Public Browser

Lets Claude Code and Cursor drive Chrome — with your real, logged-in profile. On the same 30-test benchmark page it used 30% fewer tokens, 25% less money, 41% fewer tool calls and 40% less time than Playwright MCP at the same pass rate — two runs each, 2026-09-03, driver Claude Opus 5, raw data in the repo (Benchmarks, including where it loses). Its own tool definitions are 34% smaller since v2.10.4 (7,607 → 4,990 tokens, reproduce with node scripts/token-count.mjs). Direct CDP, a11y-tree refs, multi-tab ready — 2,360 TypeScript tests, 237 Python tests.

Built for Claude Code, Cursor, and any MCP-compatible client.

Looking for an alternative to Playwright MCP, Browser MCP, or claude-in-chrome? Public Browser talks to Chrome directly via the DevTools Protocol — no Playwright dependency, no Chrome extension bridge, no single-tab limit. One command to install, zero config. See benchmark comparison below.

Why Public Browser?

Every Chrome MCP server has the same problem: bulky responses, too few reliable refs. Screenshots return 10-30x more context payload than text trees. Selector-based refs break the second the DOM rerenders. Extension bridges (Browser MCP) get stuck on the connected tab. Playwright wrappers spin up a new browser instance for every session.

Public Browser fixes this. It talks directly to Chrome via CDP (same protocol Playwright and Puppeteer use internally), returns an accessibility-tree-based reference map, and caches it across calls so click(ref: 'e5') and type(ref: 'e7', ...) survive scrolls and DOM updates.

Benchmark rows below are April 2026, 35-test suite, Opus 4.6 unless a cell also gives a September value. Cells marked Sep come from the blind September 2026 re-run (35-test page, 30 scored, driver claude-opus-5, two runs per required server; one browser-use run) — run files public-browser-run1/2.json, playwright-mcp-run5/6.json, chrome-devtools-mcp-run3/4.json, browser-use-run6.json in test-hardest/results/. Cross-suite comparison is not valid; see Benchmarks.

P95 is not computed the same way in both rows: the April values are the largest per-tool P95 (by_tool[].p95_chars, jq index floor((n-1)·0.95)), the September values are a nearest-rank P95 over all MCP calls of a run. Do not read the April and Sep P95 numbers as one series.

What changed since April. In April 2026 Public Browser's page views were 5.4x smaller than Playwright MCP's; Playwright MCP 0.0.80 has since made its snapshot format much more compact, so in the September runs its individual responses are smaller than ours (Ø 740 and 656 chars against 1,298 and 1,214) and the tool-response payload summed over a whole run is roughly a tie (109,075 / 104,432 chars against 101,478 / 99,147) — "smallest responses" is no longer a claim we can make. What the model actually consumed over the whole session is a different story: 6.3M and 6.5M tokens against 8.8M and 9.6M (~30% fewer), $3.41 and $3.35 against $4.28 and $4.78 at Opus 5 list price (~25% less). That gap comes from the second thing left standing: ~41% fewer tool calls (84 and 86 against 137 and 151) and ~40% less time to finish (281s and 296s against 468s and 493s, page timer; 34% on the full wall clock, 331s and 346s against 501s and 527s) at the same reliability, 30/30 in each of those four runs. Every call re-reads the whole conversation so far, so fewer calls means fewer re-reads — that is where the token and cost gap comes from, not from smaller responses. The advantage moved from "cheaper per look" to "cheaper per task", and the headline was changed accordingly.

September 2026 data, both runs per server, against Playwright MCP 0.0.80. Rows where the bar runs past the line are rows Public Browser loses. Method and the full table: Benchmarks.

Quick Start

Install in Claude Code

One command — installs globally for all projects:

claude mcp add --scope user public-browser npx -y public-browser@latest

Important: after claude mcp add you must fully quit and reopen Claude Code. /mcp reconnect is not enough — Claude Code reads the mcpServers config only at session start and caches it. After the restart, the first tool call auto-launches Chrome visible (no headless, no port setup). Done.

To enable parallel Python Script API access, add --script to the args: claude mcp add --scope user public-browser npx -y public-browser@latest -- --script

Install in Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "public-browser": {
      "command": "npx",
      "args": ["-y", "public-browser@latest"]
    }
  }
}

For parallel Python Script API access, use "args": ["-y", "public-browser@latest", "--", "--script"]

Install in Cline

Add to your cline_mcp_settings.json:

{
  "mcpServers": {
    "public-browser": {
      "command": "npx",
      "args": ["-y", "public-browser@latest"]
    }
  }
}

Install in other MCP clients

Any client that supports stdio MCP servers: npx -y public-browser@latest with no arguments.

