_{More badges · featured in 37 curated lists}

ClawBench: Can AI Agents Complete Everyday Online Tasks?

ClawBench is an open-source benchmark that evaluates AI browser agents on everyday online tasks — booking travel, ordering food, applying for jobs, managing email — across live websites. V1 lives in test-cases/v1/, V2 in test-cases/v2/. It measures end-to-end task success with a 5-layer recording pipeline and an agentic evaluator that compares each run against human references. Top score to date: 33.3%.

We asked frontier AI agents to do what people do every day --
order food, book travel, apply for jobs, write reviews, manage projects.
Even the best agent only completes about 1 in 3.

---

V1: 152 everyday tasks · 143 live sites | V2: 129 tasks · 63 live sites | 281 total across 163 live websites · 15 life categories

_{The paper reports 153 (V1) and 130 (V2); two ASPCA tasks were removed after publication, so the shipping corpus is 152 and 129.}

_{Built by NAIL Group · Sister project: HarnessBench — fixes the base model, varies the harness · Runs on any Chrome.}

中文

What are you looking for?

🏆 See scores Live leaderboard Pick a corpus (v1 / v2)	🚀 Run it on your model Quick start ↓ pip install clawbench-eval	📊 Browse 281 tasks Task explorer Search · filter · category	📄 Read the paper arXiv:2604.08523 Methodology · evaluator · results
🎬 Re-grade old runs V1 · V2 raw traces 5 layers per (task × model)	📦 Download the data hf download NAIL-Group/ClawBench Tasks · rubrics · metadata	🌱 Add a task / model How to contribute JSON spec + rubric	❓ Have a question FAQ · Open an issue Or ask on the HF dataset page

Quick start

git clone https://github.com/TIGER-AI-Lab/ClawBench.git &&cd ClawBench && ./run.sh

_{Clone → configure → run. Root uv package. Docker-isolated harnesses.}

Driving a coding agent instead? Point it at AGENTS.md and prompt away.

Install

uv tool install clawbench-eval

pipx install clawbench-eval and python -m pip install clawbench-eval work too. The installed commands are clawbench, clawbench-run, clawbench-batch, clawbench-rescore, clawbench-reproduce, and clawbench-harbor-adapt.

For more granular control — modifying the driver, the bundled test cases, or the container build — clone the repo and use the root uv package entrypoint instead:

git clone https://github.com/TIGER-AI-Lab/ClawBench.git &&cd ClawBench && ./run.sh

Prerequisites:Python 3.11+, uv, and a container engine — DockerorPodman. ClawBench auto-detects whichever is installed; force one with export CONTAINER_ENGINE=docker or export CONTAINER_ENGINE=podman.

Install Docker or Podman (macOS / Linux / Windows)

macOS

# Option A — Docker Desktop (easiest, includes GUI)
brew install --cask docker
open -a Docker # launch and wait for the whale icon to settle# Option B — Podman (rootless, no daemon, CLI only)
brew install podman
podman machine init # one-time: downloads the Linux VM image
podman machine start # must be running before any podman command

macOS Podman needs a VM.brew install podman alone is not enough — Podman on macOS runs containers inside a small Linux VM, so you must podman machine init && podman machine start once after install or podman info will fail with Cannot connect to Podman.

Linux (Ubuntu / Debian)

# Option A — Podman (rootless by default, recommended)
sudo apt update && sudo apt install -y podman
# Option B — Docker
sudo apt install -y docker.io
sudo usermod -aG docker $USER# log out / back in so your shell picks up the group

Rootful Docker ownership note: with classic sudo-docker, files extracted from containers land owned by root on the host. ClawBench's driver detects this after each run and chowns test-output/ back to your user automatically — but if you run other container tooling alongside, rootless Podman (or rootless Docker) avoids the issue entirely.

Windows

# Option A — Docker Desktop (WSL2 backend)
winget install Docker.DockerDesktop
# then launch Docker Desktop from the Start menu and wait for it to be ready# Option B — Podman
winget install RedHat.Podman
podman machine init
podman machine start

Run the uv run … commands below from PowerShell, WSL2, or Git Bash. Like macOS, Windows Podman requires podman machine init && podman machine start before its first use.

1. Configure models

One-time setup. If you installed from PyPI, run clawbench from the directory where you want results and editable config to live — on first launch it creates local templates under models/:

clawbench
$EDITOR models/models.yaml

From a source checkout:

cp models/models.example.yaml models/models.yaml
$EDITOR models/models.yaml

Scoring needs a judge. Add an API key for deepseek-v4-pro — the judge used for every published leaderboard row — before running any judged batch:

deepseek-v4-pro:
api_key: "sk-..."base_url: <api_base_url>api_type: openai-completions

PurelyMail credentials for disposable run emails are provided in the committed .env. You only need to edit .env to use your own PurelyMail account or to enable optional HuggingFace upload.

