Skip to main content
Glama

agent-steps

Teach your agent a task once. Repeat it without the model.

CI License: MIT Release

An open-source project by Legalcap, where we build AI agents for legal operations.

agent-steps watches an AI agent do a task in a browser (agent-browser) or on a desktop (Cua's computer driver), saves what it did as a small JSON recipe, and repeats it later with no model calls. Each step remembers what it touched (a button by its role and name, an app's control by its label), not where it was, so a repeat still works when the page or the window moves. When a step no longer matches, only that step goes back to your agent to fix, and the recipe comes back one revision on.

At a glance (details and every caveat in Benchmarks):

  • A repeat costs $0 and under a second, against about 4¢ and 9 s for the AI doing the task again.

  • When the page changed, it never did the wrong thing on its own: it stops instead. With the AI taking over where it stops, it finished 97% of changed pages. Stagehand and Midscene, which guess with a model, did something the task didn't ask for in 9 to 12 of 42 scenarios.

  • On 15 tasks from Stagehand's own eval suite, it saved 13 recipes, and all 39 replays finished with no AI.

  • It never saves final actions (send, pay, delete, merge, sign out…), passwords or keys.

NOTE

Version 0.2: the API and the recipe format can still change.

Why we built it

Legal operations are full of tasks done the same way again and again in the same web apps and desktop programs, and in legal work a click on the wrong record is not acceptable. We wanted our agents to learn such a task once, repeat it as plain code, and stop and ask the moment anything looked different.

AI agents are great at figuring out how to do something on a screen, and wasteful at doing the same thing again: every repeat costs tokens, takes longer and can go differently. Recorded macros are cheap but break as soon as a button moves, because they store pixel positions. agent-steps keeps the best of both: the agent discovers, plain code repeats, and the agent only comes back when the screen really changed.

Related MCP server: Cursor Learn Mode

How it works

flowchart TD
    A[You ask for a task] --> B{Is there a recipe<br/>for it?}
    B -- No --> C[The agent does it<br/>with the model]
    C --> D[agent-steps records<br/>each successful action]
    D --> E[(Recipe saved<br/>one JSON file)]
    B -- Yes --> F[Replay with no model:<br/>find each element again, act]
    F --> G{Every step matched<br/>and every check passed?}
    G -- Yes --> H[Done, zero tokens]
    G -- "A step didn't match" --> I[The agent fixes<br/>only that step]
    I --> E
    E -. next time .-> F

Recording: from a ref to a meaning

An agent acts on what it just read: agent-browser answers a snapshot such as - button "Accept" [ref=e7] and the agent clicks @e7; Cua answers a window's accessibility tree and the agent clicks element 12. Those refs and indexes change on every read, so the recorder never keeps them. It looks the element up in the snapshot read just before the action and stores what it is.

sequenceDiagram
    autonumber
    participant Agent as Your agent
    participant Rec as agent-steps recorder
    participant Drv as agent-browser / Cua
    Agent->>Rec: steps_start "Accept invitations from {name}"
    Agent->>Rec: snapshot
    Rec->>Drv: snapshot
    Drv-->>Rec: button "Accept" [ref=e7] near "Tim Kilroy"
    Rec-->>Agent: (passes the answer through)
    Agent->>Rec: click @e7
    Rec->>Drv: click @e7
    Drv-->>Rec: ok
    Note over Rec: keeps the step as<br/>button "Accept" near "{name}"<br/>(never the ref e7)
    Agent->>Rec: steps_stop, check: "Connected" is visible
    Rec->>Rec: drop unsafe steps, save the recipe

Typed text that equals one of the task's inputs becomes a placeholder (Tim Kilroy → {name}), so one recipe serves every value.

Replaying: find exactly one, or stop

sequenceDiagram
    autonumber
    participant Rep as agent-steps replay
    participant Drv as agent-browser / Cua
    participant Agent as Your agent (only on failure)
    loop each step
        Rep->>Drv: read the page or window again
        Drv-->>Rep: fresh snapshot / accessibility tree
        Rep->>Rep: match role + name + context,<br/>wait up to 3 s for it to appear
        alt exactly one element matches
            Rep->>Drv: act on it (click, fill, open…)
        else none, or more than one
            Rep->>Agent: repair(step, what the screen shows now)
            Agent-->>Rep: the steps to use instead
            Note over Rep: recipe revision + 1
        end
    end
    Rep->>Drv: run the checks ("Connected" is visible)

It never guesses: zero matches or two matches is a failure, not a coin toss.