Try it — your first prompt

After installing, paste this into your AI coding assistant:

Open mcp-test.second-truth.com, read the page, and fill the contact form with Name "Test User" and Email "test@example.com".

This exercises three core tools in sequence: navigate loads the page, view_page reads the accessibility tree with stable element refs, and fill_form fills multiple fields in one call. You should see Chrome open, the page load, and the form filled — all without writing a single line of code.

Uninstall

claude mcp remove --scope user public-browser

Chrome Profiles

By default, Public Browser starts Chrome with a fresh temp profile — no cookies, no logins, no extensions. For tasks like research on sites that block anonymous visitors, you can launch Chrome with your real profile instead.

List available profiles

npx public-browser profiles

Launch with a profile

Three ways — pick whichever fits your setup:

# CLI flag
npx public-browser --profile "Julian"

# Environment variable
PUBLIC_BROWSER_PROFILE="Julian" npx public-browser

# MCP tool (call BEFORE any browser interaction)
configure_session({ profile: "Julian" })

When using a real profile, Public Browser preserves extensions, cookies, logins, and sync. It creates a lightweight wrapper directory with a symlink to your real profile data — Chrome gets a "non-default" data dir (required for remote debugging) while using your actual profile.

If Chrome is already open

Public Browser detects this via lock-file inspection. If Chrome is running with remote debugging enabled, it attaches via CDP. If not, it shows a clear error asking you to close Chrome first.

Script API (Python)

A second way to use Public Browser — deterministic browser automation from Python, without an LLM in the loop. Scripts use the same tool implementations as the MCP server (Shared Core) — every improvement to click, navigate, fill_form etc. automatically benefits your scripts too. The MCP server handles AI-driven workflows; the Script API is for repeatable scripts you write yourself.

How fast that is without an LLM: a scripted run of the 24-test version of the benchmark suite finished the whole suite in 21 seconds (type: mcp-scripted, 2026-04-04). That number says what deterministic scripting costs, not how Public Browser compares to other MCP servers — every cross-server comparison in Benchmarks is LLM-driven on both sides.

Installation

pip install publicbrowser

Or, from a source checkout, install the local package:

python -m pip install ./python

Chrome.connect() auto-starts the Public Browser server as a subprocess via a local public-browser binary or the npx fallback — no manual Chrome launch or port setup needed.

Legacy single-file alternative: For quick prototyping you can copy python/publicbrowser_standalone.py into your project. This uses v1 direct CDP and does not benefit from server-side improvements — use the local publicbrowser package for the full Shared Core experience.

How it works

Python Script                        Escape Hatch (Power User)
    |                                    |
    v                                    v
HTTP POST /tool/{name}              WebSocket (CDP)
Port 9223                           Port 9222
    |                                    |
    v                                    |
Public Browser Server                    |
    |                                    |
    v                                    |
registry.executeTool()                   |
    |                                    |
    v                                    |
Tool Handler                             |
(click.ts, navigate.ts, ...)             |
    |                                    |
    v                                    v
Chrome <------------ CDP --------------->

Your script sends HTTP requests to the Public Browser server on port 9223. The server executes the exact same tool handlers that the MCP server uses — one codebase, one test suite (2300+ tests), two access paths.

Auto-Start

Chrome.connect() finds and starts the server automatically:

1. Running server — checks if port 9223 already responds, connects immediately

2. PATH binary — finds public-browser in PATH, starts it with --script

3. npx fallback — runs npx -y public-browser@latest -- --script

4. Explicit path — Chrome.connect(server_path="/path/to/public-browser") for custom setups

Example: Login + Data Extraction

from publicbrowser import Chrome

chrome = Chrome.connect()

with chrome.new_page() as page:
    page.navigate("https://competitor.example.com/login")
    page.fill({"#email": "tomek@shop.de", "#password": "***"})
    page.click("button[type=submit]")
    page.wait_for("text=Dashboard")

    for cat in ["electronics", "furniture", "toys"]:
        page.navigate(f"https://competitor.example.com/prices/{cat}")
        prices = page.evaluate(
            "[...document.querySelectorAll('tr')].map(r => r.textContent)"
        )
        save_csv(cat, prices)

chrome.close()

Methods

Escape Hatch: Direct CDP Access

For use cases the high-level API doesn't cover — network interception, console log subscriptions, performance tracing, cookie management — you can drop down to raw CDP commands:

with chrome.new_page() as page:
    page.navigate("https://example.com")

    # Enable network tracking
    page.cdp.send("Network.enable")

    # Get all cookies
    cookies = page.cdp.send("Network.getAllCookies")

    # Performance tracing
    page.cdp.send("Tracing.start", {"categories": "-*,devtools.timeline"})

The Escape Hatch communicates directly with Chrome via WebSocket (port 9222), bypassing the server. It connects lazily on the first send() call and reuses the connection for subsequent calls. Each page gets its own WebSocket routed to the correct tab.