Note

First run builds a container image (Chromium + ffmpeg + noVNC + the selected agent harness dependencies). You'll see a live progress spinner with the current build step. Subsequent runs reuse the cached layers and finish in seconds.

2. Run your first task

Tip

Recommended → interactive TUI, with guided model + test case selection:

clawbench # PyPI install
uv run clawbench # source checkout

Needs an interactive terminal. For pipes / CI / non-TTY, use clawbench-run or clawbench-batch directly.

One task against one model:

uv run clawbench-run test-cases/v1/001-daily-life-food-uber-eats claude-sonnet-4-6

Once the container starts, the script prints a noVNC URL (e.g. http://localhost:6080/vnc.html) — open it to watch the agent operate in real time. If port 6080 is taken, an alternative is chosen automatically. Results land in ./test-output/<model>/<harness>-<case>-<model>-<timestamp>/ with the full five-layer recording.

A whole corpus:

clawbench-batch --models your-model --cases-suite v2 --all-cases

your-model is a key you configured in step 1; --cases-suite v2 runs the full V2 corpus (swap in v1-lite for the 20-task subset). Add --max-concurrent N to run tasks in parallel (default 2 locally, 1 with Browserbase) and --harness <name> to pick an agent. Each task is intercepted and scored by the deepseek-v4-pro judge from step 1 — pass --no-judge to skip scoring. A batch-summary.json plus per-run recordings land under ./test-output/.

By hand, to produce a human reference run:

uv run clawbench-run test-cases/v1/001-daily-life-food-uber-eats --human

Open the noVNC URL, complete the task yourself, then close the tab. You can also leave the session open and let an external browser agent drive it while ClawBench records and intercepts.

3. Pick a harness

The harness is the agent scaffold that drives the browser; the model is a separate axis. Default is openclaw. Select one with --harness <name> on clawbench-run or clawbench-batch.

Harness	--harness	How it drives the browser	Use it when
OpenClaw	openclaw(default)	Playwright MCP bridge	You want the reference configuration used by V1 results
Hermes Agent	hermes	Native browser tools over CDP	You want the configuration behind most V2 leaderboard rows
opencode	opencode	Playwright MCP bridge	Comparing coding-agent scaffolds
Claude Code	claude-code	Playwright MCP bridge	Comparing coding-agent scaffolds
Claude Code + Claude in Chrome	claude-code-chrome-extension	Chrome extension via a local bridge (Microsoft Edge)	Testing the extension stack; any LiteLLM-routed provider works
OpenAI Codex CLI	codex	Playwright MCP bridge	Comparing coding-agent scaffolds
claw-code	claw-code	Playwright MCP bridge	Comparing coding-agent scaffolds
browser-use	browser-use	Native browser framework, routed via LiteLLM	Comparing a purpose-built web agent
Pi	pi	Pinned pi-browser-harness tools over CDP	Read-only tool allowlist, no shell
—	random-click	Random clicks, no model	Establishing a floor baseline
—	null	Does nothing	Measuring harness/recording overhead

Full registry: src/clawbench/runtime/harnesses/harnesses.yaml.

4. Other browser runtimes and frameworks

I want to…	Where
Use a managed remote browser instead of a local container	docs/browser-runtimes.md — Browserbase setup, options, recording URLs
Run V2 through the Harbor framework (and run it fast)	docs/harbor.md — conversion, judge wiring, concurrency, troubleshooting
See every CLI command and flag	docs/cli.md

Develop from source — clone + ./run.sh for contributors

Prefer the repo checkout if you want to modify the driver, the bundled V1/V2 test cases, or the container build itself.

git clone https://github.com/TIGER-AI-Lab/ClawBench.git &&cd ClawBench
cp models/models.example.yaml models/models.yaml # edit: add your model API keys# .env is already provided for PurelyMail; edit only for your own creds or HF upload
./run.sh # interactive TUI
uv run clawbench-run \
test-cases/v1/001-daily-life-food-uber-eats claude-sonnet-4-6 # single run
uv run clawbench-run \
test-cases/v1/001-daily-life-food-uber-eats --human # human mode

This path gives you live-reload on src/, src/clawbench/runtime/chrome-extension/, and all suites under test-cases/ — useful when iterating on the harness itself.