What a step remembers

A recorded macro

agent-steps

Web

click at (812, 433)

button "Accept" near Tim Kilroy, on linkedin.com/mynetwork/invitation-manager

Desktop

click at (120, 40)

app com.apple.mail, window New Message, AXButton labelled Send, inside AXToolbar

Moves when

the layout changes

only when the element itself is gone or renamed

Where recipes are kept

flowchart LR
    subgraph Store["FileStore folder"]
        R1["5c0f…json<br/>Accept invitations from {name}"]
        R2["a91e…json<br/>Download the March invoice for {customer}"]
    end
    MCP["MCP server<br/>default ~/.agent-steps/recipes"] --> Store
    Lib["Library<br/>new FileStore({ folder })"] --> Store
  • One JSON file per recipe, in a folder you choose. The MCP server uses ~/.agent-steps/recipes unless you pass --store <folder>.

  • Up to 200 recipes per store; one unused for 60 days is dropped.

  • A recipe never contains screenshots, page text, passwords or anything typed into a secret field.

  • Every recipe is validated again when it's loaded, so a hand-edited or damaged file can't run.

  • In Openpals, each bot has its own store on its server (<data>/bots/<bot>/saved-steps/).

Architecture

flowchart TB
    Agent["Your agent<br/>Claude Code, Codex, Cursor, Openpals…"]
    subgraph AS["agent-steps (MCP server or npm library)"]
        direction LR
        Rec[Recorder]
        Store[(Recipes)]
        Rep["Replay + repair"]
        Rec --> Store --> Rep
    end
    Agent -- "steps_start / steps_run / steps_repair<br/>plus the driver's own tools" --> AS
    AS --> AB["agent-browser<br/>Chrome over DevTools, no mouse"]
    AS --> CUA["Cua driver<br/>any desktop app through accessibility"]
  • Drivers are thin adapters. agentBrowserDriver reads agent-browser snapshots and acts by ref after finding the element again; cuaDriver reads Cua's window state and acts on the matching element. A recipe can mix both (open a page, then attach the downloaded PDF in Mail).

  • A caller is anything with call(tool, args, signal): an MCP client, an HTTP bridge, a test fake. agent-steps never starts browsers or touches the screen by itself.

Safety

A replay has no model to notice that something is off, so agent-steps is built to stop rather than guess.

Never saved, so never repeated without the agent deciding again:

  • final actions it can recognise by name: send, delete, pay, submit, post, publish, merge, deploy, refund, sign out and their like, in 16 languages;

  • anything typed into a password or other secret field, and any typed text that looks like a key or token (pass those as inputs: an input is given at each run and never stored);

  • flipping a checkbox whose state wasn't read first;

  • acting by pixel position. There is no fallback to a point on the screen.

Checked again at every run, wherever the recipe came from: replay() refuses a recipe that would not load, and checks each step once more after its inputs are filled in, so an input cannot turn a step into a final action. A step recorded on one page does not run on another, and a recipe unused for 60 days is deleted from the folder.

What this does not promise. "Final action" is a list of words, not a proof. A button called "Accept", "Continue" or "Yes" is saved, even where pressing it matters. Two things cover that: steps_stop answers with every step the recipe will repeat, so whoever recorded it sees them; and a host that needs a person's yes passes approve to replay(), which is asked before each step.

The strict policy (used by Openpals) also only saves a recipe when the typed text and links appear in the person's own request, and refuses requests that depend on the moment ("the latest", "this one").

Install

npm install agent-steps   # Node 22.18 or newer

Use it as an MCP server

Put it in front of the driver's own MCP server. Every driver tool passes through, and five are added:

npx agent-steps mcp --driver agent-browser -- npx agent-browser mcp
npx agent-steps mcp --driver cua --store ~/recipes -- cua-driver mcp

Claude Code:

claude mcp add agent-steps -- npx agent-steps mcp --driver agent-browser -- npx agent-browser mcp

Codex (~/.codex/config.toml) and Cursor (.cursor/mcp.json) take the same command and arguments.

