What does “no LLM in the loop” actually mean for cost?

It means the run has no token bill at all. Twin Browser bills higher-of(flat floor, metered model + compute + egress); with no model call the metered half is zero, so a replay settles at its flat floor every time. That is why the number is stable in a way a planner-driven run’s cannot be.

Can I pass values into a replay?

Yes, over REST: a compiled skill keeps {{param:NAME}} tokens rather than the literal values used at compile time, and POST /api/v1/skills/{name}/run accepts a `params` object that binds them. list_catalog tells you which params a given skill expects. The MCP tool does not currently forward params.

Core tool

Replay a compiled skill with no model in the loop

Blind deterministic replay of a stored path: the same actions, in the same order, against the same host — no planner, no tokens, one credit.

Get an API key MCP setup

run_skill

What it is

The job: Run a known-good flow again, as cheaply as it can possibly be run.

Replay is the point of compiling. run_skill loads the stored path and executes it directly — no observation-plan-act loop, no model call, no token bill. What you pay for is browser time.

Credentials and sessions are resolved server-side: the skill records which secrets it fills, and the vault supplies them inside the browser at fill time. Nothing about the credential passes through your prompt or the response.

Replay is deliberately blind, which is its strength and its limit. It does not reason about the page, so it is fast and predictable; it also cannot improvise if the page has moved. Drift is detected and repaired at the platform level rather than replayed blindly into a wrong click.

The call

Call it exactly like this.

Copied from the tool's registration and the route handler — not paraphrased.

MCPrun_skill.mcp.jsonjson

// MCP tool call — server "twin-browser"
{
  "tool": "run_skill",
  "arguments": {
    "name": "acme-login",
    "target": "https://app.example.com"
  }
}

POST /api/v1/skills/{name}/runrequest.shbash

curl -X POST https://twin-browser.com/api/v1/skills/acme-login/run \
  -H "Authorization: Bearer $TWIN_API_KEY" \
  -H "content-type: application/json" \
  -d '{"target":"https://app.example.com","params":{"query":"invoices"},"account":"work"}'

The MCP tool takes only { name, target }. The REST endpoint additionally accepts `params` (the values a parameterized skill binds) and `account` (which stored login to use). If your skill is parameterized, call it over REST — or use list_catalog first to see exactly which params it wants.

Parameters accepted by run_skill
Parameter	Type	What it does
name*	string	Name of a compiled skill in your tenant. list_skills and list_catalog both return valid names.
target*	string	Base URL to replay against. Normally the same host the skill was compiled on.

Returns

response.jsonjson

{
  "success": true,
  "path": [ /* the replayed actions */ ],
  "steps": 6,
  "replayed": 6,
  "runId": "…",
  "credits_charged": 1
}

What it costs

1 credit

Flat. There is no model in a replay, so there is no metered LLM cost to settle against — the 1 credit is the whole charge.

See the full rate card

Which one

When a different tool is the right call.

The honest answer is often the neighbouring tool. These are the trades.

dispatch

dispatch finds the right skill for you from a natural-language goal; run_skill needs you to already know its name. If your caller is an LLM improvising phrasing, dispatch. If it is your own code with a fixed workflow, run_skill is one less matching step and one less thing to go wrong.

dispatch

run_goal

run_goal re-reasons the flow every time — ten times the cost, and the right answer when the page genuinely differs run to run.

run_goal

Questions

run_skill, answered.

What does “no LLM in the loop” actually mean for cost?: It means the run has no token bill at all. Twin Browser bills higher-of(flat floor, metered model + compute + egress); with no model call the metered half is zero, so a replay settles at its flat floor every time. That is why the number is stable in a way a planner-driven run’s cannot be.
Can I pass values into a replay?: Yes, over REST: a compiled skill keeps {{param:NAME}} tokens rather than the literal values used at compile time, and POST /api/v1/skills/{name}/run accepts a `params` object that binds them. list_catalog tells you which params a given skill expects. The MCP tool does not currently forward params.

Keep going

The rest of the tool set.

compile_skill

Turn a flow you are going to run many times into a stored artifact that costs almost nothing to run again.

dispatch

Be the one call an agent makes, and get cheaper every time the agent asks for the same thing in different words.

list_catalog

Let an agent plan: pick a skill it can actually run, with values it can actually supply.

list_skills

Answer “what do I already have?” before spending anything.

The mechanism behind it:Deterministic replay Skill library Credential vault

Get an API key All tools