Browse courses · authoring studio

Teach a route. Read it again.

Create a repeatable course for a portfolio, receipts, or a website. Define the result, pin the page and fields, then rehearse with saved evidence before exporting.

Import saved capture or evidenceGet the Chrome extension Estimate timing & cost

Start with one clear outcome.

A course describes the values to return and the identity that keeps each record distinct.

Course nameCourse IDRecord identity fieldsComma-separated keys. Receipts use mailbox + message ID; balances use the account.

A course contains declared reading routes. Every route produces a record with the same field definitions. Repeated identities are compared across runs.

Course economics

Know what each run takes.

Compare browser time, model usage, and evidence size. The starting workload is illustrative. Import a capture manifest to replace assumptions with observed stages.

Run performance suite ↗

First attempt, with startup

3.41 s

Scenario latency · not p95

Per scheduled course

$0.00144

Includes expected retry usage

Per successful course

$0.001444

Spend ÷ expected validated courses

30-day usage

$1.04

720 scheduled courses

USD usage before plan fees, credits and tax. Local mode excludes your device cost. Unpriced active items make the total unknown.

Choose the capabilities a course needs

Current paths: Chrome capture, exact text / selector readings, cropped OCR, and optional hosted vision. These additional agent-browser capabilities are candidates for the package; the planner does not enable them.

Scoped snapshots ↗

Adapter planned

Find controls and read accessible content within a section. Interactive-only snapshots can omit the static values a course needs.

Measure: Commands, response bytes, model tokens.

WebMCP ↗

Experimental · adapter planned

Use a site-provided tool when it covers the requested fields. Inspect its current schema and origin. Availability and read-only claims need verification.

Measure: Discovery, invocation, validation.

Files & downloads ↗

Site-export adapter planned

Use a site's export when it provides the required data. Keep repeat screenshots scoped; change detection can reduce returned images but still needs observation work.

Measure: Download, parse, transfer, retention.

Remote Agent Browser ↗

SDK migration under evaluation

Reuse a remote session for related courses. Expired sandboxes resume with fresh Chromium state. Our scheduled cams currently use @agent-browser/sandbox.

Measure: Startup, active CPU, memory lifetime.

eve integration ↗

Optional · not installed

Run browser tools inside an eve agent sandbox. This is an agent integration, separate from the Chrome extension used to capture your signed-in tabs.

Measure: Agent turns, model usage, sandbox usage.

Video recording ↗

Diagnostic adapter planned

Record difficult routes or failures. Frame rate, duration, and optional contact sheets add encoding and storage work; keep those costs explicit.

Measure: Encoding, recording bytes, retention.

RuntimeIndependent workersScheduled courses / dayForecast daysFailure probability (0–1)Maximum attempts

1.050 expected attempts · 99.75% expected completion · 1.41 s warm attempt. Each failed attempt replays the entire course. Independent workers reduce elapsed time; their memory and CPU usage are still added together.

Reading stops

overview

assumption

https://portfolio.example/accounts/main/overview

Planned reading method

Method is a planning label. Enter its measured or assumed timings below; changing it does not run an adapter or invent savings. Blank means unknown; 0 means a step is skipped.

Navigation (ms)Assumption / unknownPage readiness (ms)Assumption / unknownText / selectors (ms)Assumption / unknownScreenshot (ms)Assumption / unknownImage processing (ms)Assumption / unknownExtraction (ms)Assumption / unknownValidation (ms)Assumption / unknownWrite evidence (ms)Assumption / unknownThumbnails / recording (ms)Assumption / unknown

Active CPU (core-ms)Billable transfer (bytes)Retained evidence (bytes)Model input tokensModel output tokens

Session lifetime & retention

Startup is per worker session. Runs sharing a session amortize startup and creation costs. Count all waiting between those runs in idle seconds. Workers wait at course boundaries.

Startup (ms)Startup CPU (core-ms)Setup transfer (bytes)Idle seconds / sessionCourses / sessionMemory GB / workerMemory billing step (s)Evidence retention days

Rates & billing units

Vercel Sandbox · iad1 · Pro usage · checked 2026-09-17 · https://vercel.com/docs/sandbox/pricing. Provider pricing ↗

CPU $ / core-hourMemory $ / GB-hour$ / sandbox creationTransfer $ / GBEvidence $ / GB-monthInput $ / million tokensOutput $ / million tokens

Cost of one scheduled course

Sandbox active CPU: $0.000025
Sandbox memory, including idle: $0.001413
Sandbox creations: $0.000001
Billable sandbox transfer: $0.000001
Model input tokens: $0.00
Model output tokens: $0.00
Evidence storage at steady state: $0.00

Import capture manifestRestore cost plan

Calculation assumptions

Independent routes assigned in order to the next free worker; each sandbox worker closes when its work and configured idle period end.
Retries replay the whole course, with identical failure probability and cost per attempt. No queueing, backoff, or contention is modeled.
Session reuse groups consecutive runs; include gaps in idle seconds. Amortization does not guarantee retained authentication or page state.
Latency is a scenario estimate, not p50/p95. Imported stages are single capture observations; missing stages remain unknown.
CPU input is active core-ms from a meter or an assumption, never wall time multiplied by allocated vCPUs.
Costs are marginal USD usage before credits, plan fees and tax. Device costs, orchestration, proxies, requests and persistent sandbox images are excluded.
Transfer input must follow provider billing rules; screenshot file size and browser response bytes do not establish billable transfer.
Evidence storage is a steady-state rolling retention estimate, with decimal GB. Failed attempts are assumed to retain the same bytes.

All calculations run locally. Exported plans include course URLs. Editing the course resets this planner so measurements cannot silently carry across revisions.

The studio runs in your browser. It makes no site requests and uses no account credentials. A course can later run through agent-browser in a dedicated session; visual discovery and automatic route learning are separate steps.