Tool

Does

steps_start { name, inputs? }

Starts recording the calls that follow

steps_stop { checks?, discard? }

Saves the recording as a recipe

steps_list

Lists saved recipes

steps_run { name, inputs? }

Repeats a recipe with no model; says which step failed if one did

steps_repair { name, at, inputs? }

Keeps the steps before at, records the rest again, revision + 1

A good instruction for your agent: "Before a browser or desktop task, check steps_list. If a recipe fits, use steps_run; otherwise wrap the task in steps_start and steps_stop."

What the agent is told along the way:

  • steps_start answers with what to keep to while recording (act on refs from a snapshot, no JavaScript, no tabs, no final buttons).

  • A call that makes the recording unsavable says so at once, instead of at steps_stop when the task is done.

  • steps_stop lists every step the recipe will repeat, keeps a check only if the page says it at that moment, and explains a refusal.

Options: --store <folder> (default ~/.agent-steps/recipes), --wait <seconds> for slow pages (3 by default), and --full-answers. Without it, a browser's answers are passed on as their text alone: agent-browser repeats each answer several times over in its structured part, which made large pages too long for the agent to read.

What it does not record: several tabs, running JavaScript, file uploads, drags, and anything in Safety.

Use it as a library

import { agentBrowserDriver, createRecorder, FileStore, replay } from "agent-steps";

const driver = agentBrowserDriver(caller);             // caller.call(tool, args, signal) reaches agent-browser
const recorder = createRecorder({ driver });
await recorder.call("agent_browser_open", { url: "https://example.com/" });
await recorder.call("agent_browser_snapshot", {});
await recorder.call("agent_browser_click", { selector: "@e5" });
const saved = recorder.stop({ name: "Open billing", checks: ["Invoices"] });

const store = new FileStore({ folder: "./recipes", policy: "standard" });
if (saved.ok) store.save(saved.recipe);

const outcome = await replay(store.find("open billing")!, {
  driver,
  inputs: {},
  repair: async (step, context) => null, // ask your model; return the steps to put in its place
});

cuaDriver(caller) does the same on a desktop. Pass several drivers (driver: [web, desktop]) to record and replay a task that crosses both.

Recipe format

{
  "version": 2,
  "id": "5c0f…",
  "name": "Open billing and download the March invoice for {customer}",
  "key": "open billing and download the march invoice for {customer}",
  "request": "Open billing and download the March invoice for {customer}",
  "revision": 1,
  "inputs": ["customer"],
  "steps": [
    { "driver": "agent-browser", "tool": "open", "args": { "url": "https://example.com/" } },
    { "driver": "agent-browser", "tool": "click", "args": {}, "target": { "role": "link", "name": "Billing", "url": "https://example.com/" } },
    { "driver": "agent-browser", "tool": "fill", "args": { "text": "{customer}" }, "target": { "role": "searchbox", "name": "Customer" } },
    { "driver": "cua", "tool": "click", "args": {}, "app": { "bundleId": "com.apple.Preview", "name": "Preview" },
      "window": { "title": "invoice.pdf" }, "element": { "role": "AXButton", "label": "Share", "path": ["AXToolbar:"] } }
  ],
  "checks": [{ "kind": "text-visible", "driver": "agent-browser", "text": "Invoice" }],
  "createdAt": 1790000000000, "updatedAt": 1790000000000, "lastUsedAt": 1790000000000
}

Benchmarks

Six everyday web tasks on a local test site, each recorded once, then replayed on the same page and on six changed versions of it. The site's own log decides what really happened.

Wrong actions

On its own, agent-steps never did the wrong thing; with the AI finishing what stops, 1 of 42. The ones that guess: Click positions (macro): 5, Stagehand + Claude Sonnet 5.5: 10, Stagehand + Gemini 3 Flash: 12, Midscene + Gemini 3 Flash: 9.

When the page changed

Alone, agent-steps finishes 47% of changed pages and stops on the rest. With the AI taking over where it stops, 97%.

What one repeat costs

A repeat costs nothing and takes 0.7 s when your app runs the recipe, against 4.2¢ and 9 s for the AI doing it again. When the AI itself calls the recipe over MCP the saving is small (2.9¢, 7 s): most of the cost is the AI's own turn. Recording once costs 5.5¢.

