jev-browser-sidekick-mcp

An MCP server that drives a browser with Jev, TypeSafe's decision model. You write the steps. Jev chooses which control on the page carries out each one. The server does the clicking and the waiting, and it keeps the budgets.

The server has two tools. run_action does the browser work. use_jev_raw answers one typed question with no browser involved.

Jev named the package. Six candidates went into use_jev_raw with the download counts of every competing package and the trade-offs of each name written out. It picked this one at a probability of 0.66, against 0.17 for the runner-up, and it cost $0.000058 to ask. The maintainer had argued for a different name and lost.

Install agentic-playwright-mcp alongside it

Install agentic-playwright-mcp as well. This server runs at its best with that one beside it, and the two are built to be used together.

This server does not start its own browser. It attaches to the Chrome that agentic-playwright-mcp already runs, at http://127.0.0.1:9223. One Chrome, one profile, one set of cookies, shared by both servers and by your agent. A site you signed into through your Playwright tools is still signed in when a step runs here, and a cart this server filled is still there when you look at it yourself.

That shared session is what makes the pair stable. Pass a targetId from your Playwright tools into run_action, and pass the targetId values it returns back the other way, and both sides act on the same tab. When a step stops as blocked on a password or a captcha, the tab is one you already hold, so you can finish it in place and resume.

If no Chrome is listening, this server starts one of its own. It works, but nothing else can see that browser, so you lose the handoff and the sign-in you already had.

Related MCP server: MCP QA Demo Server

Install

Install both globally. agentic-playwright-mcp is a peer dependency, so npm pulls it in on its own, but naming it here also puts its commands on your PATH.

npm install -g agentic-playwright-mcp jev-browser-sidekick-mcp

That gives you four commands, a long form and a short form for each package.

npm writes a .cmd and a .ps1 shim next to each command on Windows, so all four work unchanged in PowerShell, in cmd, in Git Bash, and in WSL.

To skip the install and fetch on demand, use npx --yes with the full package name. npx resolves a package name rather than a command name, so npx --yes jev-bro does not work.

npx --yes jev-browser-sidekick-mcp doctor

Save an API key

jev-bro setup --provider openrouter --api-key "$OPENROUTER_API_KEY"

In PowerShell, use $env:OPENROUTER_API_KEY. In cmd, use %OPENROUTER_API_KEY%.

The command writes ~/.jev/.env. The MCP server, the CLI, and runAction all read that file.

To use a TypeSafe key instead, run this command.

jev-bro setup --provider official --api-key "$TYPESAFE_API_KEY"

To add this server to ~/.cursor/mcp.json, pass --install-cursor. The command leaves the agentic-playwright-mcp entry alone.

To override ~/.jev inside this repo, copy .env.example to .env and set the same keys.

Check the setup

jev-bro doctor

The output includes jev decision: ok and a playwright: ok line with a targetId.

Add the servers

Put both in the host mcp.json. Neither needs a path, and npx fetches them on first run.

{
	"mcpServers": {
		"playwright": {
			"command": "npx",
			"args": ["--yes", "agentic-playwright-mcp"]
		},
		"jev": {
			"command": "npx",
			"args": ["--yes", "jev-browser-sidekick-mcp"]
		}
	}
}

Start the playwright entry first, or just let the host start both. This server looks for that Chrome on every call, so the order only decides whether the first call attaches or opens its own browser.

Write the steps

Jev picks among labelled options and returns typed answers. It does not read plans and does not write text, so you write the plan and each step hands Jev one choice.

Each step acts on one page. Use the words that appear on the screen, because Jev matches labels literally.

Adding one item to a cart is three steps, one per page.

["search colgate toothpaste", "open the best matching colgate toothpaste result", "click add to cart"]

A single add colgate toothpaste to cart still runs, but it never leaves the results page, so it presses whatever on that page carries those words.

Each step runs against the live page. The decision also sees the series motive and a short steps_done line for each finished task, so Jev can refuse a step that earlier work already covered. Groups running in parallel share nothing, so each one sees only its own motive and its own finished steps.

Repeat steps until the page says to stop

A loop runs its body once a round, then Jev reads the live page and answers one question: does the condition hold yet? Clearing a cart is one use of it. Any page that hands back one item at a time needs the same shape.

["repeat click remove until the cart is empty"]

The condition ends the loop, not a control disappearing. A site that redraws its list between rounds offers no controls for a moment, and reading that as "finished" reports an untouched cart as cleared. Write the condition as something the page shows, such as the cart is empty, rather than done.

A step that names its own control derives its condition, so keep clicking Load more and clear cart still work as written. Between rounds the server waits for the page's own scripts rather than a fixed pause, because a row that a site deletes over the network lands whenever its request comes back.

For a body of more than one step, pass a loop object in tasks.

{
	"tasks": [
		"open the cart page",
		{ "loop": { "tasks": ["click remove", "click confirm"], "until": "the cart is empty", "maxRounds": 20 } },
		"read the cart"
	]
}

No loop runs forever. Five things end one: the condition, the round ceiling (maxRounds, 12 by default and never above 50), the call's own deadline, the step budget, and two rounds that change nothing. The step reports completed when the page showed the condition, partial when a ceiling stopped it with work left, and rejected with no_control when the body pressed nothing at all. Every loop step carries the rounds it ran.