How it works

 You pick a task ClawBench spins up Agent drives the Interceptor captures
from V1 or V2 an isolated Docker browser: navigates, every action across
everyday scenarios container + Chromium fills forms, clicks all 5 layers of data
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ "Book a pet │ ──► │ Container │ ──► │ AI Agent │ ──► │ 5 layers │
│ sitter on │ │ + Chromium │ │ browses the │ │ intercepted │
│ Rover" │ │ + Agent │ │ live site │ │ & recorded │
└──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘

Live WebsitesIsolated ContainersRequest InterceptorFive-Layer Recording

Container internals

┌─────────────────────────────────────────────────┐
│ Container (Docker / Podman) │
│ │
│ ┌──────────┐ CDP Fetch/Runtime/Page events │
│ │ Chromium ├─────────────────────────────┐ │
│ │ :9222 CDP│ │ │
│ └──────────┘ │ │
│ │ │
│ ┌──────────┐ ┌────────────────▼─┐ │
│ │ Xvfb │◄──ffmpeg──►│ FastAPI Server │ │
│ │ :99 │ x11grab │ :7878 │ │
│ └──────────┘ └──────────────────┘ │
│ │ │
│ ┌───────▼─────────┐ │
│ │ /data │ │
│ │ actions.jsonl │ │
│ │ requests.jsonl │ │
│ │ screenshots/ │ │
│ │ recording.mp4 │ │
│ └─────────────────┘ │
└─────────────────────────────────────────────────┘

Datasets

ClawBench ships three Hugging Face datasets — task definitions plus full execution traces for V1 and V2. All open, downloadable in one command.

Dataset	What's in it	Get it
NAIL-Group/ClawBench(mirrored at TIGER-Lab/ClawBench)	Task definitions, rubrics, and metadata for V1 and V2 — what to attempt and how it's judged.	hf download --repo-type dataset NAIL-Group/ClawBench
NAIL-Group/ClawBenchV1Trace	One directory per V1 model run: recording.mp4, requests.jsonl, actions.jsonl, agent-messages.jsonl, interception.json, run-meta.json.	hf download --repo-type dataset NAIL-Group/ClawBenchV1Trace
TIGER-Lab/ClawBenchV2Trace	Same 5-layer bundle for V2 runs. Rolling — new models added as they're evaluated.	hf download --repo-type dataset TIGER-Lab/ClawBenchV2Trace

The trace datasets are large; use hf download --include "<pattern>" to pull a single model or a single task.

🏆 Live leaderboard:claw-bench.com/leaderboard (V2 default, two-stage scoring — interception + LLM judge). Full scoring formula in eval/scoring.md. Add your run: PR to leaderboard/results.csv.

News

• [2026.08.16] — Released RewardHarness, our self-evolving agentic reward framework: 47.4% on EditReward-Bench from just 100 preference demos, with no reward-model training. Details →
• [2026.08.03] — Added Browserbase as a remote browser runtime. Details →
• [2026.07.30] — v0.8.0: Gemini-as-judge, random-click baseline harness, EdgeBench/SForge adapter, remote-browser CDP support. Details →
• [2026.07.25] — 🏆 Our paper has been accepted by COLM 2026 WAB.
• [2026.06.22] — v0.7.0: Harbor-adapter task export; action recording moved into the CDP server. Details →

_{Earlier updates: docs/news.md · full change history: CHANGELOG.md}

Results

ClawBench leaderboard · by corpus × harness · live at claw-bench.com

V2 (Hermes) · 8 models · ds-v4-pro judge, lenient + strict

Rank	Model	Harness	Intercepted	Reward (lenient)	Reward (strict)	Pass / Total
1	claude-opus-4-7	hermes	54.6%	44.6%	24.6%	58 / 130
2	gpt-5.5	hermes	45.4%	35.4%	18.5%	46 / 130
3	glm-5.1	hermes	48.5%	34.6%	17.7%	45 / 130
4	deepseek-v4-pro	hermes	43.9%	33.9%	12.3%	44 / 130
5	openrouter-owl-alpha	hermes	14.6%	0.0%	0.0%	0 / 130
6	z-ai/glm-4.5-air:free	hermes	4.6%	2.3%	0.8%	3 / 130
7	deepseek-v4-flash:free	hermes	3.1%	2.3%	0.0%	3 / 129
8	minimax-m2.5:free	hermes	2.3%	1.5%	0.0%	2 / 130

Intercepted = final HTTP request matched the task's URL/method (Stage 1, deterministic). Reward (lenient) = additionally judged by deepseek/deepseek-v4-pro to fulfill the instruction under the "no contradiction → match" rubric (Stage 2). Reward (strict) = same judge, strict rubric ("ambiguous → mismatch"). Ranked by Intercepted; Reward as tiebreak. Totals reflect the corpus size at run time.

