Palette
Mode
Or one accent
Surface
Saved in this browser. ⌘K has them too.
By what you are building
One-click SaaS
Open code, built and deployed in one click
/agents
An agent is what it reads, what it may touch and what wakes it: AGENT.md, its skills and rules, a model, the tools it is granted one by one, and its triggers. It versions like code, because it is files.
/how-agents-work
An agent calls the same tools your people do — your systems, the Registry, the ledger — through the same gate, with the grants you give it.
A worker does one job — triage an inbox, reconcile payments, write the Monday digest. A harness routes work to other agents, each keeping its own grants: the harness never borrows them. A workflow is a harness with an ordered plan — test, staging, verify, production — so workflows need no separate engine.
An agent can call only what it was granted: read tools per system, and every tool that writes, destroys or reaches outside one by one — never a wildcard. Granting a write needs its own permission, and you can only grant what you could do yourself. On top of its grants, an agent can never do what its owner cannot.
An agent wakes on a schedule, a signed webhook, a platform event — a crash, a failed deploy, a failed backup — a mention, or another agent calling it. Anything that leaves your company is drafted and waits for a person unless the owner chooses otherwise and the organization’s policy allows it.
A new agent starts as a draft, and it cannot be activated until a test run of its current files has finished. A test run executes reads and shows what it would have written, destroyed or sent — without doing any of it. Changing its files or grants sends it back for another test before the change takes effect.
Every run is a trace: what woke it, what it read, each tool call with its decision and outcome, what waited for a person, and what it cost. The calls are rows in the same ledger your people’s calls are in.
/faq
Something else? Write to [email protected] or use the contact page — a person answers.
The same pieces as the rest of the platform: systems with their tools, the Registry, the ledger of calls, secrets and permissions.
Only if its owner chooses “acts” and the organization’s policy allows it. By default anything that leaves the company is drafted and waits for a person to approve it.
Third-party MCP servers bound to an account — a mailbox, a chat workspace, a payments account — behind the same gate as your own systems. The platform sends the connector’s own token, never the caller’s.
The one its files name, from the models your organization allows, paid from prepaid AI credits with per-run and per-day caps.
/start
Publish your systems and your agents call exactly those tools, through the same gate. Request early access to start.
