The Factory for Professional agents
Design, test, govern and run agents for whole teams without writing code, in your cloud or on your own hardware.
From a blank page to a live agent in about ten minutes.
Every part of an agent is a screen in the studio, and the studio checks it as you go.
- 01
Create
Start from a template or a blank agent. It is running in about 30 seconds.
- 02
Teach
Attach skills and knowledge bases. Upload your documents and build them into a wiki.
- 03
Govern
Choose the tools it may use, the calls that need approval and the sites it may browse.
- 04
Test
Chat with it in the studio and see every tool call, file and decision.
- 05
Publish
Put it on Slack, Teams, WhatsApp, the phone or your website.
The Copilot can take any step for you.
It asks what the agent is for, shows a plan, and builds only what you approve. Anything that reaches people or money needs its own yes.
Run all of it inside your own network.
On-premise or air-gapped, the whole platform runs on your Kubernetes with your models, and no data has to leave your network.
Open-source runtime
The runtime is on GitHub. Your security team can read every line.
Your own models
Surogate 3.7, Jackrabbit and Amami run on your GPUs, next to your fine-tunes.
Your own hardware
Train and serve on your servers over SSH, or on your own cloud account.
Your keys and records
Credentials stay in your vault. Sessions, files and logs stay in your storage.
Every part of an agent, on its own screen.
Eight tabs describe an agent. Change one and the next session uses it, with no rollout.
The soul: who the agent is, how it speaks, what it refuses. Screened for prompt injection.
Procedures in markdown or drawn on the canvas. Every version is kept, and each session records the one it used.
Your documents as a searchable wiki, with the source passage behind every answer.
MCP servers and more than 1,000 Composio toolkits, scanned before the agent sees them.
Ours, your fine-tunes, OpenRouter or any OpenAI-compatible endpoint.
The browser, Claude Code and Codex, deep research, goals, missions and schedules.
Repositories the agent can branch, change and open pull requests on.
Allowed tools, approvals, browsing rules and the EU AI Act notice.
The runtime you would otherwise build yourself.
Sessions, sandboxes, credentials and rules are part of the platform, and the runtime is open source.
It keeps working
Sessions that survive crashes
If a worker dies mid-turn, another resumes from the log.
Saga rollback
When a step fails, finished steps are undone in reverse order.
A browser you can take over
Take control for a login or a CAPTCHA, then hand it back.
Office files, edited live
Word, Excel, PowerPoint and PDF beside the chat. Your edits win.
It stays inside the rules
Keys the agent never sees
The proxy adds credentials on the way out.
Rules fixed for each turn
Text injected mid-turn cannot loosen the tool policy.
Approvals
Sensitive calls wait in the inbox. A grant covers one call, for one hour.
Checked tools
Definitions are scanned for hidden instructions and fingerprinted against changes.
Work that runs for hours, days or months.
/goal
Works until a separate evaluator, reading the evidence, says the goal is met.
/mission
Splits an objective into tasks for sub-agents and judges the results against your rubric.
/loop
Runs a prompt every few minutes, hours or days.
/deep-research
Returns a cited report, researched and written by two sub-agents.
Sub-agents that share notes
Parallel helpers write to a board of checked notes, so a dead end one of them found is not repeated by the next.
Tasks that survive crashes
A task is stored and retried up to three times, and each retry reads what went wrong before.
Programs: the agent writes first.
Enrol people in a journey and the agent follows up on the days you choose, on WhatsApp, Slack, Telegram or by phone. It writes each message, reads the replies, and alerts your team when a person is needed.
One agent on nine channels.
A person who moves from the web to Slack or WhatsApp keeps the same conversations, memory and allowance.
Web app
Always on, with your agent's name and logo.
Website widget
A chat bubble on your own site.
API
OpenAI-compatible. Every call runs the whole agent.
Slack
Channels and DMs. It can follow a channel and speak up.
Microsoft Teams
In your own Microsoft tenant.
Templates, quick replies and native forms.
Telegram
Groups, threads and private chats.
Discord
Your own bot in your servers.
Phone
Inbound calls, in the voice you choose.
Agents that improve from their own sessions.
Mark sessions as good or bad, and the studio turns them into training data, fine-tunes, scores and experts.
Skill optimisation
Rewrites a skill against generated tasks and keeps the best version. Nothing changes until you apply it.
Experts
A skill backed by its own fine-tuned model, which the agent consults in a bounded loop.
Any GPU
Connect AWS, Azure, GCP, Nebius, RunPod, Modal, Oracle, Vast.ai or your own servers.
How it compares.
Agent frameworks
LangChain, LangGraph
You write the agent in code, then build sessions, sandboxes, channels and governance around it.
Those come with the platform, and agents are built in the studio.
Coding agents
Claude Code, Codex
One person, on one machine, in one terminal.
Whole teams on nine channels, with governance and a record of every step. Claude Code and Codex can run inside a session.
Managed agent platforms
The vendor builds and runs the agents on its own cloud.
Your team builds them, in your cloud or air-gapped, with any model.
Sell an agent you built.
Put a price on it and share a link. Buyers do not need a Surogate account.
Your Stripe account
Subscriptions, usage packs or both, in seven currencies. You are the merchant of record.
Pages made for you
A buy page at your agent's address, and a landing page you assemble from blocks.
Packages
Each offer unlocks the channels, tools, knowledge and model tier you choose.
Build your first agent.
Start on the free plan with five agents and no card, or install the platform on your own Kubernetes.