V2 (OpenClaw) · 1 model

Rank	Model	Harness	Intercepted	Reward (lenient)	Reward (strict)	Pass / Total
1	glm-5.1	openclaw	0.0%	0.0%	0.0%	0 / 130

V1 (Hermes) · 6 frontier models, original paper rubric

Rank	Model	Harness	Pass Rate	Pass / Total
1	claude-opus-4-6	hermes	61.4%	94 / 153
2	claude-sonnet-4-6	hermes	56.9%	87 / 153
3	claude-haiku-4-5-20251001	hermes	30.1%	46 / 153
4	gpt-5.4-2026-03-05	hermes	25.5%	39 / 153
5	gpt-5.4-mini-2026-03-17	hermes	24.8%	38 / 153
6	kimi-k2.5	hermes	17.6%	27 / 153

V1 Pass Rate is from the original paper rubric (Claude Code agentic-eval subagent comparing each run against human reference trajectories under eval/agentic_eval.md). The two-stage Reward (interception + deepseek/deepseek-v4-pro lenient judge) for V1 will appear here once V1 trace bundles are re-judged.

V1 per-category breakdown (Sonnet 4.6 vs 6-model comparison)

Rank	Model	Overall	Daily	Finance	Work	Dev	Academic	Travel	Social	Pets
1	Claude Sonnet 4.6	33.3	44.2	50.0	19.0	11.1	50.0	23.1	38.9	18.2
2	GLM-5	24.2	30.8	16.7	38.1	16.7	28.6	0.0	16.7	18.2
3	Gemini 3 Flash	19.0	15.4	33.3	23.8	22.2	28.6	30.8	11.1	0.0
4	Claude Haiku 4.5	18.3	15.4	22.2	19.0	27.8	21.4	7.7	16.7	18.2
5	GPT-5.4	6.5	9.6	0.0	0.0	11.1	7.1	7.7	0.0	9.1
6	Gemini 3.1 Flash Lite	3.3	1.9	0.0	0.0	5.6	14.3	0.0	0.0	9.1

Task categories · V1: 15 categories, 152 tasks

Category	Tasks	Example Platforms
Daily Life	21	Uber Eats, DoorDash, Instacart, Zillow, Craigslist
Entertainment & Hobbies	15	Ticketmaster, AMC Theatres, Topgolf, Crunchyroll
Creation & Initialization	13	Squarespace, Wix, Webflow, Ghost, Substack
Rating & Voting	10	Trustpilot, G2, Goodreads, RateMyProfessors
Travel	9	Booking.com, Expedia, Airbnb, TripAdvisor
Education & Learning	9	Coursera, Udemy, Khan Academy, Duolingo
Office & Secretary	9	Google Calendar, Slack, Notion, Trello
Beauty & Personal Care	9	Sephora, Ulta, Glossier
Job Search & HR	8	LinkedIn, Greenhouse, Lever, Workday
Pet & Animal Care	7	Chewy, Petco, Rover
Personal Management	6	Mint, YNAB, Todoist
Shopping & Commerce	6	Amazon, eBay, Etsy, Target
Nonprofit & Charity	6	GoFundMe, DonorsChoose
Academia & Research	5	Google Scholar, Semantic Scholar, OpenReview
Finance & Investment	4	Robinhood, Fidelity, Coinbase
Others	15	Automation, Dev & Tech, Government, Home Services, Automotive

_{Codex and Claude Code runs on V2, and the V1 OpenClaw aggregate, are still in flight — they land on the live leaderboard first.}

Reproduce the leaderboard

Our scores are stable: two independent runs of the same model under the same judge (deepseek/deepseek-v4-pro, lenient rubric) reproduce Intercepted and Reward within ±2 pp on the V2 corpus.

There are two ways to verify this on your own machine.

Path A — Re-run the agent, then score

Confirms the full pipeline (your agent + our judge) lines up with our leaderboard row.

clawbench-batch --models deepseek/deepseek-v4-flash --cases-suite v2 \
--all-cases --harness hermes --no-judge --output-dir ./my-run
clawbench-rescore ./my-run --judge-model deepseek-v4-pro --rubric both

Path B — Skip the run, re-judge our published traces

Confirms just the judge matches ours (cheap, no agent compute, useful for sanity-checking your judge config).

hf download --repo-type dataset TIGER-Lab/ClawBenchV2Trace \
--include "batch-aligned-*/deepseek-v4-flash-free/**" --local-dir ./reproduce
clawbench-rescore ./reproduce --judge-model deepseek-v4-pro --rubric both

One-shot equivalent of Path B for any model in the leaderboard:

clawbench-reproduce --model deepseek-v4-flash --tolerance 2.0

Pass criterion

For deepseek-v4-flash:free × hermes × v2, the published row is Intercepted 3.1% / Reward-lenient 2.3% / Reward-strict 0.0% (3 / 129). Path A or B counts as reproduced when all three metrics land within ±2 pp. Larger gaps usually mean a different judge model, a different rubric prompt, or a harness configuration drift — diff your eval_results/<batch>/summary.json against the published row to localize the cause.