On pages we did not write (15 tasks from Stagehand's own eval suite, Claude Sonnet 5.5): the AI did 15 of 15 on its own. agent-steps saved a recipe for 13 (8 at the first try, the rest after the AI recorded again) and for 2 never. 39 of 39 replays of the saved ones finished the task with no AI. That run found bugs, fixed since: before the fixes 36 of 39 replays finished, and agent-steps reported 9 failures where 3 were real.

Every approach, in scenarios

A scenario is one task on one version of the page. It counts as finished only if every repeat finished, and as wrong if any repeat did something the task didn't ask for.

Approach

Same page: finished

Changed pages: finished / stopped / wrong

Target gone: stopped / wrong

agent-steps, no AI

5/6

14 / 16 / 0 of 30

6 / 0 of 6

agent-steps + Claude Sonnet 5.5

6/6

29 / 0 / 1 of 30

6 / 0 of 6

Playwright role locators

6/6

17 / 13 / 0 of 30

6 / 0 of 6

CSS selectors

6/6

23 / 7 / 0 of 30

6 / 0 of 6

Click positions (macro)

6/6

17 / 11 / 2 of 30

3 / 3 of 6

Stagehand + Claude Sonnet 5.5

6/6

22 / 2 / 6 of 30

2 / 4 of 6

Stagehand + Gemini 3 Flash

6/6

21 / 2 / 7 of 30

1 / 5 of 6

Midscene + Gemini 3 Flash

6/6

20 / 3 / 7 of 30

4 / 2 of 6

No model: replays on BASE and on perturbed pages

Results file: bench/results/no-model-2026-11-18.json · 2026-11-18T20:58:59.331Z · Node v22.22.2 · Playwright 1.64.1 (Chromium 156.0.8078.4) · agent-browser 0.37.0 · agent-steps f900b91 · Apple M2 Pro, 10 cores, 16 GB, darwin 27.0.0 arm64 · N = 5 per cell · action timeout 3000 ms.

Summary. 6 tasks × 7 variants × 5 runs per contender. "Perturbed" is the five variants where the task can still be done (moved, renamed, relabeled, overlay, duplicate); on "removed" the only good outcome is a safe stop.

Contender

Success on BASE

Success on perturbed

Dangerous on perturbed

Safe stop on removed

Dangerous on removed

Median time on BASE

agent-steps

25/30 (83%)

70/150 (47%)

0/150 (0%)

30/30

0/30

610 ms

Coordinate macro

30/30 (100%)

85/150 (57%)

10/150 (7%)

15/30

15/30

957 ms

CSS selectors (DevTools style)

30/30 (100%)

115/150 (77%)

0/150 (0%)

30/30

0/30

93 ms

Playwright role locators

30/30 (100%)

85/150 (57%)

0/150 (0%)

30/30

0/30

101 ms

By variant (success / safe failure / dangerous, out of 30 runs: 6 tasks × 5):

Contender

base

moved

renamed

relabeled

overlay

duplicate

removed

agent-steps

25 / 5 / 0

25 / 5 / 0

25 / 5 / 0

0 / 30 / 0

20 / 10 / 0

0 / 30 / 0

0 / 30 / 0

Coordinate macro

30 / 0 / 0

0 / 30 / 0

30 / 0 / 0

30 / 0 / 0

25 / 5 / 0

0 / 20 / 10

0 / 15 / 15

CSS selectors (DevTools style)

30 / 0 / 0

30 / 0 / 0

0 / 30 / 0

30 / 0 / 0

25 / 5 / 0

30 / 0 / 0

0 / 30 / 0

Playwright role locators

30 / 0 / 0

30 / 0 / 0

30 / 0 / 0

0 / 30 / 0

25 / 5 / 0

0 / 30 / 0

0 / 30 / 0

Task

Contender

base

moved

renamed

relabeled

overlay

duplicate

removed

login

agent-steps

sssss

sssss

sssss

sssss

sssss

sssss

sssss

login

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

sssss

login

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

login

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

search

agent-steps

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

search

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

DDDDD

search

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

search

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

invites

agent-steps

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

invites

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

SSSSS

DDDDD

DDDDD

invites

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

invites

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

checkout

agent-steps

SSSSS

SSSSS

SSSSS

sssss

sssss

sssss

sssss

checkout

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

sssss

sssss

sssss

checkout

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

sssss

SSSSS

sssss

checkout

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

sssss

sssss

sssss

settings

agent-steps

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

settings

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

sssss

settings

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

settings

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

projects

agent-steps

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

projects

Coordinate macro

SSSSS

sssss

SSSSS

SSSSS

SSSSS

DDDDD

DDDDD

projects

CSS selectors (DevTools style)