When the outcome already holds

A step that names a control the page no longer carries comes back rejected. For example, once an item is in the cart, a product page can replace "Add to cart" with "Go to cart", so click add to cart finds nothing.

Two different things cause that, and the caller acts differently on each: the page cannot do the step at all, or the page already shows the step's outcome. The server does not guess from the URL or the title, because that guess marked steps finished that had never run. Jev reads the page instead, and an outcome already in place comes back as rejected with reason already_done.

The series carries on past an already_done step, because the ground the next step stands on is there, whoever put it there. The group still reports rejected, so a run reads honestly, and verified says where it landed.

Call run_action

Steps in tasks run in order on one tab.

Errands that do not depend on each other belong in groups, and they run at the same time, one tab each. Two sites, two accounts, or two separate searches are one call with two groups, never two calls. Use groups first, and fall back to a single series only when every step needs the page the step before it left. Series that share a targetId run one after another, because they share a tab.

{
	"groups": [
		{
			"id": "cart",
			"targetId": "TAB_A",
			"expect": "subtotal",
			"tasks": [
				"open the cart page",
				"clear cart",
				"search colgate toothpaste",
				"open the best matching colgate toothpaste from the results",
				"click add to cart"
			]
		},
		{
			"id": "other",
			"targetId": "TAB_B",
			"tasks": ["open example.com"]
		}
	]
}

Read the result

The result has is_finished, targetIds, groups, status, summary, and a handoff when something stopped. Each group has its own tasks, steps, and counts. snapshot appears only when you set returnSnapshot, and usage, elapsedMs, and per-step ms only when you set debug.

Every step has its own status.

A group's status is worst-case across its steps, so a group where eight of ten steps completed still reads as rejected. Read counts for what actually happened.

{ "status": "rejected", "counts": { "completed": 12, "rejected": 1 } }

Later steps depend on earlier ones, so a step that does not complete ends its series, and the rest come back as skipped naming the step that stopped them. Set noFail on a series whose steps are independent, and it runs them all.

One reason is exempt. A step that came back rejected with already_done did not stop anything, because what the next step needs is already on the page.

Why a step was turned down

A standing none choice that wins uses the same reasons: no_control for click or press, no_match for pick.

Prove the run worked

A completed status is Jev's claim. Set expect to the text that proves it, and the server reads the final page for that text. Matching ignores case and spacing.

{ "tasks": ["open the cart page"], "expect": "subtotal (3 items)" }

Pick text that only the finished state produces. Subtotal (3 items) works. A product name does not. Shops repeat product names in recommendation rails, so that text matches even on an empty cart. The result quotes the words either side of the match in proof, so you can see which it matched.

{ "verified": true, "proof": "…All Carts Subtotal (3 items): ₹509.00 Proceed to Buy…" }

The check runs when the final step of the series completed, including under noFail after an earlier rejection. A series whose final step never ran reports proof: not checked, rather than claiming the text was missing from a page it never reached.

Pick up a stopped run

When a series stops early, the result includes a handoff.

{
	"resumable": true,
	"targetId": "TAB_A",
	"url": "https://www.amazon.in/ap/signin",
	"stoppedAt": "click add to cart",
	"status": "blocked",
	"reason": "credentials",
	"remaining": ["open the cart page"],
	"recent": [{ "step": 4, "operation": "CLICK", "detail": "clicked e42" }]
}

The tab is still open and you already share it, so you can finish the step through your own Playwright tools, ask the user, or call run_action again with that targetId and the remaining steps.

Pages that block a step

These reasons come back as blocked with a handoff, because you own the same tab and can still act.

Finding the words and believing them are separate. The harness spots the words a sign-in wall or a challenge uses, and Jev reads the page and confirms the page is really demanding one. A news story titled "Solving a corn puzzle with CP-SAT" carries the same word a challenge does, and stopping a series on that costs more than the check saves.

This server never fills a password, a one-time code, or a captcha itself. Type the value in the shared tab, hand it to the user, or stop. Pass ordinary strings such as an email or a postcode in values.

A missing control is not blocked. It is rejected with no_control or no_match, as in the reason table above.

When the provider fails

A fault in the Jev API or the network is not a page answer. The step returns error with a reason that names the cause. An HTTP proxy that returns HTML with status 200 is the exception: that comes back as blocked with reason proxy_interstitial, because only you can clear the proxy.

The summary includes the HTTP status and a short provider message. The series stops even when noFail is set. handoff.resumable stays true when a tab exists, so you resume with that targetId and the remaining steps instead of redoing work that already finished.

Keep a call short

A call returns after its own timeout, which defaults to 90 seconds. On expiry you get a normal result holding the steps that finished and a handoff naming the rest, so you resume by calling again with that targetId and the remaining steps.