MCP Coexistence

When the MCP server and Python scripts need to run at the same time, add --script to the MCP config. Chrome.connect() handles the rest automatically — each script works in its own tab, MCP tabs are never touched.

Enabling --script in MCP Config

Claude Code:

claude mcp add --scope user public-browser npx -y public-browser@latest -- --script

Cursor / Cline (mcp.json):

{
  "mcpServers": {
    "public-browser": {
      "command": "npx",
      "args": ["-y", "public-browser@latest", "--", "--script"]
    }
  }
}

See python/README.md for the full API reference and advanced examples.

Node Library API (multiple instances in one process)

The MCP server and the Python Script API both drive exactly one Chrome per process. When you need several browsers at once — say a read-only research browser and a separate action browser per agent — spawning one npx public-browser per instance costs 4–6 s of start-up each. createSession() runs the same session inside your own Node process instead:

import { createSession } from "public-browser";

const research = await createSession({
  cdpUrl: "http://127.0.0.1:9333",          // or cdpPort: 9333
  userDataDir: "/var/agents/a1/research",   // created if missing
  headless: true,
  stealth: false,                            // stay identifiable — see below
  downloadDir: "/var/agents/a1/quarantine",  // never deleted by us
  downloadHash: true,                        // adds sha256 to every download
  downloadNaming: "suggested",               // real filenames, not GUIDs
  cortexDir: "/var/agents/a1/cortex",        // per-instance pattern store
  inheritEnv: ["HTTPS_PROXY"],               // opt in — see Environment below
});

const action = await createSession({ cdpPort: 9334, userDataDir: "/var/agents/a1/action" });

await research.callTool("navigate", { url: "https://example.com" });
const page = await research.callTool("view_page", {});

await research.close();
await action.close();

callTool(name, params) takes the same tool names and parameters as the MCP tools (navigate, view_page, click, type, fill_form, run_plan, download, ...) and routes through the identical handlers (Shared Core).

Isolation. Each session runs in its own worker thread by default, so the module-level caches (element refs, selector cache, viewport state, stealth flag, cortex matcher) exist once per session rather than once per process — two sessions can never hand each other stale element refs.

Measured on macOS with isolation: "process", attaching to a Chrome started outside Public Browser (a worker thread saves ~40 ms):

Most of the attach cost is Chrome starting a renderer for the tab Public Browser opens for itself — an attached session never takes over tabs that belong to someone else.

A thread is not a security boundary: same process memory, same file descriptors. isolation: "process" forks one OS process per session instead — separate heap, separate descriptors, separate crash domain — for integrators whose trust model draws the line there. isolation: "inline" skips isolation altogether and is only correct when the thread runs exactly one session.

No listening CDP port (transport: "pipe"). By default Chrome is launched with --remote-debugging-port, which is what makes --attach, the Script API and reconnect-after-crash possible — and which also means every other process on the machine can drive that browser. For a session holding real logins that is a way around any permission check you perform yourself.

const action = await createSession({
  transport: "pipe",                      // no --remote-debugging-port at all
  userDataDir: "/var/agents/a1/action",
  headless: true,
});

CDP then travels over the child's stdio pipe, which only Public Browser holds: lsof shows nothing listening and a second process finds no way in. The price is everything the port paid for — no reconnect after a Chrome crash, no second client, no attach, and no named profile (Chrome rejects the pipe with a real user profile). Both contradictions fail at createSession() rather than at the first tool call. session.transport reports which mode is in use, and session.cdpPort is undefined — there is no port, and reporting the default would name whatever Chrome the user has open on 9222.

Environment. A session does not start from the host environment. It starts from a documented minimum and you widen it deliberately — an orchestrator holding cloud credentials, API keys and tokens should not hand them to a browser session just because the two share a process tree.

