprise customers and a Windsurf acquisition behind it. Here's how they actually differ.
The short version
AgentMachinist runs on your own machine, costs nothing, and forces a disciplined specification-approval-testing pipeline before a pull request ever gets its green light — the tradeoff is that it's a CLI tool built for a solo developer or small team, with no hosted dashboard. Devin runs in the cloud, plans and executes multi-file work with far less hand-holding, and is aimed at enterprise teams doing large-scale migrations — the tradeoff is opaque, sales-negotiated pricing and higher-stakes autonomy where mistakes can be more consequential and harder to catch mid-task.
AgentMachinist: a disciplined pipeline you control
AgentMachinist takes a GitHub issue and walks it through specification, implementation, and testing, then gates the resulting pull request behind an approval step tied to a specific commit (SHA-bound) — so what got approved is exactly what ships, not a slightly-different version that snuck in after review. It's local-first, running on your own machine rather than a cloud service, and it works across several AI coding harnesses (Claude, OpenCode, Pi, Codex) instead of locking you into one vendor's agent.
It ships as a CLI (the machinist command) with task queue controls — pause, resume, defer — plus a repository registry for managing multiple projects and diagnostic tooling. It's free, open source, and installable via PyPI.
For who: solo developers or small teams on GitHub who want a disciplined issue-to-PR pipeline without handing full autonomy to a black-box cloud service.
Price: free, open source, no cost.
Strengths: SHA-bound approval is a real checkpoint before code ships, not a formality; local-first and compatible with multiple AI coding harnesses; free and open source; task queue controls for pausing, resuming, and deferring work.
Limits: CLI only, no hosted dashboard; built more for a solo developer than a large team; you have to configure and manage your own AI harness rather than getting one bundled in.
Devin: autonomous, cloud-hosted, enterprise-facing
Most AI coding tools work turn by turn — you ask, it suggests, you decide what happens next. Devin was built to run further ahead on its own: given a task, it plans the approach, writes and tests code across a whole project, reviews its own work with visual QA, and can carry a piece of work from a ticket all the way to an opened pull request, checking in with a human at key decision points rather than at every single step.
Devin handles autonomous multi-file coding, testing and deployment, automated PR review with visual QA, large-scale code migration and refactoring, auto-generated documentation and system diagrams, issue triage and bug fixing, and multi-repository project management — integrating with 50+ tools including GitHub, Linear, Slack, Datadog, and AWS. It's available via cloud, desktop app, CLI, or a Windows VM, and Cognition's 2025 acquisition of Windsurf extended its footprint from a standalone agent into IDE-level tooling as well.
For who: enterprise software teams and large institutions doing code migration at scale, who want an agent that runs ahead with less oversight and are willing to pay for it.
Price: freemium — a free trial is available, but paid tiers aren't published; enterprise pricing requires a sales conversation.
Strengths: runs further autonomously than a suggestion-based assistant — plans, codes, tests, and can open a PR unsupervised; owns both a standalone agent and a popular IDE (Windsurf); enterprise partnerships with risk-averse buyers like Goldman Sachs and Mercedes-Benz; built for large-scale migration and refactoring, not just small suggestions.
Limits: higher-stakes autonomy means mistakes can be more consequential and harder to catch mid-task; its efficiency claims (8x, 20x) come from single-customer case studies, not guaranteed universal results; no published pricing means budgeting requires a sales call.
Side-by-side
| AgentMachinist | Devin | |
|---|---|---|
| Where it runs | Local, on your machine | Cloud (also desktop app, CLI, Windows VM) |
| Approval model | SHA-bound approval before shipping | Human checkpoints at key decision points |
| AI harness | Bring your own (Claude, OpenCode, Pi, Codex) | Built-in, own agent + Windsurf IDE |
| Price | Free, open source | Freemium, enterprise sales for paid tiers |
| Target user | Solo developer / small team | Enterprise team, large-scale migration |
| Dashboard | CLI only | Cloud dashboard, IDE, desktop, CLI |
Which one should you pick?
Pick AgentMachinist if you're a solo developer or small team who wants a free, local tool that forces a real approval checkpoint before AI-generated code ships, and you're comfortable running it from a CLI with your own choice of AI harness.
Pick Devin if you're an enterprise team tackling large-scale migrations or refactors, you want an agent that can run further ahead with less step-by-step supervision, and you're prepared to go through a sales conversation to find out what that costs.
The honest read: these aren't really competing for the same buyer. AgentMachinist is a control-first tool for developers who don't want to give up the review step. Devin is a scale-first tool for organizations that want to delegate more and are willing to pay enterprise prices — and accept enterprise-grade risk — to get it.