Your assistant builds them. orangeforge.ai tests, versions and publishes them — without taking away your files' role as the source of truth.
Every change leaves artifacts your team can review, compare and build on.
The agent you publish can call on other specialized agents, give them a short step or a longer-running mission, then bring their results together. Every delegation stays bounded and visible in your traces.
your ai coder
Codex, Cursor, OpenCode, OpenClaw, Hermes or your favourite assistant work just as well: the Forge skill and CLI give them the complete method.
Your assistant:
Output schema, custom tools, prompts, the best-suited model for each phase. Your assistant puts together a real precision instrument, not just a prompt.
It runs your defined cases on the real infrastructure, reads the traces, fixes issues and reruns the bench. Stored verdicts show what passes and what still needs work.
Once the bench is green, it publishes. Your agent is live on Cloudflare's global network, behind one URL and one API key.
Build your agents on the Claude Code plan
you already pay for.
The long crafting loops run on your subscription, not on platform credits.
Start for freemacOS · Linux · no credit card
our stance
Each of these deserves its own article, they are coming to the blog. Meanwhile, the documentation goes into detail.
the process
Scroll: the agent builds itself on the left, at the pace of your reading.
01
One command: magic-link sign-in, an agents/ folder created, kit and skills installed. Your AI agent now knows the whole cycle.
02
Your assistant (Claude Code here, but the skill is the same for Codex, Cursor or OpenCode) applies the built-in methodology: output schema first, declared tools, a golden set of test cases. In the end you review files.
you → claude code
“Create an agent that answers questions about our opening hours, with strict JSON output.”claude code
I created agents/opening-hours/ : agent.ts, prompt.md, output.schema.json and 5 golden cases. Pushing now and running the bench.03
The golden set runs on the real infrastructure, every run traced (tools, tokens, cost, latency). No production release without proof.
04
One command publishes the proven version in an isolated Worker, with its secrets and URL. It streams from Cloudflare's global network.
05
The same published agent answers in Chat, over HTTPS, as an MCP tool or as an A2A agent. Try it here: edit the question and get a live production response.
Understand the four access methodsHello. Ask me about the shop's opening hours.
Loading…
MCPLoading…
06
Every push freezes a complete version, with the prompt, the code and the config. Every run knows which one produced it. Compare, pin, restore in one command.
Execution traces, golden sets, versions, secrets: full details in the documentation.
new — missions
Your agents answer in seconds. Mission agents take their time: you hand them a job, they settle in, do it, check their own work — and hand you finished files.
“Analyse these forty customer reviews and bring me a summary.” One sentence is enough. The mission starts within the limits of your plan and has a maximum duration. You can track its spending while it runs.
In an isolated workshop, under contract: the platform checks spending between steps and stops the mission when the limit is reached; a request already in progress can exceed it. The documents it needs are copied into the mission's isolated workshop, so it can work without general access to your systems. After an interruption, it resumes from its work log.
A report, a summary, a document: finished files, produced by an agent that reviewed its own work and tried again until it satisfied your brief.
A mission is forged like everything else: describe the job to your assistant, it prepares the agent — all that's left is to hand over the work.
trust
An agent platform touches your prompts, your data and your API keys. It has to earn its place.
Start on our discovery credit, then plug in your Vercel AI Gateway or OpenRouter key. It is only visible to the egress worker that calls your provider, the interface shows only its last four characters, and your provider bills you directly.
The CLI reads your workspace files and sends the necessary operations to the platform. Announced project: we plan to publish its source code so this path can be audited independently.
Your tool keys are set as Cloudflare secrets of the agent's worker (forge secret set), write-only: neither our database, nor the logs, nor the API can read them back. Sign-in by magic link with an anti-phishing code.
pricing
to evaluate properly, no credit card
1 user · 5 active agents
5,000 executions/month · 1 project
7-day traces · community
or EUR 990 excl. tax/year
5 users · 50 active agents
250,000 executions/month · 5 projects
90-day traces · white label · 1 Gate
Email support within 48 hours
Forge must be open. Otherwise, download the app.
contractual volumes, governance and SLA
Users, projects and agents per contract
Gates, identities and voice per contract
Negotiated SLA · partner support
Language models are billed separately, with no markup from us: your own key, or prepaid credit deducted at the price we pay for each execution. Each top-up adds a 5% fee to the requested net credit.
See full pricing →integrators
White label, a reassignable licence pool, margin on every licence and on your services. Your clients never see orangeforge.