ClawBench-Lite

New here? Run this first.test-cases/v1-lite/ is a 20-task curated subset of V1, selected for household-name sites, real-world relevance, difficulty, and category diversity. It matches the 20-tasks-per-source convention of browser-use/benchmark and gives you a credible signal at a fraction of the full-benchmark cost.

Tier distribution: flagship 9 / core 8 / wildcard 3 — spanning daily life (OpenTable, DoorDash, Instacart, TaskRabbit), entertainment (Eventbrite, Goodreads, Fandango), creation (Asana, Mailchimp, Squarespace), travel (Airbnb), education (LeetCode), dev-tech (GitHub), academia (Overleaf), personal management (1Password), and more. All Lite tasks are judged by eval/agentic_eval.md regardless of url_pattern shape.

Run it with --cases-suite v1-lite, or inspect the link-backed task files in test-cases/v1-lite/.

Example walkthrough

Curious what one task actually looks like, start to finish? Here's task 001 end to end. Each run also produces a full MP4 session recording — see the project page for V1 task recordings.

The task — from test-cases/v1/001-daily-life-food-uber-eats/task.json:

{
"instruction": "On Uber Eats, order delivery: one Pad Thai, deliver to home address, note \"no peanuts\"",
"time_limit": 30,
"eval_schema": {
"url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
"method": "POST"
}
}

The agent gets this instruction verbatim, plus read-only access to /my-info/alex_green_personal_info.json (the dummy user's name, home address, phone, date of birth) and a disposable email account for any sign-in prompt. It has 30 minutes to reach a POST request — any longer and the container is killed.

What the agent does (the happy path):

1. Navigates to ubereats.com
2. Reads the dummy user's home address from /my-info/alex_green_personal_info.json and enters it in the delivery-address box
3. Searches for "Pad Thai" in the food search
4. Picks a restaurant that has Pad Thai available for delivery to that address
5. Opens the item detail page, finds the customization or special-instructions field, enters "no peanuts"
6. Adds one to cart, opens the cart, and handles any sign-in prompt using the disposable email credentials
7. Reaches checkout, taps Place Order

What the interceptor catches — that final Place Order tap fires a POST request. ClawBench's request interceptor sits in front of the browser and captures the outbound request before it reaches Uber Eats's servers, so the dummy user is never actually charged. At the exact moment of interception, all five recording layers (MP4 video, PNG screenshots, HTTP traffic, browser actions, agent messages) are frozen into /data/.

How the judge decides PASS / FAIL — task 001's url_pattern is the intentional sentinel __PLACEHOLDER_WILL_NOT_MATCH__, which means no request path can mechanically match. The verdict comes from the agentic judge in eval/agentic_eval.md, which replays the five-layer recording against a human reference run and checks four things:

• Did the agent actually reach the final checkout step?
• Is the cart exactly one Pad Thai (not two, not a combo)?
• Is the delivery address the user's home address from alex_green_personal_info.json?
• Does the order carry the "no peanuts" note in the instructions field?

All four must hold for a PASS. Miss any one and it's a FAIL with evidence from the recording pinned to the failing criterion. This per-task rubric is what makes ClawBench judge-sensitive rather than URL-regex-sensitive — see eval/README.md for the full rubric format and eval/agentic_eval.md for the judge prompt.

Evaluation

Evaluation is a post-session step — first run agents to collect trajectories, then evaluate them against human reference runs.

 1. Run agents (root uv package) 2. Evaluate (eval/)
───────────────────────── ────────────────────────────────
./run.sh / clawbench-batch ──► Claude Code subagents compare
produces test-output/ agent vs human trajectories
with 5-layer recordings under eval/agentic_eval.md rubric

The evaluator compares each agent trajectory against a human reference trajectory across all five recording layers (video, screenshots, HTTP traffic, browser actions, agent messages), then outputs PASS/FAIL with evidence-backed justification.

See eval/README.md for the full evaluation guide and Claude Code prompt template.

CLI

./run.sh # interactive TUI
uv run clawbench-run <case-dir><model># one task
uv run clawbench-run <case-dir> --human # human reference run
uv run clawbench-batch --models <model> --cases-suite v2 --all-cases # a whole corpus

Every command, flag, and suite selector: docs/cli.md.