What a session always gets is ESSENTIAL_ENV_VARS: PATH, HOME, the temp dir, CHROME_PATH, locale/timezone, the Linux display variables and the Windows process basics. Everything else is opt-in:

// PATH/HOME/CHROME_PATH plus the proxy — and nothing else from the host.
await createSession({ inheritEnv: ["HTTPS_PROXY", "NO_PROXY"] });

// Full inheritance, the pre-2.8 behaviour.
await createSession({ inheritEnv: true });

Proxy variables are deliberately not essential: a proxy URL can carry credentials, so it is allowlisted on purpose rather than inherited by accident.

On top of that, a session never inherits Public Browser's own SILBERCUE_* / PUBLIC_BROWSER_* configuration variables — in any inheritEnv mode. Each of them has an option here, and a host-level variable, usually set for the host's own Chrome, silently redirecting a configured session is a bug, not a feature: with SILBERCUE_CHROME_HOST=10.9.9.9 in the orchestrator's environment, a session created with cdpPort: 9450 still talks to 127.0.0.1:9450. Use env to set one back deliberately.

Shutdown. close() resolves only once Chrome is actually gone — SIGTERM, SIGKILL after 5 s — so the port and the user-data-dir are free for the next launch instead of racing a process that was merely asked to exit.

One session per Chrome. Some CDP settings are browser-wide rather than per-session, Browser.setDownloadBehavior among them: two sessions attached to the same Chrome share one download directory, and whichever connected last wins.

This fails silently and it corrupts the record: the losing session keeps reporting paths under its downloadDir, but the file was written to the other one. path then points at nothing, with no error to notice. Give each session its own Chrome — its own port (or transport: "pipe") and its own user-data-dir — whenever downloadDir matters.

Multiple instances via the CLI

The same thing without a Node host — one process per Chrome, each on its own port:

public-browser --port 9333 --profile research --download-dir /q/research
public-browser --port 9334 --profile action   --download-dir /q/action

--profile <name> uses one of your real Chrome profiles. For a throwaway per-agent Chrome, point at a raw directory instead — it is created if missing:

public-browser --port 9335 --user-data-dir /var/agents/a3/chrome

--attach connects to an already-running Chrome on the configured port instead of launching one. SILBERCUE_CHROME_PORT and SILBERCUE_SCRIPT_PORT are the environment equivalents of --port and --script-port and are part of the stable public contract.

Identifiable automation (--no-stealth)

By default Public Browser masks navigator.webdriver (it reports undefined) and launches Chrome with --disable-blink-features=AutomationControlled. That is the right default for consumer automation, but the wrong one when your integration must be transparently identifiable as a bot — compliance-driven crawling, internal agent fleets, or sites whose terms require honest signalling.

Turn the masking off completely:

public-browser --no-stealth
# or
SILBERCUE_STEALTH=0 npx public-browser

await createSession({ stealth: false });

With stealth off, navigator.webdriver stays true and keeps its native getter (Object.getOwnPropertyDescriptor(Navigator.prototype, "webdriver").get still reports [native code]) — permanently, across navigations and tab switches, with no post-correction needed on your side. No masking script is injected at any point and the launch flag is omitted.

Downloads

Downloads land in a per-session temp directory that is removed on shutdown. Point them at a directory of your own — a quarantine dir, a shared volume — with --download-dir / PUBLIC_BROWSER_DOWNLOAD_DIR / downloadDir. A directory you supply is created if missing and never deleted by Public Browser.

With --download-hash (or downloadHash: true) every completed download also carries a sha256, so the download tool returns path, size and digest:

{"filename":"report.pdf","path":"/q/research/A1B2...","size":48213,"sizeKb":48,
 "url":"https://example.com/report.pdf","sha256":"9f86d081884c7d659a2f..."}

Filenames. Chrome writes downloads under their internal GUID, so the file on disk is called A1B2... and only the filename field carries the real name. That is fine when you read the JSON, and useless when something else has to walk the directory. --download-naming suggested (or downloadNaming: "suggested", PUBLIC_BROWSER_DOWNLOAD_NAMING=suggested) renames each finished file to the server-supplied name:

{"filename":"report.pdf","path":"/q/research/report.pdf","size":48213,"sizeKb":48,
 "url":"https://example.com/report.pdf","sha256":"9f86d081884c7d659a2f..."}

The name is sanitised before it touches the disk — basename only, no control characters, never hidden, length-capped — and a collision gets a -1, -2, ... suffix rather than overwriting an existing file. filename always reports the name the file actually has, so join(downloadDir, filename) equals path. If the rename fails, the GUID path and the raw server name are kept and reported; a download is never lost to a naming problem.

