The weekly research, the recurring digest, the sweep nobody gets to, the ticket that has waited a quarter. Describe it once in plain English and a team of agents picks it up — product, architect, developer, QA, reviewer — with an independent manager checking each step before the next one starts. The whole record lands in your git repository.
Add a capability to the product you already have — or build the product itself, on your own domain and under your own name. Either way your customers deal with you, and they never see us.
Sites, publishing and voice agents, called from your backend as your organisation, in a cloud account you own.
One POST from your backend, an HTTPS URL back. Hosting, certificate and DNS are inside the run.
Learn more → weekly, without a writerResearched, drafted, illustrated, opened as a pull request. Your customer's repository, your customer's voice.
Learn more → the call they bookedA briefed agent joins the meeting your lead booked and leaves a transcript behind.
Learn more →Atrium is the builder: describe what your customers should be able to do, and the portal is composed from parts that already run in production. It opens on your domain, and it is yours.
It shipped, and people started using it. The next job is a different job — a deploy any change can go through without you, secrets out of the repository, and the next feature still landing this week.
One declarative stack definition per service, provisioned into a cloud account you own — Amazon, Google, Yandex Cloud or your own servers. The core that does it is open source and MIT-licensed.
a team, not a promptFive roles take a ticket through to a pull request, with a management agent reading every handoff. Each change proves itself on a throwaway stack first, and a person presses merge.
nothing to unwind laterThe work, the reasoning and the verdicts land in your git repository. Self-host the open-source toolchain, or run it managed and bring your own provider keys.
An assistant speeds up a person who is sitting there waiting for the answer — so the work is bounded by their time. Forge takes the brief and does the work when nobody is there.
Perfect for a question you can answer in one sitting. You hold the thread, you steer every step, and the session ends when you close the tab.
AsynchronousBrief it the way you would brief someone who works for you, then close the laptop. It reads twenty sources, sweeps your telemetry, or takes a ticket to a pull request — on a schedule or right now — and it is still working during your commute. When you come back you get two things: the finished work, and a full account of how it was done — every step, every decision, every dead end.
The question isn't which model is smarter. It's whether you have to be present while the work happens — or can delegate it and read the report.
A team of agents runs the job. A supervisor reads every handoff and decides whether the next step gets it — so weak work stops instead of travelling.
Ours and our partners'. Each one with the evidence behind it.
A researcher gathers sources; a separate analyst re-opens each one and confirms the quote. On a live run: 18 gathered, 11 verified by direct fetch, 2 figures shipped marked unverifiable rather than passed off.
Pinned sources checked on a schedule, filtered by an analyst, posted to your channel. The publication that quietly costs an analyst-day a week — running without one.
Agents read your metrics and logs, correlate findings to the responsible code, and file a deduplicated issue — or a fix PR a human still merges. Credentials never enter the model.
Five roles with a supervisor between them take a ticket to a PR a human merges. We know it works because it is how Forge builds itself — every run leaves its branch, handoffs and verdicts behind.
The parts you meet in week two, once the first run is behind you.
A role is a named character with a voice, a tool policy of its own and a memory that outlives the run. The reviewer who caught a regression last month still knows what to look for this month.
Cron, a GitHub issue, a chat message, an inbound webhook. The run dispatches without you and asks for a person only when the supervisor escalates.
Secrets are bound to the roles that need them and injected at call time — never pasted into a prompt. Every outbound model call is logged with its provider, model and destination.
Run the agents on your own machines — a spare workstation, a rack, a private cloud — against the managed control plane, or host the whole thing. The toolchain is MIT-licensed and the runners are yours.
Five things that hold against every tool we get compared with.
Everywhere else, the agent that asked for the work also decides whether it was any good. Forge puts an independent management agent on each edge between roles: it can pass the work, send it back with specific rework instructions, or escalate to a person — before the next role builds on it.
Every run leaves a branch in your repository: the spec, the brief and each role's handoff as commits, authored by the persona. Readable with git log, months later, by someone who was not there.
No canvas, no nodes, no edges to keep in sync. Name the roles, say what done looks like, and the engine assembles the run — then keeps it working when the job changes.
Learn more → your keys, your metalSelf-hosted runners keep execution on your machines. Inference routes across a provider pool you control — priority weights, automatic failover, any OpenAI-compatible endpoint, including models you host yourself. Agents never hold your keys.
Research that verifies its own sources, a recurring digest, a content desk, an ops rotation. Same roles, same review loop, same audit trail — software delivery is one archetype, not the product.
Learn more →Each page quotes the other product in its own words, then names the thing it does not do.
The open-source toolchain on your own machines, the managed cloud, or an install inside your perimeter. The workflows are the same; what differs is who runs the infrastructure.
The whole toolchain is MIT-licensed. Run it on your own machine or in your own cloud — no account, no invoice, nobody from us in the loop. Code and data never leave your perimeter.
Workflows, schedules and the run log live on our side. The models and provider keys stay yours, and the runners can stay on your own machines.
A full install inside your perimeter: your own cloud, a private network or your own servers. Contract, SLA and named support.
One workflow — the one that builds this platform — counted from the control plane's own records.
Every run the workflow made in that window — not a sample, not a pilot, not a case study we cannot show you. A person performed every merge. OpenSSF Scorecard 10/10, MIT licensed.
If your question isn't here, write to us and we'll answer it directly.
An AI editor speeds up a person sitting in front of it: the work happens while they watch. Forge takes the whole job and runs it when nobody is there. An independent management agent checks every handoff between roles, so the result arrives reviewed rather than needing a read from scratch.
Wiring one run is not the hard part. What comes after is: handoffs between roles, an independent check on each one, role memory, access policy, model routing with failover, and a log that still shows six months later why an agent did what it did. The core is MIT-licensed, so you can size the work before deciding.
You don't have to. The runners stay on your machines, inference can point at models you host, and the toolchain can be self-hosted outright — it is MIT-licensed. Agents never hold your provider keys: secrets are injected at call time and never enter a prompt. Every outbound model call is logged with its provider, model and destination.
Three boundaries. An independent management agent reads every handoff and can send the work back for rework or escalate to a person. A change proves itself on a throwaway stack before it touches your environment. And a person presses merge — across twelve weeks that was true of all 202 pull requests.
Self-hosted is free — the toolchain is MIT-licensed. The managed cloud and the in-perimeter install are priced above: the number is on the page, with no mandatory request-a-quote step.
Yes. Inference routes through a provider pool you control — any OpenAI-compatible endpoint, including models you host yourself. Each provider has its own priority weight and automatic failover, so the pool can be composed entirely of models in your region.
Describe your workflow in plain English. Forge runs it — across SDLC, research, content, and ops.
A person reads it next and replies to the address you gave. Nothing is sent to you automatically — no confirmation mail, no list.
Nothing has been sent yet — this page has no inbox of its own. Send it as an email and a person reads it next.
Send it as an email