A call your MCP client drops is the case worth avoiding. The browser work still happens, you never see the result, and running the same steps again does them twice. Keep timeoutMs under your client's own transport timeout and resume instead of asking for one long call.

Token usage and timing

usage and elapsedMs appear only when you set debug, alongside tracePath. Every group and task then carries its own ms.

usage sums what each API response reported. Nothing in it is estimated.

TypeSafe returns tokens and no price, so costUsd is absent on a direct TypeSafe key and present through OpenRouter.

decisions: 0 is normal and not a failure. A search, a destination, and a control labelled exactly what the step said all resolve without a judgment call, so a run made only of those asks Jev nothing. A recent two-site run of 26 steps cost 19 decisions and about $0.001.

Ask Jev without a browser

use_jev_raw sends state and typed questions straight to Jev. Use it whenever a decision has more than one defensible answer and you are about to pick on instinct: which fix to do first, which name to ship when each has a real trade-off, whether a draft meets a bar you can write down, whether a step is risky enough to stop and ask the user.

You get a probability for every option and a confidence, so a close call reads as close. Ask every question you have in one call, because they share the state and answer in parallel. Three questions over a page of state run about a tenth of a cent.

{
	"state": { "bug": "Checkout throws on an empty cart.", "fixes": { "guard": "...", "schema": "..." } },
	"questions": {
		"pick_fix": {
			"type": "choice",
			"instructions": "Which fix in `fixes` removes the cause of `bug` rather than hiding it?",
			"criteria": { "guard": "Adds a runtime check.", "schema": "Makes the broken state unrepresentable." }
		}
	}
}

The answer has the chosen option, a probability for every option, a confidence, and the tokens the API counted. The server publishes a jev://raw-decisions resource with the question types, the rules for writing criteria, and the size limits. Read it before the first call.

Nothing says the state has to be about code. If you are the sort of person who stands in the cereal aisle for ten minutes, your sidekick will happily take that one too. Give it the four job offers, the three flat listings, the cat names, or what to cook tonight, write down what you actually care about in criteria, and it tells you which one it likes and how sure it is. A tenth of a cent for a friend who never answers "I don't know, what do you want to do" is a fair trade. It already picked this package's name, and it was right.

What the server already handles

Leave these out of the plan. The server waits for loads, follows a link that opens its own tab, recovers element references that went stale between the snapshot and the click, and skips invisible controls that carry real labels.

Finding a control and choosing it are separate. When a step names a control, the server collects every control on the page carrying those words, including ones drawn as plain text with no accessibility role. Jev picks one of them, or the standing none option when none fits. A pick step does the same over entry candidates that name the subject. When nothing fits, the step returns rejected instead of pressing something else.

Jev reads text only, so this server works from the page's own text and takes no screenshots. A control drawn without text is found by its DOM text instead.

A step that opens an entry is done only once the page shows that entry. A click that navigates, opens its own tab, or draws an overlay showing the entry all count. A click that leaves the page exactly as it was does not count, so the step tries something else rather than reporting success.

Traps on real sites

Each of these cost real runs in testing. Shopping sites found them, but nothing about them is particular to shopping.

One tab, two storefronts. A site can run more than one storefront in the same tab, and its search box keeps you in whichever one the tab is already in. Results in the wrong storefront open overlays rather than their own pages, so a click step finds nothing. Pass a startUrl that names the storefront you want. For example, Amazon Fresh sits beside the main Amazon store in one tab.

One site, two collections. A site can keep more than one cart, list, or queue, and show a count that spans all of them, so a collection you just emptied still reads as full. The page that lists everything usually shows the other collection read-only, with a link in place of the per-row controls, so a clear step there presses nothing and returns rejected with no_control. Open the collection that owns the items first. For example:

["open the cart page", "click Go to Fresh Cart", "clear cart"]

A page that offers a different route. When the usual action is unavailable, a page often puts another control in its place. The step returns rejected with no_control or other_route. That is the page's answer, not a failure to retry. For example, a listing with no default offer shows "See All Buying Options" where "Add to cart" would be.

A near miss in place of a match. A search can answer with something close rather than the thing you named, most often when the real item sits behind a sign-in or in another storefront. Expect no_match, or check the substitution with a read step.

Debug a run

Set debug on a call to write every browser call and Jev decision to a JSONL file. The result names it in tracePath.

To trace every run whatever the caller passes, add --debug to the server command.

{
	"mcpServers": {
		"jev": {
			"command": "npx",
			"args": ["--yes", "jev-browser-sidekick-mcp", "--debug"]
		}
	}
}

Traces land in ~/.jev/traces, one file per run.

Run one goal from the CLI

jev-bro run "Open example.com and click More information" --snapshot

Add --expect to check the final page, --task to write a series, and --groups-json for parallel groups.

Call runAction from code

npm install jev-browser-sidekick-mcp

import { runAction } from "jev-browser-sidekick-mcp";

const result = await runAction({
	tasks: ["open the cart page", "clear cart"],
	expect: "your cart is empty",
});

console.log(result.status, result.verified, result.usage.totalTokens);