Timing. action: "status" waits up to 250 ms for a download to start before reporting that there is none, because Chrome fires downloadWillBegin a few milliseconds after the click that triggers it — without the window, the first call after a click misses a file that is already on its way. Adjust it per call with settle ({"action":"status","settle":0} for an instant check, 5000 for a slow server). Once a download has started, status waits for it to finish, bounded by timeout.

For polling loops use action: "list" — it returns the full session history immediately and never waits, for either a start or a completion.

Tool Overview

Why an MCP server and not a CLI?

Several browser-automation projects now ship a CLI and tell coding agents to call it from the shell; Microsoft's Playwright README recommends that route. A CLI adds no tool definitions to the context. The trade-off is that the model has to learn the command surface from --help output and error messages. In one practitioner's side-by-side of Chrome DevTools MCP and the agent-browser CLI, the MCP tool surface came out better and the models "do not seem deeply fluent with it yet" (Pasi Huuhka, 28 Jan 2026) — one comparison, not a study.

Public Browser keeps the MCP surface and keeps it under a fixed budget: its 25 tool definitions take about 4,990 tokens of context as delivered over the wire (characters / 4 of the tools/list response, npm run token-count; measured the same way with node scripts/token-count.mjs --cmd npx --args "-y @playwright/mcp@0.0.80", Playwright MCP 0.0.80 takes 4,626 and Chrome DevTools MCP 1.8.0 takes 6,290). Chars / 4 is a coarse proxy — your client's /context figure will differ, but it is the same proxy for all three servers. A test enforces the budget, so it cannot creep back up. Getting there cost nothing in the benchmark: the two acceptance runs of the shortened definitions solved 30/30 with 80 and 68 calls, against 84 and 86 for the previous wording (raw data in test-hardest/results-local/).

The other argument for a CLI — "MCP needs a round-trip per step" — is what run_plan is for: N steps in one call, executed server-side with variables, conditions and suspend/resume. In the same two acceptance runs that was 49% fewer tool calls than Playwright MCP (80 and 68 against 137 and 151, same pass rate); the headline's 41% comes from the published September runs with the previous wording (84 and 86 calls, see Benchmarks).

Coming from Browser MCP?

Browser MCP (@browsermcp/mcp) has had no release since 0.1.3 on 11 April 2025, and its extension bridge works on one tab. If you picked it for its four promises, here is what Public Browser does for each: Fast — talks to Chrome directly over CDP, no extension bridge, no cloud hop; Private — runs on your machine, telemetry is opt-in; Logged In — can drive your real, logged-in Chrome profile (see Chrome Profiles); Stealth — sends real CDP input events (not synthetic JS events) and masks navigator.webdriver; --no-stealth makes the automation identifiable when you want that. Install with one command (Quick Start). Tool names differ: browser_snapshot → view_page, browser_click → click, browser_type → type; view_page returns the refs that click and type take. Multi-tab works.

Benchmarks

Two data sets, measured on https://mcp-test.second-truth.com: a September 2026 blind re-run (current) and the April 2026 runs kept for history. Cross-suite and cross-model comparisons are not valid — compare rows only inside one data set. Raw run JSONs, the full method and the environment matrix are in test-hardest/README.md.

September 2026 (current)

2026-09-03, 35-test page with 30 scored (T5.3–T5.6 can only be started by the page's own runner; T4.7 grades a self-reported token count and was dropped for everyone). One fresh blind Claude Code session per run in print mode, driver model claude-opus-5, exactly one MCP server per session, built-in tools cut down to Write, identical prompt for every server. Two runs each for Public Browser 2.10.1, Playwright MCP 0.0.80 and Chrome DevTools MCP 1.8.0; one run for browser-use 0.12.5. All seven runs are pinned to the same page version (suite.html_sha256 = 81e4b7aa…bed2). Metrics are counted post-hoc from the session transcript via test-hardest/measure-tool-calls.sh — nothing is self-reported by the servers. Table below is the output of node test-hardest/blind-run.mjs compare; method, profiles and losses are documented in test-hardest/README.md.

What these two runs show: Public Browser needed 84 and 86 tool calls where Playwright MCP needed 137 and 151 and Chrome DevTools MCP needed 156 and 172, and it finished the page in 281s and 296s against 468s/493s and 547s/558s. Duration here is summary.duration_s, the page's own timer from the first test to the export click — server start, Chrome start and the first navigation are not in it; the full wall clock per run (harness.wall_clock_s) is 331s/346s against 501s/527s and 583s/597s. Pass rate is a tie with Playwright MCP (30/30 in both runs each) and one test ahead of Chrome DevTools MCP: its only miss is T5.2, a CDP-fingerprint check (navigator.webdriver is true under chrome-devtools-mcp) rather than a browser capability; on the other 29 tests it is a tie. browser-use-run6 has complete: false — two tests were never started, so its 24/30 is an incomplete run, not a clean loss.

