The seams, and what each one makes possible
Code shape: generate the plumbing instead of writing it
A forge project has two sources of truth: protobuf for the API and SQL migrations for the schema.forge generate projects them into handlers, the ORM, typed frontend hooks, pages and tests, so an agent spends its tokens on your logic rather than on wiring it. The layout is fixed, so an agent asks the project where things are (forge project shapes) instead of grepping for them.
- Two sources of truth: what is generated from proto and migrations
- Adding a feature: the everyday generate, lint and test loop
- Contracts and dependency injection: generated mocks and tracing for every package boundary
Architecture: lint for whole classes of bugs
Because forge knows what every RPC, table and boundary is,forge lint can check things a generic linter can’t: a list filter that silently filters nothing, a server-owned field nothing writes, a trace cut by a missing context, a migration that rewrites a table under an exclusive lock. Every migration also declares whether the previous release survives it, and lint holds it to that. New rules warn for a release before they gate, so upgrading forge never turns your build red overnight.
Releases: gate on measured health, never past a breaking migration
A release is built once and promoted by digest, so prod runs exactly the bytes staging ran. After every deploy,forge obs check judges the release against the window before it, per service, and answers with an exit code a script or an agent can branch on. Missing telemetry is reported as undetermined, never as a pass. With the opt-in health gate, prod promotion is refused until staging holds a passing verdict, and a rollback floor stops anyone deploying a release older than a breaking migration. There is no automatic rollback: a failed deploy never replaces what is serving, and you recover by deploying an earlier release or rolling forward.
- Release promotion: releases, plans, approvals and the ledger
- Release gates and regression checks:
obs check, the health gate and rollback floors - Multi-environment deploys: dev, staging and prod from the first commit
Runtime: tie an error to the release that caused it
Every boundary of a forge app records logs, traces and metrics from the first commit: inbound RPCs, component calls, database queries, outbound HTTP and the browser. Errors are grouped into issues and tied to the release that opened them. Your agents read all of it through plain commands (forge obs status, obs issue, obs trace, obs logs) with JSON output, and read your database through one audited, read-only command that never shows secret columns.
- What every boundary records
- Dashboards, alerts and issues
- Querying with forge obs
- Why this helps an agent
Operations: hand an alert to an agent with the evidence attached
On Reliant hosting, a workflow trigger declared aswebhook: {source: alerting} receives your organization’s critical alerts and new or regressed error issues with no URL, token or alert route to set up. The built-in Autopilot on-call workflow uses it: each alert becomes one deduplicated GitHub issue, investigated read-only with the traces, logs and data the agent can reach. When a person labels the issue ready, the issue worker writes the fix and opens a pull request. Autopilot never merges; you do.
- Your organization’s alerts as a trigger
- Autopilot: review, issues and on-call
- AI workflows in YAML: mix agents with deterministic steps, approvals and tests
Secrets: used by agents, never seen by them
Hosted secrets are write-only: you can set and rotate one, but never read it back. Integration steps in a workflow use saved credentials for GitHub, Gmail, Slack, Twilio and any HTTP API without the agent ever seeing the value.Opinionated defaults, no walls
Every default is a starting point, not a lock-in:- Start anywhere. Use any piece on its own, in any repo and any language. An existing app deploys from its own Dockerfile, and monitoring or release recording works without forge deploying anything. See adopt an app that already exists.
- Bring your own. Point an environment at your own database, identity provider or OpenTelemetry backend in its config. See bring your own.
- Deploy where you want. Reliant hosting, your own Kubernetes cluster, Docker Compose or plain host processes. See self-hosted delivery.
- Any model. Use your Claude, ChatGPT or Copilot subscription, your own API key, or local models, and mix them per workflow step. See API keys and subscriptions.
- Try it without an account. Deploy a Vite or Next.js app with no signup and claim it later.
Where to go next
Reliant quick start
Run your first agent and workflow
Forge quick start
A prompt to a deployed URL
Hosting
What Reliant hosting runs