Check our work.
For technical evaluators. How the platform is built, what we test, the raw benchmark table, and what is not finished. The driver API, keys, and pipeline routes are in the integration guide.
The AI drives. The platform executes deterministically.
An AI model drives. It works out what a task requires and writes the steps based on user requirements. The platform reads the page deterministically, names every step before anything runs, carries out each step the same way every time, and ends with a typed outcome and reason - barring the AI from spinning. Each task is a reproducible set of steps, so outcomes are predictable and failures can be reproduced and labeled.
Isolation for each tenant
A multi-tenant platform. Each tenant runs in its own isolated browser environment.
Lease based scheduling
A task holds a time limited lease on a browser environment while it runs.
Known good images
Environments start from a known good image. Failed sessions recover automatically from one.
Health monitoring
Runs are monitored while they happen.
Shared control handoff
A secure, expiring handoff link. An authorized person can take over a step, then signal done so the run resumes.
Machine readable failures
Every failure records what failed, why, and what to do next.
Your data stays yours. We never train on it or send it to an AI provider.
Your team can define every step before it runs, and every run ends with a clear result and a reason, so problems can be repeated and fixed instead of guessed at. We never train on your data or send it to an AI provider, and you bring the AI you already trust.
End to end tests on our own hardware.
We run end to end tests on the running service, on our own hardware.
- End to end tests, 15 of 15 passed (September 2026). Critical user journeys on running systems.
- Known good image builds. Each new browser environment starts from a known good, instrumented image.
- Isolation for each environment. Each environment is a separate, independent workspace.
- Handoff. The handoff to a person is tested as part of the end to end suite.
Measured results, published method.
Like for like: the same prewritten task run as Cinch steps and as raw CDP or Playwright code. No large language model in the loop. Runs alternated. Terms are defined below the table.
| Metric | Result | Evidence | Date |
|---|---|---|---|
| First-attempt success rate | 100% (10/10, both methods) | benchmark-suite-2026-09-22 BENCH-2 | 2026-09-22 |
| Warm-start latency | ~1.1s median | benchmark-suite-2026-09-22 BENCH-3 | 2026-09-22 |
| Cold-start penalty | +65% on extract; ~7.7s Chrome boot | benchmark-suite-2026-09-22 BENCH-3 | 2026-09-22 |
| Session awareness (authed workflow) | Cinch succeeds, raw CDP hits login wall | benchmark-suite-2026-09-22 BENCH-4 | 2026-09-22 |
| Concurrency speedup | 1.84x with 3 heads (18 runs, 100% success) | benchmark-suite-2026-09-22 BENCH-5 | 2026-09-22 |
| Token cost for verified tasks | $0.00 (0 model calls, strace-verified) | benchmark-suite-2026-09-22 BENCH-6 | 2026-09-22 |
- Cinch
- The Hayseed Live engine that carries out task steps deterministically.
- CDP
- Chrome DevTools Protocol. Low level commands sent straight to the browser. "Raw CDP" means code written directly against it.
- Playwright
- A widely used open source browser automation framework.
- LLM
- Large language model, the kind of AI model that drives each task.
- Heads
- Isolated browser environments. "3 heads" means three running side by side.
- Warm and cold start
- Warm means the browser is already running. Cold means it has to start first.
- Script green
- A task whose steps have already passed their checks.
- strace
- A Linux tool that traces system calls. We used it to confirm zero model calls.
What we do not claim yet.
What is in development, stated plainly.
- Integrity checking is research. Reconciling records against source screens, completeness checks across paged and filtered results, and evidence saved for each record are being built and measured.
- Escalation before a run is being measured. We are measuring how many blocked or failed outcomes the pre run check names in advance, and adding a typed reason for every miss.
- Recovery across several systems in one workflow, with task state kept outside the browser, is research.
- Credential vault is in development. Encrypted storage for credentials per client and per site.
- Dedicated servers for each tenant are in progress. Environments are isolated today and run on shared hosts.
- Context efficiency is a single pilot. On 2026-09-27, one 10 step workflow on one warm environment: a fixed Cinch run used 77 script tokens, 2.4 s, and $0.00. An agent loop used 600 model tokens and 8.6 s on the same task. One run. Indicative, not statistically significant.
How to check our working.
A person can step in on a live task when judgment or approval is needed.
Status link
Check any task's health and status through a secure, expiring web link.
Handoff to a person
When a step needs a person, such as an approval or a new sign in, the run pauses and offers a Done button. You take over, finish the step, and signal the run to resume.
Typed outcomes
Every task ends as success, partial, blocked, or failed, with a typed reason and a screenshot when it stopped.
Proof artifacts
Our proofs are kept as content addressed files that do not change after they are written. Ask us for them when you evaluate.
Try it on a real workflow.
Sign up for an account, or talk to us about an evaluation.