Where Public Browser loses in these runs: Playwright MCP 0.0.80 returns the smaller responses (Ø 740 and 656 chars against 1,298 and 1,214; browser_snapshot Ø 1911 and 2269 against view_page at 2841 and 3398, whose P95 reaches 9,734), total response volume is a near tie slightly in Playwright's favour (101k/99k against 109k/104k), and per-call click latency is mixed rather than a win — by_tool.avg_ms for click is 90 ms in run 1 but 435 ms in run 2, against 251 and 260 ms for Chrome DevTools MCP. P95 response size over all calls goes against Public Browser in both runs and against both competitors (6077 and 6479 chars against Playwright's 3617 and 1587 and DevTools' 5676 and 5271), and so does the evaluate response (Ø 1913 and 966 chars against browser_evaluate at 1223 and 755 and evaluate_script at 1069 and 488). The April claim that view_page is 5.4x more compact than browser_snapshot does not hold against Playwright MCP 0.0.80.

What changed since April, and where the remaining lead comes from. The April 2026 numbers (35-test suite, Opus 4.6) had Public Browser's page views 5.4x smaller than Playwright MCP's. Playwright MCP 0.0.80 has since made its snapshot format considerably more compact, and in these September runs its single responses are the smaller ones (Ø 740 and 656 chars against 1,298 and 1,214). Summed over a whole run the context payload is about the same on both sides — 109,075 and 104,432 chars for Public Browser against 101,478 and 99,147 for Playwright MCP — so "smallest responses" is not an argument we still have, and we say so. What the model consumed over the whole session is a different measure, and there Public Browser is ahead: tokens.delta (input + output + cache writes + cache reads from the Claude Code transcript) is 6,288,593 and 6,518,119 against 8,783,693 and 9,592,264 — ~30% fewer — and cost_usd_list is $3.41 and $3.35 against $4.28 and $4.78 — ~25% less. Almost all of that volume is cache reads of the growing conversation (4.4M and 4.3M against 6.0M and 6.9M); fresh input is 170–304 tokens per run on every side, so the shared overhead that made session totals useless for comparison in April is not present in these blind runs. Cache reads are billed cheaper than fresh input, which is why the cost gap (25%) is smaller than the token gap (30%) — quote the cost figure when in doubt. What drives both is ~41% fewer tool calls (84 and 86 against 137 and 151) and ~40% less time to finish (281s and 296s against 468s and 493s on the page's own timer), at identical reliability: 30/30 in all four runs. That gap does not come from smaller snapshots — both servers took page snapshots about equally often (view_page 16 and 16 against browser_snapshot 17 and 14). The design hypothesis is that it comes from run_plan, which bundles several steps into one call and was used 22 and 29 times, and from the 1,601 characters (2,358 since v2.10.4, without the Cortex line) of handshake instructions Public Browser sends to tell the model how to work (it is the only server in this field that sends any). Neither was isolated in this benchmark, so treat both as plausible contributors, not as a proven cause. The advantage moved from "cheaper per look" to "cheaper per task", and the headline was changed accordingly. The chart at the top of this file plots these six ratios, wins and losses on one scale; it is regenerated from the run JSON by scripts/make-benchmark-chart-2026-09.py.

April 2026 (historical)

Measured on the same page against the 35-test version of the suite (April 2026) — 5 levels (Basics, Intermediate, Advanced, Hardest, Community Pain Points). Four of the 35 tests are runner-only and are excluded from every score, so all pass rates in this section are out of 31 scorable tests. An extended 42-test version exists locally and is not yet published; the numbers here are not measured against it. Driver model was Claude Opus 4.6 and competitor versions were not recorded. Each run is independent, values on the benchmark page are randomized per page-load, all runs started in a fresh Claude Code session out of /tmp (no project context bias), and all metrics measured post-hoc from the session JSONL via test-hardest/measure-tool-calls.sh — no self-reporting, no MCP-side instrumentation, just counting tool_use blocks and tool_result char lengths.

April 2026 data, 24- and 35-test suites, superseded by the September 2026 rerun above.

