Your host. Your keys. Your repo.
Fragua is the opposite of a done-for-you platform. It's opinionated, Rails-deep, and self-serve — agents run on a host you manage, with your own credentials, and we never sit between you and your provider. The orchestration is ours; the keys, the code, and the repository stay yours.
Orchestration, the live timeline, cost roll-ups, the audit trail. Run it as managed SaaS or on your own hardware.
Your gh and Claude credentials, your repo, isolated git worktrees. Your keys and code never leave it.
Tokens billed straight to your account. Fragua can't see the bill.
AI traffic runs host → provider directly. Fragua is never in the path.
A control plane you watch. A host you control.
Fragua is built around the two-plane split above. Orchestration and observability live in the web app; the agents — and the credentials they use — live on a host you run. A connector on your host picks up the work, runs it locally, and streams back only the events and cost. The web app directs; your host executes, and your access and your AI stay on infrastructure you own.
You register a machine by running fragua login on it. The web app sees it come online, bind to its workspaces, and report capacity — and you can forget it at any time.
Connected machines
2 machines-
daniela-mbp.local Onlinelast seen less than a minute ago · 1 / 3 sessions · darwin-arm64 · v1.0.0 · runner 2.1.170almighty-monolith marigold
-
mateo-studio.local Onlinelast seen 2 minutes ago · 0 / 3 sessions · linux-x64 · v1.0.0 · runner 2.1.170nimbus
- Runner name
- daniela-mbp.local
- Runner uuid
- 03ga8dtatew7a6mmqt2a9w1lh
- Control plane
- https://fragua.app
- Workdir
- ~/fragua-workdir
- Concurrency cap
- 3
- Authentication
- token present
- Daemon
- running (since 2026-06-10T13:54:10)
- Server reachable
- true
- Active sessions
- 1 / 3
- Pending reports
- 0
- Errors
- 0
- Managed workspaces
-
brujula-6396/almighty-monolith (cloned)brujula-6396/marigold (cloned)
We never sit between you and your AI provider.
Agents run through the provider's own CLI on your host, using the credentials already configured there. Fragua orchestrates the run and parses the usage — it never carries your AI traffic.
Your AI calls go host → provider directly. Fragua isn't in the path.
Fragua never holds your key — the CLI reads it on your host.
Tokens bill straight to your provider account. We can't see your bill.
Agents run on a host you manage; only events and cost flow back.
Pushes and pull requests use your own access.
When an agent commits work, pushes a branch, and opens a pull request, it does so through the GitHub access already present on your host — your gh and git credentials. Fragua doesn't hold a persistent token to your repositories.
The code stays in your repositories the whole time. Every network step soft-fails: if a push or a PR can't complete, the committed work stays safely on disk in its worktree rather than disappearing.
Each feature runs in its own git worktree on its own branch — work on different features never overlaps.
Agents run in a directory outside the Rails app, sharded per account and workspace — never inside Fragua's own tree.
The agent process is launched with Fragua's own Bundler context scrubbed, so it runs against your project's toolchain, not ours.
A scoped callback token — never your AI key.
When an agent needs workspace context — like the Knowledge Base — it calls Fragua's own API with a per-user token that's stored hashed, scoped to that user, and revocable at any time. It's a Fragua-internal credential: it never touches your AI provider, and it carries none of your code or keys.
Self-serve, on your terms.
Fragua is in private beta. A human reads every access request before a seat opens.