Which MCP browser server uses the fewest tokens?

Owl Browser MCP, in this benchmark. Across six tasks and 60 runs per stack it spent a median 38,590 tokens per task, against 71,437 for Chrome DevTools MCP and 73,420 for Playwright MCP, roughly 47% cheaper than either. It also used fewer tool calls: 2.0 per task against 3.0 and 4.0. All three stacks answered 100% of tasks correctly, so correctness does not separate them and cost does.

Why is Owl Browser cheaper per task for an AI agent?

Three reasons. Pages are returned as a compact structural text view called OwlMark with handle tokens for every interactive element, instead of screenshots that cost 20 to 100 times more tokens. The default agent tool profile advertises 20 tools rather than 80, and an MCP client re-sends its whole tool catalogue every turn, so each unnecessary tool is taxed on every turn. One-shot tools like browser_go and browser_search create a context, act, extract and close in a single call, so simple reads do not pay for a full session lifecycle.

Can I run a browser automation task without an AI agent CLI?

Yes. Owl Browser has a built-in natural language loop, browser_nla, that observes the page, takes one action, and observes the result, all inside the browser. There is no external agent harness and no tool catalogue re-sent per turn. On the same six benchmark tasks it answered 18 of 18 correctly at a median 4,844 tokens, 2 steps and 6.7 seconds per task. Use the MCP server instead when the browser is one tool among many and an agent must reason across all of them.

Are the Owl Browser benchmark results reproducible?

Yes. The harness for both suites is open source at github.com/Olib-AI/owl-detection-report, and the agent-efficiency harness drives any MCP browser server, not just Owl Browser. The model is pinned to a local qwen3.6-27b at Q4_K_M quantization so the same weights and context length can be loaded by anyone. The raw result JSON behind every figure is published alongside the page.

Does the benchmark report a dollar cost per task?

No, deliberately. The model runs locally, so there is no price to measure, and any dollar figure an agent CLI prints is a hosted pricing table applied to local token counts. Tokens, tool calls and wall clock are reported instead because those are the quantities actually measured.

BENCHMARK

Same Answers. Fewer Tokens.

Agentic, detection, and performance benchmarks - what an AI agent spends to finish the job on Owl Browser MCP versus the alternatives, how it scores against CreepJS, and how it keeps pace with bare-metal automation tools under load.

What Does the Same Answer Cost?

The same agent CLIs, the same tasks, the same machine - only the browser MCP server changes. Every stack gets the answer right, so correctness is the entry requirement, not the score. What separates them is what the agent had to spend to get there.

Owl Browser MCP

v1.3.2 · 20 tools

48.1ktokens per task

100% of runs answered correctly

Cheapest per task

What it cost to get thereall runs

Tool calls per task

2.3

client 2.3 · transport 2.3

Tokens per task

48.1k

Page payload read back

2.2k tokens

Images sent to the model

0 blocks

0 B

Time to first action

6.4s

Wall clock per task

17.2s

Cheap only counts when the answer is right. A stack that quits early looks fast and thrifty on every line above.

Chrome DevTools MCP

v1.6.0 · 29 tools

75.8ktokens per task

100% of runs answered correctly

What it cost to get thereall runs

Tool calls per task

3.3

client 3.3 · transport 3.3

Tokens per task

75.8k

Page payload read back

5.2k tokens

Images sent to the model

0 blocks

0 B

Time to first action

6.8s

Wall clock per task

17.6s

Cheap only counts when the answer is right. A stack that quits early looks fast and thrifty on every line above.

Playwright MCP

v0.0.79 · 24 tools

82.7ktokens per task

100% of runs answered correctly

What it cost to get thereall runs

Tool calls per task

4.6

client 4.6 · transport 4.6

Tokens per task

82.7k

Page payload read back

4.0k tokens

Images sent to the model

0 blocks

0 B

Time to first action

6.3s

Wall clock per task

19.9s

Cheap only counts when the answer is right. A stack that quits early looks fast and thrifty on every line above.

Task by Task

Correct answers first, then what the run cost (all runs). Averaged across every agent CLI until you filter above.

Task	Owl Browser MCP	Playwright MCP	Chrome DevTools MCP
Find the cheapest book in the Mystery category `list-extraction`	100%correct 4.5calls 92.3ktokens	100%correct 6calls 141.4ktokens	100%correct 4.5calls 119.7ktokens
Read the last item on the last page of a 1000-item catalogue `deep-pagination`	100%correct 1.5calls 35.7ktokens	100%correct 4calls 94.2ktokens	100%correct 2calls 49.4ktokens
Read a value out of an HTML product table `table-read`	100%correct 1calls 24.4ktokens	100%correct 2calls 38.5ktokens	100%correct 2calls 43.3ktokens
Open a detail page from a listing and read its stock count `detail-drilldown`	100%correct 2calls 43.4ktokens	100%correct 4calls 67.0ktokens	100%correct 4calls 99.4ktokens
Fill and submit a six-field form, report the confirmation code `form-fill`	100%correct 4calls 68.7ktokens	100%correct 9.5calls 117.4ktokens	100%correct 5.5calls 100.1ktokens
Negative control: report that a figure is not available `negative-control`	100%correct 1calls 24.2ktokens	100%correct 2calls 38.0ktokens	100%correct 2calls 42.8ktokens