Head-to-Head (24-test suite, April 2026 — historical suite version)

All rows LLM-driven by Claude Opus 4.6 on the same test page, one recorded run each. Public Browser ran 2026-04-05, the other servers 2026-04-02. This is the older 24-test version of the suite — do not compare these rows against the 31-scorable-test numbers below.

In this one April 2026 run each (24-test suite, Opus 4.6), Public Browser needed 71 tool calls where Playwright MCP needed 138 — roughly half the roundtrips for the same 24 passes. The September 2026 re-run above is the current figure: 84 and 86 calls against 137 and 151. Raw data: test-hardest/benchmark-*.json (Public Browser row: benchmark-silbercuechrome_mcp-llm-2026-04-05.json, type: llm-driven).

Pass Rate + Duration (31 scorable tests, LLM-driven)

Every row is one recorded run; the run id is named so each number is traceable to a single run JSON in test-hardest/results/. No averaging across runs.

Servers with several recorded runs, so you can see the spread rather than only the row above: Playwright MCP ranges 29–30/31 across three runs (Runs 2–4), its best being 30/31 in 449s (Run 3); Chrome DevTools MCP ranges 27–29/31, its best 29/31 in 518s (Run 1). Run 2 is quoted for both because that is the run the tool-efficiency analysis below instruments end to end. On pass rate this field is effectively a tie — the durable difference is response size, and that holds across every Playwright run measured (avg 1,216–1,467 chars in Runs 2–4).

Tool-Efficiency (the fair metric)

We measure each tool call's response char length directly, group by tool name, estimate tokens via chars/4. Why this metric: in these April runs, session-level token deltas were dominated by LLM overhead (system prompt + CLAUDE.md + conversation history = ~80-90% of the budget) and only showed 5-15% differences between MCPs — untrustworthy for comparing browser servers. Tool-response size is the part the MCP server actually controls. (The September 2026 runs are different: they are blind sessions out of /tmp with no CLAUDE.md and 170–304 fresh input tokens per run, so their session totals are comparable and are quoted above.)

Public Browser Run 5 vs Playwright MCP Run 2 — the same two runs as the pass-rate table above.

Per-Tool Breakdown (where the difference comes from)

¹ recorded as read_page in the April 2026 runs; the tool was renamed to view_page afterwards.

The Ambient-Context trade-off

Ambient Context — Claude sees DOM changes for free, no extra view_page needed

Public Browser's click is 2.8x larger than Playwright's because every click response embeds the DOM diff (NEW/REMOVED/CHANGED lines). Playwright returns a bare confirmation, so the LLM typically follows up with a browser_snapshot or browser_evaluate to see what happened. Over a full benchmark run, Playwright MCP spends 47 browser_evaluate calls averaging 2,155 chars against Public Browser's 33 at 510 chars. Public Browser delivers the diff inline. Net result: PB's click+read_page+evaluate total is 120k chars vs Playwright MCP's 170k — 30% less response content overall.

April 2026, Opus 4.6: view_page was 5.4x more compact than Playwright MCP's browser_snapshot (superseded — against Playwright MCP 0.0.80 in September 2026 it is not)

Measured on the 35-test benchmark (2026-04-09): Public Browser's view_page averages 1,124 chars per call vs Playwright MCP's browser_snapshot at 6,084 chars. Same page, same test suite, same LLM driver. The a11y-tree compression + Ambient Context pipeline meant we only sent what the agent actually needed — smaller responses, less context pressure, cheaper runs. That was the April 2026 picture. Against Playwright MCP 0.0.80 it no longer holds — that release made the snapshot format much more compact, and in the September runs browser_snapshot averages 1,911 and 2,269 chars against view_page at 2,841 and 3,398; see September 2026 (current) above.

See test-hardest/README.md for the full protocol, per-test breakdown, and raw JSON runs with tool_efficiency blocks.

Cortex — Self-Learning Pattern Engine

Public Browser includes a lightweight learning layer called Cortex. It observes which tool sequences work on different page types and feeds that knowledge back as hints to the LLM agent. No ML model, no training step — just deterministic pattern recording and Markov-chain predictions.

How it works

1. Page Classification — Every page is classified by its accessibility tree into one of 16 functional types: login, signup, mfa, search_form, search_results, data_table, form_simple, form_wizard, article, navigation, dashboard, settings, media, checkout, profile, error. The classifier is rule-based (ARIA roles, landmarks, keyword signals) — no domains or URLs are involved.

