Glossary
The language of agentic browser automation
Definition-first entries for the concepts behind Twin Browser and the wider field — from the guardrails that keep you in control of an authorized run to the vault that holds the accounts you connect.
Terms
61 definitions, grouped by what they explain
Each entry is a short, citable definition plus why it matters when you run agents over the browser at volume.
Cost engine
Semantic dispatch cache
A vector cache that matches a new, re-phrased request to a skill you already compiled — so a repeated task skips the LLM.
Deterministic replay
Re-running a compiled skill as a fixed sequence of indexed actions, with zero LLM calls, so the same goal produces the same path every time.
Skill compilation
Turning a successful agent run into a reusable, replayable skill — a compact program of indexed actions you can call again without re-planning.
Cross-tenant skill corpus
A shared, sanitized pool of compiled action structure that lets a flow solved once seed a first run for someone else.
Skill sanitization
Stripping inputs, secret values and identifiers out of a compiled path so only the shape of the actions can be shared.
Late binding
Deferring the variable parts of a compiled skill to call time, so one program serves every customer, date or record.
Skill brick
A named sub-flow inside a compiled skill — the segment of the path that accomplishes one phase, reusable on its own.
Cache hit rate
The share of runs that resolved to a compiled skill instead of a fresh compile — the single number that says whether the cost engine is working.
Cold start vs warm start
A cold start plans the flow from scratch and pays the model; a warm start dispatches to something already compiled and does not.
Marginal cost per run
The cost of one more execution of a workflow. On most browser infra it’s flat; here it falls as cache hits replace model calls.
Credit
The unit every action is billed in, settled as the higher of a flat per-action floor and the metered cost the action actually incurred.
Page model
DOM indexed state
A compact, numerically-indexed map of a page’s interactive elements that an agent acts on by index instead of raw HTML.
Token-efficient DOM
A page representation stripped to just the interactive, indexed elements under a token budget — so prompts stay small and cheap.
Token budget
A hard ceiling on how much page state may reach the model, so a huge page costs about the same to plan against as a small one.
CSS selector drift
The slow failure mode of selector-based automation: a class rename or a moved container silently breaks a script that used to work.
Structured extraction
Reading a page into typed JSON you specified — by schema, by field list, by template, or by a per-row selector map.
Extraction template
A per-host or generic-metadata recipe that reads a known page shape into structured data with no model call at all.
Row expansion
Pointing at one row of a repeating listing and having every structurally-alike row extracted from it.
XHR capture
Collecting the page’s own background API responses instead of parsing the markup it built from them.
Data & retrieval
Crawler
A bounded, breadth-first reader that follows links from a start URL under path, depth and page limits.
Site map (discovery)
A flat-fee inventory of a site’s URLs from sitemap.xml, robots.txt and a shallow link scan — URLs only, no page content.
robots.txt
The file at a site’s root declaring what automated clients may fetch, how fast, and from where.
Autothrottle
A crawler pacing itself from the site’s own responses — slowing when latency rises or a 429 arrives, relaxing as it recovers.
Deep search
A search that does not stop at the ranking — it scrapes the top results for content, as a background job.
Embedding
A numeric vector representing the meaning of a piece of text, so two differently-worded things can be compared by distance.
Chunking
Splitting a document into overlapping segments small enough to embed and precise enough to retrieve.
ETL (extract, transform, load)
The pipeline that takes a source, cleans and shapes it, chunks and embeds it, and persists it into a queryable store.
Semantic search
Retrieving stored content by meaning rather than keyword — the query is embedded and matched against embedded chunks.
Change monitoring
Reading one value on a schedule, hashing it, and firing a webhook only when the hash actually moves.
Execution
Human-in-the-loop handoff
Pausing a run at a step that needs a person — an approval or a verification code — then resuming from the same step once it is cleared.
Run parking
The paused state a blocked run enters instead of failing — browser still open, session id returned, credits refunded.
Verification wall
A second-factor step — an SMS or emailed code, an authenticator code, or a phone approval — standing between a sign-in and the account.
Co-drive
Clicking and typing on a run’s live browser yourself, with coordinates expressed as fractions of the viewport.
Session video
A durably-stored recording of a browser run — opt-in per run, tiered by plan, and ending on the run’s actual end state.
Server-sent events (SSE)
A one-way HTTP stream of named events — how a live run, a parked session and a job’s status are all delivered.
Async job
A run submitted for background execution that returns an id immediately, to be polled, streamed or delivered by webhook.
Success condition
The structured test that decides whether a free-text goal actually succeeded — not the model’s own opinion of its work.
Headless vs headful
Headless runs a browser with no visible UI; headful runs a real, rendered browser. The choice affects compatibility and how some sites behave.
Identity & access
Credential vault
Write-only per-tenant storage for the secrets a run needs — referenced by name, resolved inside the browser, never returned by any read.
Credential injection
Substituting a secret’s value inside the browser at the moment a field is filled — after the planner has already decided what to type there.
Connect handoff
A single-use link where the account’s real owner signs in by hand, so only the resulting session is captured — never the password.
Session persistence
Storing and restoring a browser’s authenticated cookies per account, so the next run starts logged in instead of signing in again.
Multi-tenant isolation
The guarantee that one tenant’s runs, skills, sessions and secrets are unreachable from another — enforced in the database, not the application.
Row-level security (RLS)
A database policy that decides per row whether the current identity may read or write it — the enforcement point for tenant isolation.
API key scope
A named permission on a key — run, skills:read, skills:write, tenants:provision — that decides which endpoints it may call.
Sub-tenant
A tenant provisioned under your account, with its own key and library, billing either to its own balance or to your pool.
Network & evasion
Proxy egress
The network a browser session leaves from — a managed residential or mobile pool, or a proxy URL you supply yourself.
Sticky egress
Pinning one account’s traffic to a stable exit IP across runs, so an identity does not appear to move continents between sessions.
Geo pinning
Setting one country that drives both the exit IP and the browser’s locale and timezone, so the two halves cannot disagree.
Browser fingerprinting
Identifying a client from the combination of signals its browser exposes — and noticing when those signals do not fit together.
Bot detection
The vendor systems sitting in front of a site that decide whether a visitor is automated — and what to do about it.
Challenge wall
An interactive block — a CAPTCHA widget or an interstitial — that can sometimes be cleared inside the session.
Score wall
An invisible reputation check that no client-side change reliably passes — the case where the honest answer is a human sign-in.
Cadence governing
Pacing runs per account so one identity is not re-poked faster than a real user would — and backing off after a push-back.
Rate limiting
A cap on request volume — the service-level kind that protects the API, distinct from the account-level pacing that protects your login.
Integration
MCP server
A Model Context Protocol server exposing the browser layer as 35 tools to clients like Cursor, Claude Desktop, Claude Code and Cline.
Webhook
An HTTP POST the platform makes to your endpoint when something finishes or changes — the push alternative to polling.
Webhook signing
An HMAC over the raw delivery body, sent as a header, so your endpoint can prove the POST really came from the platform.
Concepts
Browser agent
An LLM-driven system that perceives a live web page and acts on it toward a goal, rather than following a recorded script.
Browser execution layer
The infrastructure between an LLM agent and a real browser that turns a goal into deterministic, replayable action.
RPA vs agentic automation
RPA follows brittle, hand-recorded scripts; agentic automation has a model plan against the live page — then compiles the result so it replays cheaply.
Put your agent to work. Keep the decision.
Hand your agent any site, authorize the accounts it may use, and set the guardrails. Repeated workflows compile into skills that replay without a model.