V1 tasks are in test-cases/v1/ (152 tasks), V2 in test-cases/v2/ (129), Lite in test-cases/v1-lite/ (20), and converted Claw-Eval tasks in test-cases/claw-eval/ (19). All suites use test-cases/task.schema.json. For test case authoring, see CONTRIBUTING.md; for output structure and evaluation guidance, see eval/README.md.

How ClawBench compares

Benchmark	Domain	Environment	Task count	ClawBench difference
WebArena	Synthetic web apps	Self-hosted replicas	812	Live consumer sites, not admin UIs on hosted replicas
GAIA	General assistants	Closed-book text + tools	466	Browser-centric; end-to-end task execution
SWE-bench	Software engineering	GitHub repos	2,294	Non-code; everyday consumer workflows
BrowserGym	Web agents	Headless sandbox	—	Cloud-parity; records real user journeys
Mind2Web	Web navigation	Static traces	2,350	Dynamic live websites, not replayed traces
Online-Mind2Web	Live web navigation	Real websites	300	4× more tasks (V1+V2: 281 vs 300 — comparable), with full 5-layer recordings
VisualWebArena	Visual web tasks	Self-hosted (3 sites)	910	Real websites with full visual layer (vs 3 hosted apps)
WebVoyager	Real-website nav	Real websites (15)	643	Interception-graded vs LLM-judge-only, 143 sites covered
TheAgentCompany	Office workflows	Self-hosted (6 platforms)	175	Consumer everyday tasks instead of enterprise sandbox

ClawBench's niche: live consumer websites, everyday tasks, end-to-end recording. If you want a controlled sandbox or replayed traces, the projects above are excellent. If you want to know whether your agent can actually order food or book a flight today, this is the benchmark for that.

FAQ

What it is

What is ClawBench?

An open-source benchmark for AI browser agents — the systems (GPT-based, Claude-based, or open) that drive a real web browser to complete a user's task. V1 measures whether the agent actually finishes 152 everyday online tasks across 143 live websites; V2 adds a 129-task corpus in test-cases/v2/. It measures completion, not whether the agent produces the right-looking text.

What kinds of tasks does it cover?

Fifteen life categories: food delivery, travel booking, job applications, shopping, housing search, email and calendar management, academic research, software development, learning platforms, and more. Every task is something a normal person might do in a normal week, on a real website.

Are ~150 tasks enough for evaluation?

Yes for a V1 benchmark signal: the tasks span 143 live websites and 15 life categories, and each full run is expensive because it uses isolated containers, real websites, five-layer recording, and post-session judgment against human references. V2 adds another 129 tasks. For cheaper iteration, start with the 20-task test-cases/v1-lite/ subset.

What's the current top score?

33.3% — roughly one task in three — from the strongest frontier model we evaluated on V1. The majority of tasks still defeat every model we've tested; the headroom is real, and the benchmark is not saturated.

How does ClawBench relate to HarnessBench?

Same scoring pipeline, orthogonal axis. ClawBench fixes the harness and varies the model; HarnessBench fixes the model and varies the harness. They share the V1 corpus, the five-layer recording, and the agentic evaluator — so numbers are directly comparable.

Running it

What data does each run produce?

Each session records five layers of synchronized data under /data/:

Layer	File	Description
Session replay	recording.mp4 or run-meta.json recording URL	Local H.264 video or Browserbase Session Inspector replay
Action screenshots	screenshots/*.png	Throttled timestamped PNGs captured after browser actions
Browser actions	actions.jsonl	Every DOM event (click, keydown, input, pageLoad, scroll, etc.)
HTTP traffic	requests.jsonl	Every HTTP request with headers, body, and query params
Agent messages	agent-messages.jsonl	Full agent conversation transcript (thinking, text, tool calls)

For the Pi harness, agent-messages.jsonl is filtered Pi JSON mode output, including message_start/message_end events, tool_execution_* events, tool-call content blocks, and thinking blocks when the selected model emits reasoning. Streaming message_update fragments, including *_delta rows, are omitted because complete assistant messages are already preserved in message_end events.

Harness diagnostic logs such as Pi's agent.log and proxy.log are not copied into the final data/ directory. The interceptor result is saved to interception.json.

What is the synthetic user profile?

Each container gets a /my-info/ directory with a dummy user identity (Alex Green): personal info JSON, email credentials, and a resume PDF. The email is a fresh disposable PurelyMail address generated per run. The agent reads these files when it needs to fill forms, register accounts, etc.

Source templates: src/clawbench/runtime/shared/alex_green_personal_info.json (profile) and src/clawbench/runner/run_support/resume_template.json (resume).

How do account login, registration, and initial task state work?

