Terragon

There is nothing to sign up for. The hosted service stopped, the sandboxes were terminated and what remains is a source snapshot, so the deciding fact in daily use is that there is no daily use. If you liked the shape of it, delegating a task and receiving a branch, that shape exists elsewhere and with somebody answering the phone.

Pick Codex Cloud if you want cloud agents that open pull requests, or Devin if you want the most complete autonomous stack and can bring a budget alarm. Read this repository if you are building something similar and want a worked example.

An as-is release is not a product. The documented self-hosting path lists seven separate dependencies before the first task runs, and every one of them is an operational commitment with its own failure modes and its own bill. There is no maintainer to file an issue against and no release to pin. What was handed over is the architecture, not the service.

What it did right while it ran: each task got an isolated container with its own copy of the repository and its own branch, so agent work never shared state with anything a human was holding.

Two design decisions are worth preserving. 1. Output arrived as commits and a pull request on a branch, which means the acceptance step was the review process the team already trusted and already staffed, rather than a bespoke approval interface nobody reads twice. 2. Work was triggered by events, including mentions in chat and on the forge, so the agent was invoked from where the request naturally appeared instead of from a separate console.

No evaluation was ever published. The interesting contribution was the integration surface, and that survives as source for anyone rebuilding this pattern.

The billing structure explains the ending. Selling the metered component at cost is generous to customers and fatal to margins, because it puts every dollar of gross profit on a modest subscription while the expensive part of the product scales with usage. Add a trial that gave away credit up front and the unit economics were negative on exactly the customers who used it most.

The wind-down was orderly, with subscriptions cancelled and outstanding balances refunded, which is better behaviour than this category usually manages. No acquirer took the asset; the code was released instead. Position: closed, and the pricing page is the lesson.

The historical arithmetic is simple and academic: sixty seats at the professional tier came to $3,000 a month before any usage. What matters now is that there is no vendor, so there is no contract, no support terms, no retention policy and nobody to send a security questionnaire. Self-hosting turns a subscription into a staffed internal service, which is a headcount decision dressed as a licence saving.

Onboarding cost is irrelevant when there is nothing to onboard onto. Not yet, and there is no version of later. Anything still pointing at this should be disconnected.

Releasing the whole thing under a permissive licence when the lights went off is the correct way to leave, and I want to say so before I complain. The snapshot includes the command-line handoff tool and an MCP server, so the pieces I would reuse are the pieces that were designed to be talked to rather than clicked.

What I would actually fork is small: the trigger plumbing and the branch handling. Standing up the rest is somebody's full-time job, and it will not be mine. Grudging respect for a company that let the code outlive the company.