Straight answers about running an agent.

Twin Browser is the browser layer for AI agents: hand it a plain-English goal and it reaches any site, pulls any data, and signs into the accounts you connect — doing the real work in a live browser while you set the guardrails. Repeated tasks then compile into skills that replay deterministically at near-zero cost.

Most browser infrastructure re-runs the LLM on every execution, so cost climbs with usage. Twin moves the LLM to the planning stage: the first run compiles a reusable skill, and subsequent similar runs hit a semantic cache and replay deterministically. A cache hit is roughly 5x cheaper, so the marginal cost per run trends toward zero the more your agents run.

An exact-match cache only helps when the identical action repeats — it is keyed on a selector or a parameter hash, so a re-worded request is a cold start. Twin embeds the request and matches by meaning, so "schedule a call" finds the skill compiled for "book a demo." That is why Twin keeps saving where exact-match caches fall back to a full LLM run.

When a goal succeeds, Twin freezes the action path into a skill keyed to the page structure. On a cache hit it replays that skill deterministically — the same indexed clicks and inputs — with no model call. It is production-ready, not a debug-only recording.

The cross-tenant skill corpus reuses the structural action plan of a skill, not your data, credentials, or run inputs. Secrets stay in your per-tenant credential vault, the backend enforces default-deny row-level security, and every call is authenticated and audit-logged. The shared layer is the "how to operate this page" knowledge, which raises the cache-hit rate for everyone.

Three ways: the REST API under /api/v1/* with Bearer-key auth; the MCP server, which exposes the tools run, compile_skill, and run_skill to Cursor, Claude Desktop, Claude Code, and Cline; and native LangChain and AutoGen tool adapters. See the integrations pages for copy-paste snippets.

The Twin MCP server exposes three tools: run (execute a goal on a URL), compile_skill (freeze a successful run into a reusable skill), and run_skill (replay a compiled skill). Any MCP client — Cursor, Claude Desktop, Claude Code, Cline — can drive the full pipeline without writing API glue.

Every request carries a per-tenant API key as a Bearer token in the Authorization header. Keys are scoped to a tenant, and authentication, usage billing, and audit logging run on every call before any browser action is taken.

Yes — in a per-tenant credential vault. A flow that needs to log in pulls secrets from the vault at run time rather than embedding them in the goal or your code. Vault contents are isolated per tenant under default-deny RLS.

A sign-in run never just fails on a 2FA wall — it PARKS by default: it returns status:"paused" with a sessionId and holds the browser session open (resume via POST /api/v1/runs/:id/resume, the MCP submit_verification tool, or the dashboard). If you connect an email inbox or a phone number, Twin auto-fills an emailed or texted code itself and resolves the pause with no human in the loop; an authenticator-app code, a phone approval, or a code only you hold is supplied by a person, then the run resumes from the exact step. Pass hitl:false to opt out and get a plain miss instead.

Every run streams a real-time live view and is recorded to durable session video, so you can watch a flow live or review exactly what happened afterward. Proxy support (IPRoyal) is available for runs that need it.

Pricing is usage-based credits, free to start, with an entry plan from $29/mo. Each call is billed the higher of its flat action price or the metered cost of that run, against a transparent rate card published at /api/v1/pricing. Pay-as-you-go auto top-up keeps long-running agents from stalling.

Because Twin is built so usage gets cheaper. As your cache-hit rate rises, more runs replay deterministically at near-zero LLM cost, so your blended cost-per-1k-runs falls. A flat per-seat price would hide that; usage credits let the savings flow straight to you.

Twin automates the web where you are authorized — first-party sites, operator-approved automation, internal RPA, accessibility, and authorized testing. The run target URL is the authorization signal, and auth, billing, and audit logging run on every call. Twin is not a CAPTCHA-bypass-for-hire or anti-bot evasion service.

A multi-tenant Supabase backend with default-deny row-level security, per-tenant API keys, a credential vault, and an audit log on every call. Tenants are isolated by default; access is granted explicitly, never assumed.

They are all good at running a browser, but their default loop re-runs the LLM (or caches only exact, single-workflow repeats), so cost stays flat or rises with usage. Twin adds a semantic dispatch cache that matches re-phrased requests and a cross-tenant skill corpus that compounds savings. See the side-by-side breakdowns on the compare pages.