Type one sentence. Relay reads the windows you have open, writes a plan you can edit, drives the real apps, checks every step, and stops before anything consequential. Then it remembers how, so next time costs zero model calls.
Recorded from the Relay app. Demo company and data; the windows are stand-ins so nothing private is on screen.
The job: export the open invoices, reconcile them against the ledger, write Finance a summary, and don't send it without me. Here is every stage of that run, frame by frame from the app.
Relay tags the intent as you type (a loop, a URL, an email) and notices it has done this before. A saved, verified skill is one click away, with no re-planning and no model calls.
Relay snapshots the windows you track, injects what it already knows about them, and compiles a structured mission: phases, targets, steps, failure policies, and approval gates. Edit any step inline. Nothing has run yet.
Native clicks, keystrokes, and focus changes across the browser, Explorer, the spreadsheet, and mail. Each step targets a control by name, so a button that moved doesn't break the run.
Sending is consequential, so the plan carries an approval gate. Relay pauses with the draft on screen and waits for you. Approve and it finishes. Cancel and nothing was sent.
Every step gets a verdict. Cheap UI-Automation probes run first; vision only steps in when needed. The run report shows what passed, what was skipped, and why, so a run that only looks finished never becomes a saved skill.
The exact mission from the recording, step for step. Press run to replay it: watch verdicts land, the windows change, and the gate wait for you.
Each window, before the run and after it. The export landed in the right folder, the ledger flagged two problems, and the email sits drafted and unsent.
Relay added the portal amounts by invoice id and a Reconcile column. Six rows tie out. Meridian Freight is off by $630.00 (likely transposed digits) and Tessellate Design is missing from the ledger.
Relay doesn't just click. It captures what it finds into the Data Vault, so the next run, a watch, or another mission can use it as {{recon.summary}}.
| Vendor | Invoice | Ledger | Portal | Result |
|---|---|---|---|---|
| Northwind Supply | INV-20918 | 2,480.00 | 2,480.00 | ✓ match |
| Bluefin Logistics | INV-20922 | 6,912.40 | 6,912.40 | ✓ match |
| Kestrel Print Co. | INV-20927 | 1,145.00 | 1,145.00 | ✓ match |
| Meridian Freight | INV-20931 | 4,180.00 | 4,810.00 | ✕ Δ $630.00 |
| Halcyon Cloud | INV-20935 | 9,600.00 | 9,600.00 | ✓ match |
| Orchard Office | INV-20938 | 782.20 | 782.20 | ✓ match |
| Pinecrest Facilities | INV-20941 | 8,113.00 | 8,113.00 | ✓ match |
| Tessellate Design | INV-20944 | — | 4,580.00 | ⚠ not in ledger |
ap.open_total$38,412.60ap.mismatches2recon.summary6 of 8 match · 2 to reviewwatch → ap.mismatches increasedqueued vendor follow-upsA verified run becomes a skill: a parameterized, replayable version of the job. Replays take the deterministic path first and only call a model to repair a step that drifted.
In the demo workspace, the AP skill has run 14 times. 97% of its steps executed deterministically; two repairs were learned and folded back in as new versions.
Schedule any mission or skill, with window-readiness checks, missed-run policy, and auto-pause after repeated failures. Watches fire when a captured value changes.
Live mode and the macro recorder capture what you do, anchored to UI elements rather than pixels, and turn it into an editable mission.
Every screen below is the shipping app, populated with demo data. Click a screen to enlarge it.
The relay CLI and MCP server let Claude Code, Cursor, Codex, VS Code, or Gemini CLI list windows, read text, click named controls, and run your saved skills. You decide each agent's tier, and consequential steps still wait for a human.
Describe the job, inspect the plan, approve what matters, and keep what works. Relay runs on your machine with the models you choose.