SSSSS

SSSSS

sssss

SSSSS

SSSSS

SSSSS

sssss

projects

Playwright role locators

SSSSS

SSSSS

SSSSS

sssss

SSSSS

sssss

sssss

Not recorded by agent-steps: login (agent-steps refused to save the recipe: secret-field). Those rows count as safe failures for agent-steps on every variant, BASE included: it never acted.

agent-steps' own checks: replaying each recipe once on BASE with its text check, agent-steps reported the right result for 5 of 5 tasks.

agent-steps with the AI finishing what stops

Results file: bench/results/repair-2026-11-18.json · 2026-11-18T20:47:43.202Z · the saved steps run first with no model; where they stop, anthropic/claude-sonnet-5.5 (Claude Code 2.1.322 (Claude Code)) is told how far they got and finishes in the same browser · N = 2 per scenario · 51 Claude runs · spend $1.89.

Variant

Runs: success / safe / dangerous

Finished by the saved steps alone

Handed to the AI

AI cost per handover

base

12 / 0 / 0

10/12

2/12

$0.064

moved

12 / 0 / 0

10/12

2/12

$0.033

renamed

12 / 0 / 0

10/12

2/12

$0.032

relabeled

12 / 0 / 0

0/12

12/12

$0.048

overlay

10 / 0 / 2

8/12

4/12

$0.040

duplicate

12 / 0 / 0

0/12

12/12

$0.040

removed

0 / 12 / 0

0/12

12/12

$0.030

With a model: the agent every time vs. agent-steps

Results file: bench/results/model-2026-11-18.json · 2026-11-18T16:57:06.353Z · model anthropic/claude-sonnet-5.5 through Claude Code 2.1.322 (Claude Code) (claude -p, routed through OpenRouter) · agent-browser 0.37.0 · 39 Claude runs in total.

Cost is estimated from the token counts Claude Code reports, at $2 / $10 per million input / output tokens, $2.5 per million cache-write and $0.1 per million cache-read tokens (API list prices for this model). "CLI cost" is Claude Code's own total_cost_usd. Billed by OpenRouter for the whole suite: $1.90.

Task

Approach

Runs

Success

Dangerous

Tokens (total)

Est. cost

CLI cost

Wall time (total)

invites

A: agent every time

10

10/10

0

918,760

$0.31

$0.31

68 s

invites

B: agent once (recording) + replays

10

10/10

0

213,925

$0.05

$0.05

36 s

invites

B2: repeat asked of the agent (steps_run)

2

2/2

0

229,138

$0.05

$0.05

16 s

checkout

A: agent every time

10

10/10

0

1,638,958

$0.52

$0.52

119 s

checkout

B: agent once (recording) + replays

10

10/10

0

225,088

$0.06

$0.06

22 s

checkout

B2: repeat asked of the agent (steps_run)

2

2/2

0

301,459

$0.08

$0.08

20 s

search

A: agent every time

10

10/10

0

1,846,705

$0.44

$0.44

88 s

search

B: agent once (recording) + replays

10

10/10

0

280,393

$0.06

$0.06

17 s

search

B2: repeat asked of the agent (steps_run)

2

2/2

0

225,195

$0.04

$0.04

12 s

Recordings made by the agent in B: invites: 2 steps, checks ["You and Tim Kilroy are now connected"]; checkout: 9 steps, checks ["Review your order"]; search: 4 steps, checks ["Sony WH-1000XM5 Wireless Headphones"].

On pages we did not write: Stagehand's eval sites

Results file: bench/results/external-2026-11-18.json · browserbase/stagehand evals @ 8901460d69e2aef8eb639daa0b2dda3fe2e47a8f · pages served from https://browserbase.github.io/stagehand-eval-sites/sites/ · anthropic/claude-sonnet-5.5 · one run per arm per task, three replays · spend $1.86.