Each run receives that synthetic profile plus a fresh disposable address. If a task requires sign-up, the agent normally starts from scratch and registers during the run. If a task needs starting files or workspace context, those live under the task's extra_info/ directory and are mounted for the agent at runtime.

What tools can the agent use?

All supported harnesses run inside the same container recording and interception environment. CLI/MCP harnesses expose the browser tool plus a restricted set of read-only shell commands (ls, cat, find, grep, head, tail, jq, wc, etc.); commands that could bypass the browser (curl, python, node, wget) are blocked. Hermes and Pi use native browser/file tools attached to the same ClawBench Chrome CDP endpoint. The Pi harness intentionally allowlists only read-only file tools and browser interaction tools; bash, write, edit, browser_http_get, and browser_run_script are not enabled. The agent instruction also explicitly requires browser-only task completion.

Can I use Podman instead of Docker?

Yes. Set export CONTAINER_ENGINE=podman. The framework auto-detects whichever is available. Podman works without root privileges. (Harbor runs are the exception — they use Harbor's Docker provider.)

Is ClawBench tightly coupled to OpenClaw? Can it evaluate CLI agents?

No, and yes. OpenClaw is the default harness, but harnesses are interchangeable — see the table in Quick start and the registry at src/clawbench/runtime/harnesses/harnesses.yaml. CLI and coding-agent harnesses drive the same instrumented Chromium session using native tools or MCPs.

Scoring and safety

How is a task judged successful?

Each task runs in an isolated browser container with a five-layer recording. For the original V1 results, an evaluator compares the agent trajectory against human reference runs and assigns PASS/FAIL with evidence from the recording. For V2 and newer leaderboard rows, scoring is two-stage: first, the request interceptor checks whether the final blocked HTTP request matches the task's URL/method schema; second, an LLM judge checks whether the captured request payload fulfills the natural-language instruction.

How does the request interceptor work — and is this safe to run against live websites?

The interceptor blocks critical, irreversible HTTP requests (checkout, form submit, email send) to prevent real-world side effects. It connects to Chrome via CDP's Fetch domain and matches requests against the eval schema (url_pattern regex + method + optional body/params). When triggered, it saves the blocked request to interception.json, kills the agent, and stops recording. Tasks that need to simulate an irreversible action (e.g. "add to cart and checkout") terminate at the last reversible step; you can relax the interceptor per-task if your research requires it.

The interceptor does not validate task completion — that is handled separately post-session. For tasks behind payment walls (the agent has no valid credit card), the eval schema uses a placeholder pattern that never matches, so the session runs until timeout.

Which harness are the published model results based on?

The repo default is openclaw, but leaderboard rows include their harness explicitly. V1 results used OpenClaw; newer runs may use Hermes or other supported harnesses. Use the harness column when comparing models, because model and harness changes are separate experimental axes.

What happens when live websites change?

Live-site change is part of the benchmark's target: ClawBench measures whether agents can handle production websites rather than frozen snapshots. That also means some runs can be affected by layout changes, availability, anti-bot systems, or alternate flows. Reproducibility comes from publishing task definitions, eval schemas, run metadata, and five-layer traces; repeated runs over time are still useful for measuring site drift.

Do CAPTCHA or bot checks dominate failures?

If an agent encounters a CAPTCHA, it must attempt it. We have seen cases where frontier models are able to solve some CAPTCHAs. CAPTCHA failures can reflect model behavior, browser-control stack limits, or site defenses. The trace datasets make these failures inspectable.

Contributing and coverage

How do I add a new test case?

See CONTRIBUTING.md. In short: create a directory under the target corpus (test-cases/v1/ or test-cases/v2/) with a task.json conforming to test-cases/task.schema.json, define the eval schema, test with human mode, and submit a PR. Harness definitions live in src/clawbench/runtime/harnesses/harnesses.yaml.

How do I reproduce a published score?

See Reproduce the leaderboard for the two verification paths and the pass criterion.

Will newer models be added?

Yes. New model runs can be submitted or requested through the contribution flow and issues. Public rows are added as complete or clearly marked partial runs, depending on what has finished.

✨ Awesome Works using ClawBench

We are delighted to see ClawBench used across the research community.

The growing community includes authors from Google DeepMind, Stanford, UC Berkeley, Google, Microsoft Research, Harvard, ETH Zürich, Oxford, Northwestern, ByteDance Seed, HKUST, and more.

If we missed your work, please open an issue or submit a pull request.

