Describe it. Writ builds it.
Say the goal in one sentence. Scribe, our AI concierge, browses the page live, records the steps, proposes the selectors, and wires up a callable workflow.
Would rather just describe it? Meet Scribe.
The same describe-a-goal engine powers Scribe, our flagship AI concierge, a chat assistant you can open from anywhere in the app. Tell Scribe an outcome in one sentence and it drives every part of Writ for you: finding the page, creating the monitor, drafting the workflow or the buy, and wiring the whole automation, pausing only to confirm what matters.
On it. I opened the orders page and drafted the steps.
Deterministic replay vs. AI navigation, and when to use each
Writ gives you two execution modes, and they're explicit. Recorded recipes give you certainty; AI navigation gives you resilience. You can mix them in one workflow: record the stable parts, let AI handle the parts that move. AI is used where adaptivity helps; the rest runs as deterministic replay. No hand-waving.
| Criterion | AI navigation (adaptive) | Recorded replay (deterministic) |
|---|---|---|
| How a step is decided | A model reads the live page and chooses | Replays the exact recorded step |
| Model in the loop at run time | Yes, per step | No, none |
| Best for | Flows that shift or vary by account | Stable flows that need certainty |
| When a target moves | Finds it by meaning, adapts | Fails unless AI recovers it |
| Strength | Resilience | Predictable, fast, auditable |
| Use both | Record the stable parts; let AI handle the parts that move. | Record the stable parts; let AI handle the parts that move. |
AI navigation compared to deterministic recorded replay
Survives UI changes
Selector scripts break the moment a site ships a redesign (a renamed class, a moved button, a new modal) and you're back to maintenance. AI navigation reads the page the way a person would and finds the right element by meaning, not by a fragile CSS path. When a recorded step can't find its target, AI can step in to recover the flow instead of failing it.
css: .btn-buyNot foundbutton renamed to "Purchase"
ai: "the buy button"Adapted- Find elements by meaning, not brittle CSS paths.
- AI recovery when a recorded step's target moves.
- Less selector maintenance after every redesign.
Works on the hard web
The web that breaks other tools (login walls, two-factor prompts, content that loads as you scroll) is the web Writ is built for. On your own authenticated accounts, Writ logs in, reads a TOTP code from an authenticator or a one-time code from your mailbox, completes the verification step, and waits for dynamic content before acting.
Authorized access, not circumvention. Every one of these runs on your own authenticated accounts, on sites you are authorized to use. Writ automates work a person could do by hand in their own browser; it does not break into anything. See Personas, reusable logged-in identities with built-in 2FA.
Your keys run unbilled. Managed AI is metered, plainly.
Bring your own AI keys and every run is free of compute charges: the tokens go to your provider, not to us. Prefer not to manage keys? Use managed cloud AI and pay only for what a run actually uses.
AI on your terms
Bring your own keys: unbilled
- Point Writ at your own Anthropic, OpenAI, or Ollama key.
- AI sessions, assist, and optimize run with no Writ compute charge.
- Local and BYO runs stay on your machine; data never leaves.
Managed cloud AI: metered
- No keys to manage; Writ supplies the model in the cloud.
- Billed by tokens used; the price is shown before you run.
- Drawn from a prepaid balance, alongside cloud run time and managed AI.
AI auto-repair (the cloud rewriting a broken recipe) runs on managed cloud AI. Self-healing that re-locates a moved element runs everywhere, on your keys or ours.
Three ways AI builds the workflow
You don't have to script anything. Build a callable workflow three ways, then publish it as a REST endpoint and an MCP tool.
Want it fully autonomous? An AI session can drive the live browser to author a list, detail, and pagination web data extraction workflow for you.
AI assist and optimize
While you build, AI assist helps shape and fix steps; AI optimize reviews a recorded workflow (including the network calls behind it) and proposes a faster, more robust version. Optimize previews its changes so you can apply or discard; it never silently rewrites a working flow.
AI optimize previews its changes. You apply or discard; nothing silently breaks.
AI step types
Let AI act inline within an otherwise deterministic flow.
Deterministic step types
A slice of the 30+ step types that run exactly as recorded.
Observable runs you can audit
Autonomy is only trustworthy if you can see what happened. Every AI run is logged in your unified Runs log: the goal, each action it took, the page state, and the result. You get observable, auditable runs, not a black box.
The Runs log for an AI session shows each action the AI took. It reflects whether recent real runs are succeeding, not an independent "verified" badge.