How it compares
Four products put an AI agent to work in your repository. They differ on the things that decide whether one fits you: where it runs, who reviews the work, who merges it, and what you pay. Every cell below links to where it is stated. Where a vendor’s own pages do not say, the cell says “not stated” rather than guessing.
The comparison
| FULCRUMAXE | Devin | GitHub Copilot coding agent | OpenAI Codex | Cursor Cloud Agents | |
|---|---|---|---|---|---|
| How you give it work | A GitHub Discussion in your repo. faq | Web app, Slack or Teams, Linear, Jira, API, CLI, or /devin on a PR. source | Assign an issue, the agents panel, @copilot on a PR, Slack, Teams, Jira, Linear, VS Code. source | chatgpt.com/codex, @codex on a PR, Slack, Linear, the CLI. source | The Cursor app, cursor.com/agents, Slack, @cursor on GitHub, Linear, API. source |
| Where it runs | Claude Code sessions on your own machine. safety | Planning and models in Cognition’s cloud; the dev environment in their cloud or a dedicated VPC. source | GitHub Actions runners. source | OpenAI-managed containers; the CLI runs on your machine. source | Cursor-managed virtual machines. source |
| Self-hosted? | Yes — there is no hosted service. safety | Partly: Outposts run commands on your machines; inference and planning stay in Devin’s cloud. source | Self-hosted Actions runners are supported; the service is GitHub’s. source | The cloud agent, no. source | Partly: your own workers run commands; Cursor runs the agent loop and the models. source |
| What it hands back | Pull requests, written in isolated worktrees. safety | Pull requests. source | Commits on a copilot/ branch and a draft pull request. source | A diff; you choose to open a pull request. source | A branch and a draft pull request. source |
| Who merges | Its own loop, and only once the gates the change needs are present. Keep branch protection on. safety | Human engineers review and choose; Devin can merge where branch protection allows it. source | A human: Copilot cannot approve or merge. source | You: “review before you merge”. source | A human reviews before anything merges. source |
| Review built in | Separate code, security, acceptance and browser reviewer agents; each gate is a label. safety | Devin Review: a bug catcher and security scanning. source | CodeQL, a dependency advisory check, secret scanning and Copilot code review, on by default. source | Codex code review on pull requests, flagging high-priority issues. source | Bugbot, a separate product billed by usage. source |
| Cheapest plan with the agent | Free software; you bring a Claude subscription or API key. cost | Pro, $20 a month. source | Copilot Pro, $10 a month. source | ChatGPT Plus, $20 a month. source | Pro, $20 a month. source |
| How usage is paid for | By your own Claude account. Our run measured a median of $29 per merged pull request. how | A quota, then prepaid credits. source | Actions minutes plus AI credits, usage-based since June 2026. source | Limits per five-hour window and per week, then credits. source | API pricing for the chosen model, with a spend limit set first. source |
| Models | Claude, through Claude Code, on your subscription or API key. start | Not stated for the cloud agent. | A choice including Claude, GPT and Gemini models. source | GPT-5.6 Sol in the cloud, which needs a ChatGPT sign-in. source | A curated selection, including Cursor’s own. source |
| Open source | Yes, AGPL-3.0. licence | Not stated. | Not stated. | The CLI is Apache-2.0; the cloud agent is not open source. source | Not stated. |
Where the others are stronger
They are hosted: nothing to keep running on a machine of your own, and work continues when your laptop is shut. They take work from where teams already talk — Slack, Jira, Linear, the editor — rather than from GitHub Discussions alone. They are products with support behind them, where FULCRUMAXE is an experimental open-source project built for one developer. And they choose from several model vendors, where FULCRUMAXE runs on Claude.
Where FULCRUMAXE is different
Review is the core of it rather than an add-on. A change is written by one agent and has to pass separate code, security, acceptance and browser reviewers, recorded as labels the merge step reads. Because of that, its own loop merges the work once the gates are present — a deliberate trade-off the others make the other way, by leaving the merge to a person. How the gates work, and where they fail.
It runs on your hardware and your own Claude credential, the source is AGPL-3.0, and the team’s own work is public: every pull request it merged is on how the work ships, and what it cost is measured rather than quoted on the cost page.
It is the wrong choice if you want a hosted service, work from an editor or a chat tool, need vendor support, or cannot keep a Linux machine running the loop.