Scroll the table sideways to compare every stack

Methodology: Driven by qwen/qwen3.6-27b. 5 runs per task, per stack, per agent. Agent CLIs under test: Claude Code, Codex CLI. Identical prompts and tasks on one machine, only the MCP server swapped. Raw per-run answers are published, so the grading can be audited. Owl is handicapped here and it is not adjusted for: it runs as a server the agent reaches over HTTP, so every call pays network latency, while the other two run in-process on the same machine. That is also the trade - one Owl instance serves many agents concurrently, where a local MCP server is one browser for one agent. Figures cover every run in the set. Dollar cost is not reported - the model runs locally, so any figure the agent CLIs print is fabricated from a hosted pricing table. Owl Browser v1.3.2. Last run: August 7, 2026. The harness, the exact prompts and the grading rules are in the benchmark repository; every raw per-run answer, with its expected value and the rule that graded it, is in the report data this page loads.

Fingerprint controls, tested in public

We ran the same fingerprint collector and performance procedure for each automation stack, then published the versions, screenshots, and raw output. Detection on other sites can differ from these results. This suite includes CreepJS, then pushed under load to show stealth costs nothing in speed.

Playwright

baseline

Runs on Linux - cannot spoof platform

Headless

100%

Like Headless

94%

Platform

Linux

GPU

ANGLE (Google, Vulkan 1.3.0 (S...

Fonts

1/51

Devices

User Agent

HeadlessChrome

Confidence

moderate

creepjs - fingerprint

Puppeteer

baseline

Runs on Linux - cannot spoof platform

Headless

100%

Like Headless

50%

Platform

Linux

GPU

ANGLE (Google, Vulkan 1.3.0 (S...

Fonts

1/51

Devices

User Agent

HeadlessChrome

Confidence

moderate

creepjs - fingerprint

Owl Browser

spoofed

Unique fingerprint per context - each session gets its own identity

Headless

Like Headless

31%

Platform

Windows

GPU

ANGLE (Intel, Intel(R) UHD Gra...

Fonts

5/51

Devices

User Agent

Mozilla/5.0 (Windows NT 10.0...

Confidence

high

creepjs - fingerprint

Fingerprint Uniqueness

Playwright and Puppeteer expose identical device fingerprints - they leak the real machine. Owl Browser generates unique fingerprints per profile at the C++ source level.

Category	Playwright	Puppeteer	Owl (Windows)
Canvas	`f5ae5edb`	`f5ae5edb`	`fdcdbf14`
WebGL	`208a9379`	`208a9379`	`b0017f68`
Audio	`9048f6b9`	`9048f6b9`	`ae05bdc6`
Fonts	`f936b9a1`	`f936b9a1`	`023b83e2`
DOMRect	`c4c4b540`	`c4c4b540`	`8e02378c`
SVGRect	`cbd2ed1a`	`cbd2ed1a`	`cca10e0b`
Screen	`5ccbceb6`	`b0313e03`	`39974c4d`

Puppeteer

522.8ms

p95: 574.9

Methodology: Each browser performs 1000 sequential iterations of: create/launch → navigate to target URL (wait for domcontentloaded) → capture screenshot → close. All times in milliseconds. Playwright and Puppeteer run inside the benchmark container (local process). Owl Browser runs in a separate container and is accessed via REST API over localhost, adding ~5-10ms of network overhead per API call - Owl's real times are faster than shown. Sequential execution, same machine. Owl times include the full cost of stealth - fingerprint generation, per-context isolation, and REST API overhead to an isolated container.

Concurrency Scaling

All sessions launch simultaneously. Each performs a full cycle - create context, navigate, screenshot, and close.

OwlPlaywrightPuppeteer

1 session:Owl 1/1Playwright 1/1Puppeteer 1/1

10 sessions:Owl 10/10Playwright 10/10Puppeteer 10/10

25 sessions:Owl 25/25Playwright 25/25Puppeteer 25/25

50 sessions:Owl 50/50Playwright 50/50Puppeteer 50/50

100 sessions:Owl 100/100Playwright 99/100 (1 failed)Puppeteer 100/100

Methodology: At each concurrency level (1, 10, 25, 50, 100 sessions), all sessions launch simultaneously and each performs: create → navigate (domcontentloaded) → screenshot → close. Playwright and Puppeteer each launch a separate browser process per session. Owl Browser creates lightweight contexts within a single running engine. Total time = wall clock from first launch to last completion.

Run the same tests in your environment

Our published run shows the measured fingerprint and performance results above. It does not guarantee that every detector, site, account, or proxy will accept a session. Clone the harness, reproduce the run, and test the sites that matter to your workflow.

Get Started

Tested by the Olib engineering team·Owl Browser v1.3.0·Playwright (chromium)·Puppeteer (Chrome/145.0.7632.6)·Last run: July 7, 2026

Detection powered by CreepJS · View report generator source code

Methodology maintained by Akram H. S., Founder and CTO · Architecture and limitations