Record or typeShow it the task, once.
Record a browser task with narration — screen plus mic — or skip recording and
just type it in plain English at /teach. Either path produces the same starting draft.
Local · Chrome · Your keys
Hand an AI a browser task and it improvises — every run is a fresh gamble. Realtime Teach watches you do the task once, distills it into a skill any agent can follow, then certifies that skill against your own Chrome until it replays the same way every time — with zero LLM calls in the loop.
Install — macOS, Apple Silicon
curl -fsSL https://realtimeteach.com/install | sh
then realtime-teach up
→ open registrar → search "{query}" → add to cart → checkout
// durable intent + landmarks, never pixel coordinates
trial 01 trace compiled → recipe validated
The problem
Great for a one-off. Miserable for a workflow you run every day — because the agent re-reasons the whole task from scratch on every single run.
Re-explain the task, watch it click the wrong button, retry, and pay for a full model run whether it works or not. Yesterday's success guarantees nothing today.
One certified skill, replayed from a compiled recipe — deterministic, fast, and with no model in the loop. Teach it once; it holds.
The artifact
Realtime Teach produces a plain-English skill.md: durable intent and
on-screen landmarks — never pixel coordinates or click ordinals. That's what lets it
survive a redesign and run on a fresh agent that never saw your recording.
{variables} before anything touches disk.--- name: buy-domain inputs: [query] --- # Buy a domain # Landmarks, not coordinates. 1. Open the registrar's homepage. 2. Type {query} into the domain search. 3. On the results, add the .com to the cart. 4. Go to checkout and confirm the order. oracle: "Order confirmation shows the {query}.com domain."
How it works
Four stages. Only the first one needs you.
Record a browser task with narration — screen plus mic — or skip recording and
just type it in plain English at /teach. Either path produces the same starting draft.
A distill pass renders your demo into a structured draft and then a readable
skill.md you can edit. Sensitive data is masked before storage — nothing
secret is ever written down.
A local daemon runs fresh agents against your Chrome over CDP — no recording context,
only the skill.md. Replay-first: the first unaided pass whose trace compiles
into a validated recipe certifies on the spot; otherwise it falls back to five straight vision
passes. A failed trial is classified environment / agent / skill — a patcher rewrites the
prose and you re-run. Login walls stop the run; they're never patched away.
Run the certified skill and it replays the compiled recipe with zero LLM calls — fast and deterministic. If the page has drifted, one vision-agent trial finishes the job and recompiles the recipe for next time.
The cockpit
Open a run in your browser and follow every trial as it happens: per-step screenshots, the trial timeline, and the reflection and patch after each attempt.
Pause to inject a hint — it feeds the agent and the patcher at the same time, so you only ever teach it once.
Local-first
This is a local app, not a hosted service. Your browser data has no reason to travel, so it doesn't.
Recordings, screenshots, cookies, and certified skills stay on your laptop. No upload path exists.
It binds 127.0.0.1 and rejects foreign hosts and origins — a DNS-rebinding guard, so no random web page can drive it.
Keys live in ~/.realtime-teach/config.json at chmod 600. Env vars override, and the app never echoes them back.
Nothing to sign up for and no server to pay for. Bring your own Anthropic + Gemini keys and go.
Get started
Google Chrome · macOS (Apple Silicon) · Anthropic + Gemini keys
curl -fsSL https://realtimeteach.com/install | sh
then realtime-teach up
The first run walks you through a quick setup wizard — Chrome check, keys, profile — and ships an example skill you can certify immediately.
Release updates and early access. No spam, ever.