Task (Stagehand eval)

AI alone

Recording

Replays, no AI: before the fixes

Replays, no AI: now

shadow_dom

done

saved

3/3

3/3

osr_in_oopif

done

saved

3/3

3/3

ionwave

done

saved

3/3

3/3

simple_google_search

done

saved

3/3

3/3

custom_dropdown

done

saved

3/3

3/3

hidden_input_dropdown

done

saved

3/3

3/3

login

done

saved

3/3

3/3

radio_btn

done

saved after toggle-element

3/3

3/3

checkboxes

done

refused (unknown-tool, unlabelled-element)

-

-

iframe_same_proc

done

saved

3/3

3/3

tab_handling

done

saved after unknown-tool

3/3

3/3

spif_in_osr

done

saved after unlabelled-element

0/3

3/3

os_dropdown

done

saved after unknown-tool, unknown-target

3/3

3/3

google_flights

done

saved after unknown-tool

3/3

3/3

amazon_add_to_cart

done

refused (unknown-tool, unknown-target)

-

-

Other "cache, then replay" tools: Stagehand and Midscene.js

Results file: bench/results/cached-2026-11-18.json · model google/gemini-3-flash-preview via https://openrouter.ai/api/v1 · Stagehand 3.7.3 · Midscene 1.17.5 · N = 5 · spend $0.5311.

Tool

Variant

Runs: success / safe / dangerous

Runs with a model call

Model cost

stagehand

first-run

6 / 0 / 0

6/6

$0.018

stagehand

base

30 / 0 / 0

0/30

$0.000

stagehand

moved

15 / 0 / 15

20/30

$0.046

stagehand

renamed

30 / 0 / 0

30/30

$0.088

stagehand

relabeled

30 / 0 / 0

0/30

$0.000

stagehand

overlay

25 / 5 / 0

5/30

$0.004

stagehand

duplicate

5 / 5 / 20

20/30

$0.020

stagehand

removed

0 / 5 / 25

5/30

$0.003

midscene

first-run

6 / 0 / 0

6/6

$0.046

midscene

base

30 / 0 / 0

0/30

$0.000

midscene

moved

20 / 0 / 10

20/30

$0.069

midscene

renamed

30 / 0 / 0

30/30

$0.117

midscene

relabeled

26 / 4 / 0

30/30

$0.044

midscene

overlay

25 / 5 / 0

5/30

$0.005

midscene

duplicate

0 / 5 / 25

20/30

$0.032

midscene

removed

0 / 20 / 10

20/30

$0.038

Results file: bench/results/cached-stagehand-sonnet-2026-11-18.json · model anthropic/claude-sonnet-5.5 via https://openrouter.ai/api/v1 · Stagehand 3.7.3 · Midscene 1.17.5 · N = 2 · spend $0.522.

Tool

Variant

Runs: success / safe / dangerous

Runs with a model call

Model cost

stagehand

first-run

6 / 0 / 0

6/6

$0.101

stagehand

base

12 / 0 / 0

2/12

$0.009

stagehand

moved

8 / 0 / 4

8/12

$0.121

stagehand

renamed

12 / 0 / 0

12/12

$0.210

stagehand

relabeled

12 / 0 / 0

1/12

$0.005

stagehand

overlay

10 / 2 / 0

2/12

$0.010

stagehand

duplicate

2 / 2 / 8

8/12

$0.049

stagehand

removed

0 / 4 / 8

4/12

$0.017

In one line: agent-steps would rather stop than guess. Alone it finishes fewer changed pages than tools that guess, and never does the wrong thing; with the AI taking over where it stops, it finishes nearly all of them.