2. Pattern Recording — Successful tool-call sequences (e.g. navigate → view_page → fill_form → click on a login page) are recorded into a local append-only Merkle log (~/.public-browser/patterns.jsonl). Only page type, tool names, a content hash, and a timestamp are stored — no URLs, no page content, no PII.

3. Markov Predictions — Recorded patterns are ingested into a first-order Markov table that models P(next_tool | last_tool, page_type). When the agent lands on a page, the Cortex returns the most likely next tools with probabilities. Stale entries decay automatically (0.95/week, removed after 30 days).

4. Community Markov Table — A hand-curated transition table (community-markov.json) ships with every installation. It contains baseline probabilities for common page types so that new installations benefit from community knowledge immediately, without needing local history. The table is SHA-256 verified at load time and merged with local patterns (local data takes precedence).

Privacy by design

The Cortex stores and transmits only structural metadata — page types (not domains), tool names (not arguments), and content hashes (not content). A login pattern reveals nothing about which login page was visited. The telemetry payload is built via explicit field allowlist (no spread operator), preventing accidental leakage of future fields.

Opt-in telemetry

Telemetry is disabled by default. To contribute your anonymised patterns back to the community table, set PUBLIC_BROWSER_TELEMETRY=1. Uploads go via HTTPS only; non-HTTPS endpoints are rejected. Each pattern is rate-limited to prevent duplicate uploads.

Local friction log (developer opt-in)

For development of Public Browser itself there is a second, fully local opt-in: SILBERCUE_CHROME_FRICTION_LOG=1 makes the server count tool calls, tool errors and detected fallback spirals per run in ~/.silbercue-chrome/friction-queue.json. It records counters, timestamps and the working directory — never page content, URLs or user input — and nothing ever leaves the machine. Without the variable the code path is not entered at all: no file, no counters, no hints.

Architecture

Public Browser (Node.js MCP server, public-browser)
+-- @modelcontextprotocol/sdk (stdio transport)
+-- CDP Client
|   +-- WebSocket transport (existing Chrome on :9222)
|   +-- Pipe transport (auto-launched Chrome with --remote-debugging-pipe)
+-- Auto-Launch: Chrome + optimal flags, visible by default
+-- A11y-tree cache + Selector cache
+-- Session Manager (OOPIF support for iframes and Shadow DOM)
+-- Tab State Cache (URL/title/ready across tabs)
+-- Cortex (self-learning pattern engine)
|   +-- Page Classifier (16 page types from a11y-tree)
|   +-- Pattern Recorder + Merkle Log (local persistence)
|   +-- Markov Table (transition predictions)
|   +-- Community Table (shipped baseline, SHA-256 verified)
|   +-- Hint Matcher (delivers predictions to tool responses)
|   +-- Telemetry Upload (opt-in, HTTPS, rate-limited)
+-- Script API (Python, source install from ./python)
|   +-- Shared Core via HTTP (:9223) — same tool handlers as MCP
|   +-- Escape Hatch via WebSocket (:9222) — direct CDP for power users
+-- 25 tools
    Reading - Interaction - Navigation - Scripting - Observation

Connection priority:

1. Auto-Launch (default, zero-config) — starts Chrome as a child process via --remote-debugging-pipe, visible as a window, with all flags set for reliable screenshots and keyboard focus.

2. WebSocket (optional) — if you already run Chrome with --remote-debugging-port=9222, Public Browser connects to that instead. Use this to control your own browser with its extensions and login sessions.

Requirements

• Node.js >= 18

• Google Chrome, Chromium, or any Chromium-based browser (auto-detected on macOS/Linux/Windows; override with CHROME_PATH)

Environment Variables

Invalid values fail loudly: an unparseable port or naming mode aborts startup with a named error instead of silently falling back to 9222. Sessions created through the Node library ignore every variable in this table except CHROME_PATH and the telemetry pair — see Node Library API.

License

MIT licensed — see LICENSE. Use it however you want, commercially or otherwise.

Contributing

Issues and pull requests welcome at github.com/Silbercue/public-browser.

Privacy

Public Browser runs entirely on your machine. All browser automation happens locally via CDP. The Cortex learning layer stores only structural metadata locally (page types, tool names, content hashes — no URLs, no domains, no page content, no PII). Telemetry is off by default. If you opt in via PUBLIC_BROWSER_TELEMETRY=1, only the same structural metadata is uploaded via HTTPS — the payload is built from an explicit field allowlist to prevent accidental leakage.

Related

Building for iOS too? SilbercueSwift is the same idea for the iOS Simulator.

Links

• GitHub Repository

• npm Package

• Benchmark Test Site