😊 Google DeepMind, University of Oxford & Columbia University, The Recipe for Intelligence in Natural and Artificial Systems (DOI)

😊 Stanford, UC Berkeley, Microsoft Research & UCSB, Auditing Agent Harness Safety (Code, Project)

😊 Google, Agentic Coding Needs Proactivity, Not Just Autonomy (Google Research Blog)

😊 Harvard Kempner Institute, Massachusetts General Hospital & CUHK, NeuroClaw Technical Report (Code, Project)

😊 ETH Zürich & Handshake AI Research, Verifying Agents in Rubric-Graded Environments (Code, Workshop)

😊 University of Oxford, NUS & Peking University, OpenClaw Research: A Systematic Survey of Large Language Model Agents in Open Deployment (Project, Resources)

😊 Northwestern University, A Survey of the OpenClaw Ecosystem: From Platform Extensibility to Constraint Design (Project & Resources)

😊 UC Davis & UT Dallas, Toward Trustworthy Computer-Use Agents: Risk Propagation, Evaluation Gaps, and Human Governance (Code & Project, Resources)

😊 ByteDance Seed & HKUST, Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context (Models)

😊 Tencent Hunyuan & Fudan University, TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training

😊 Unipat AI, VibeSearchBench: Benchmarking Long-horizon Proactive Search in the Wild (Code, Project)

😊 Tsinghua University & CUHK, WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation (Code, Project)

😊 NUS, HKUST, Tsinghua University & Peking University, Towards Long-Horizon Agents: A Survey (Project, Resources)

😊 HKU MMLab, UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks (Code, Project)

😊 Tsinghua University & SJTU, MacAgentBench: Benchmarking AI Agents on Real-World macOS Desktop (Code, Project)

😊 Peking University & CUHK, π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows (Code, Project)

😊 SJTU, AcademiClaw: When Students Set Challenges for AI Agents (Code, Project)

Contributing

We welcome contributions -- especially new test cases. If you've ever ordered groceries, booked an appointment, or filed a form online, you already know how to write one. Most PRs are a single JSON file and land in under a day.

Quick wins:

• Add a new test case (~30 min, no container expertise needed)
• Add a new category of 10+ tasks → co-author invitation on the next paper revision
• Submit a new model to the public leaderboard
• Browse good first issues

See CONTRIBUTING.md for the full guide and contributor recognition policy.

Community

Come hang out with researchers, builders, and contributors working on real-world browser agents.

Questions & bugs Fastest route to a maintainer

中文社区 研究者、开发者、贡献者交流

Dataset & leaderboard talk On the Hub, next to the data

Citation

If you use ClawBench in your research, please cite:

@misc{zhang2026clawbenchaiagentscomplete,
title = {ClawBench: Can AI Agents Complete Everyday Online Tasks?},
author = {Yuxuan Zhang and Yubo Wang and Yipeng Zhu and Penghui Du and Junwen Miao and Xuan Lu and Wendong Xu and Yunzhuo Hao and Songcheng Cai and Xiaochen Wang and Huaisong Zhang and Xian Wu and Yi Lu and Minyi Lei and Kai Zou and Huifeng Yin and Ping Nie and Liang Chen and Dongfu Jiang and Wenhu Chen and Kelsey R. Allen},
year = {2026},
eprint = {2604.08523},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2604.08523}
}

Contact

Questions, suggestions, or research collaboration? Reach the maintainer:

• Yuxuan Zhang — reacher [at] cs.ubc.ca (UBC, NAIL Group) · Homepage ↗
• For bug reports or feature requests, please open a GitHub issue — it's faster than email and gets seen by all maintainers.

Core Contributors

Yuxuan Zhang

Yubo Wang

Perry Zhu

Penghui Du

Junwen Miao

Advisors

Kelsey R. Allen

Wenhu Chen

Dongfu Jiang

Liang Chen

Support ClawBench

If ClawBench is useful for your research or product work, the single most helpful thing you can do is star the repo — it surfaces the benchmark to other AI-agent researchers and helps us justify continued dataset curation.

Open to contributions — new test cases, bug fixes, or evaluation submissions for a model we haven't scored yet. See CONTRIBUTING.md.

Star History

License & Acknowledgments

Apache 2.0 -- see LICENSE.

The converted Claw-Eval suite in test-cases/claw-eval/ is derived from claw-eval/claw-eval and the claw-eval/Claw-Eval dataset, which are released under the MIT License. Third-party package notices are in NOTICE.

Built with OpenClaw, opencode, Claude Code, the Claude in Chrome extension, OpenAI Codex CLI, browser-use, claw-code, Hermes Agent, Pi with pi-browser-harness, and WebBrain (selectable harnesses), Microsoft Playwright MCP (browser control bridge for the opencode, claude-code, codex, and claw-code harnesses), LiteLLM (API translation proxy for the claude-code, claude-code-chrome-extension, codex, browser-use, claw-code, and pi harnesses), noVNC (MPL 2.0), and websockify (LGPL 3.0).