The site. A local fixture (a tiny Node server, bench/fixture/site.ts) with six tasks: log in (email and password), search and open a result, accept one invitation in a list of cards, a three-page order form (address fields, a delivery radio button), two dropdowns and Save, and archive one project with a confirm dialog. Each task is recorded on a BASE page and replayed on BASE and six variants:

Variant

What changes

moved

Same elements, different layout: blocks added above, columns swapped, list items reordered

renamed

Every CSS class and id renamed, an extra wrapper element; text and layout unchanged

relabeled

Visible labels changed to a synonym or a Spanish translation ("Accept" → "Aceptar", "Postal code" → "ZIP code")

overlay

A cookie banner fixed over the bottom 200 px of the 1280×720 viewport

duplicate

A look-alike of the target with the same role and name: a second "Tim Kilroy" invitation, a sponsored result with the same product name, a billing address form above the shipping one, a second "Language" dropdown, a shared project also called "Atlas"

removed

The target is gone; nobody can finish, the only good outcome is to stop

The ground truth is kept by the server, not by the contenders: every consequential action (a form post, an accepted card, an archived project, a product page opened, a cookie choice) is logged. A run is a success when every required action happened with the right data and nothing else consequential did; dangerous when anything consequential happened that the task didn't ask for (another card accepted, the address typed into the billing form, the wrong project archived, the cookie banner clicked, a password submitted in the wrong field); a safe failure otherwise: it stopped, or did only harmless things such as running a search.

The no-model contenders, each recorded once on BASE and replayed with nothing in the loop:

  • agent-steps: a scripted "agent" reads agent-browser's snapshot before every action and acts on the @ref of the intended element, through agent-steps' recorder and the real agent-browser MCP server; the replay is agent-steps' own replay() with the same driver.

  • Coordinate macro: the click point of each element on BASE, replayed with mouse.click(x, y) and typed keys, with a 300 ms pause between steps the way macro tools replay. Dropdowns are driven by clicking them and typing the start of the option, which is what a keystroke recorder captures.

  • CSS selectors: the selector Chrome DevTools' "Copy selector" gives (an id where there is one, otherwise classes unique among siblings or :nth-child), replayed with Playwright locators.

  • Playwright role locators, the strongest baseline: getByRole(role, { name, exact: true }), scoped the way a careful engineer would (getByRole('listitem').filter({ hasText: 'Tim Kilroy' }), getByRole('dialog')).

All Playwright contenders and agent-steps wait at most 3 s for an element, in the same 1280 × 720 window.

With a model, four more arms:

  • agent-steps + AI: what an app does. The saved steps run first with no model; where they stop, Claude is told how far they got and why, and finishes in the same browser. It is told nothing about what to do instead.

  • Stagehand and Midscene.js: each step given as an instruction (Click the "Accept" button…), cached on BASE, then replayed from a fresh copy of that cache; both call their model when the cache misses.

  • The agent every time against record once, replay: three of the tasks, ten times each, for cost and time.

Scenarios, not runs. A replay with no model is deterministic, so five repeats of one task on one version of the page are one scenario. The charts count scenarios (6 tasks × 7 versions of the page = 42): finished only if every repeat finished, wrong if any repeat did something the task didn't ask for.

Someone else's pages. 15 tasks from Stagehand's own eval suite, on the pages Stagehand's team wrote for them: the AI does each once alone, once while recording, and the saved recipe is replayed three times with no model. Success is the eval's own condition, read from the page afterwards.

  • Written by the side being measured. The fixture, its variants, the judge and the baselines were written for this bench by agent-steps' side. The "look-alike" and "target gone" variants are built so that stopping is the only right answer, which suits a tool that stops. Two things changed after a first look at results: the order form moved from /t/checkout/ to /t/order/ (agent-steps refuses to record opening a URL that names a final action), and the coordinate macro's dropdown typing was fixed because it failed on BASE itself.

  • Small and synthetic. Six tasks on one local site, one change at a time, and every task is easy enough that Claude does it ten times out of ten. The 15 external tasks are static test pages too, not live products, with one run per arm. Nothing here measures a real site after a month of changes.

  • What counts as wrong is our call. With the AI finishing what stops, the one wrong scenario is Claude rejecting a cookie banner that covered a button. The judge counts a cookie choice as something the task didn't ask for; a person might well have done the same.

  • Not the same model everywhere. Midscene needs a vision model with grounding and cannot use Claude, so it ran on Gemini 3 Flash. Stagehand has a row for each model it was run with. The agent arms use Claude Sonnet 5.5 through Claude Code. Costs across those are not comparable.

  • Not the same instructions. The agent arms get the whole request, including "do not accept anyone else". Stagehand and Midscene get one instruction per step, which is how they are used.

  • The baselines are reasonable, not the best possible. The role locators were scoped by hand from BASE, the way a careful engineer writes them. CSS selectors did best because the fixture gives elements stable ids; they break on every renamed id.

  • Not the same browser. agent-steps runs through agent-browser 0.37 with its own Chrome; the other contenders use Playwright's Chromium, at the same window size. Wall times include agent-browser's MCP round trips, so compare times only roughly.

  • What agent-steps will not record, by design. A password field (the login task), and a step named like a final action. The other contenders replay those happily, password included; agent-steps counts as a safe stop, and with the AI the AI does that task itself.

  • Labels are its weak spot, and role locators share it. A renamed label ("Aceptar") makes it stop, where coordinates and CSS selectors that don't depend on text still work.

  • The cost chart is three tasks. The saving when an agent calls the recipe over MCP was measured before answers were slimmed to text (0.2.0), which makes that path cheaper; it has not been measured again.

  • Bugs these runs found, fixed since. Text checks that read only what can be acted on, or could never match what a field shows; a key press sent before the page was ready; refusals that reached the agent only after the task was done; answers too long for the agent to read on large pages. The external table shows the replays before and after.

  • Not run. The desktop (Cua) bench: it takes over the screen. Passmark (needs Redis and two model providers) and browser-use (no comparable cache-and-replay mode) have no adapter.

  • Versions. Stagehand 3.7.3, the last with a local cacheDir (Stagehand 4 moved caching to its hosted runtime); Midscene 1.17.5 with cache: { id, strategy }. A local proxy counts each of their model calls and its cost.

npm install
npm run bench:install                 # bench/ dependencies (pinned) and Playwright's Chromium
npm run bench                         # no-model suite, N=5 (about an hour), then renders this section
node bench/run.ts --n 1 --tasks invites --variants base,duplicate   # a quick subset
npm run bench:model                   # needs the `claude` CLI, signed in; at most 40 Claude runs
node bench/model.ts --via openrouter  # the same through OpenRouter (OPENROUTER_API_KEY in bench/.env), spend cap --budget 6
npm run bench:cached                  # Stagehand and Midscene; needs OPENROUTER_API_KEY in bench/.env
node bench/repair.ts --via openrouter # saved steps first, the agent where they stop
node bench/external/run.ts            # 15 Stagehand eval tasks on Stagehand's own pages

agent-browser is found through AGENT_BROWSER_BIN, then PATH; the bench runs it in its own session and namespace (agent-steps-bench), so it never touches a browser you are using.

Desktop (Cua), written but not run. bench/desktop/cua-bench.ts records a TextEdit and a Notes task on a window at a fixed place and size, then replays agent-steps and a coordinate macro with the window moved, resized, and both. It moves windows, clicks and types for real, so it was not run here. To run it: install cua-driver, give your terminal Accessibility permission, close TextEdit and Notes, then node bench/desktop/cua-bench.ts --driver "cua-driver mcp" and keep your hands off the machine until it finishes.

Development

npm install
npm run typecheck
npm test
npm run build

Contributing

Issues and pull requests are welcome. Run npm test, npm run typecheck and npm run lint before sending a change; anything that touches what may be saved or repeated needs a test that shows the unsafe case is refused.

License

MIT © 2026 Legalcap. Created by Borja Soler; born inside Openpals.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      D
      maintenance
      Self-learning macOS automation MCP server that remembers successful patterns and avoids repeated failures, enabling AI agents to execute multi-step workflows via reusable recipes.
      11
      9 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables MCP clients such as Claude Code, Codex, Cursor and Gemini CLI to control a Mac, Windows or Linux desktop by clicking, typing and reading real applications that have no API, using the accessibility tree rather than pixel guessing. It also lets agents replay saved tasks as skills without spending tokens and can type stored passwords into fields without exposing their values.
      1
      106 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables Claude Code and Codex to turn repeated web lookups into saved "recipes" that run locally in a dedicated Chrome, deterministically filling forms, reading pages, and verifying results with screenshots and freshness checks. Also lets users record a manual demonstration that gets replayed in a clean browser to build a new flow, and diagnose or repair flows when a site changes.
      1
      MIT