NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1523 most downloaded on PyPI
Browser agent on Pydantic AI: one persistent Python REPL on raw CDP, cloud or local Chrome
Latest stable
1.0.1
Last release 11 days ago
10 Nov 2026
Ships fairly regularly
a new release about every 3 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
2 versions withdrawn
withdrawn after publishing
2 years old
146 releases · first in 2024
Drag across the ruler to compare, or tap a release to open it.
> Release candidate for 1.0.2. Install it explicitly: pip install 'browser-use==1.0.2rc1'. Plain installs still get 1.0.1.
Release candidate for 1.0.2. Install it explicitly:
pip install 'browser-use==1.0.2rc1'. Plain installs still get 1.0.1.
--cli-mcp with Chrome remote debugging off now drives its own Chromium and tells the model how to get your Chrome (toggle at chrome://inspect, restart the server).--cli-mcp CDP calls wait 20 s; a page frozen by a JS dialog fails fast with a clear error.open() in browser_exec code defaults to UTF-8.report_issue in the agent, browser-use feedback, and the browser_feedback MCP tool.PR: https://github.com/browser-use/browser-use/pull/6376
MCP without Chrome: browser-use --cli-mcp on servers, containers and CI boxes now drives its own Chromium, installed on first use. When no browser can
Fixes on 1.0.0.
browser-use --cli-mcp on servers, containers and CI boxes now drives its own Chromium, installed on first use. When no browser can run, you get one clear error.http_get in the CLI and MCP always fetches from your machine.--mcp with allowed_domains can no longer be bypassed through a cloud browser.timeout=: the final answer arrives before the deadline, and provider retries respect it.hasattr(browser_use, '<removed 0.x name>') returns False.uv add 'browser-use==1.0.1'
Your coding agent can read these notes before it upgrades. Set up the MCP server →
browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools (browser_exec, browser_screenshot, browser_login). browser-use init was r…
<img src="https://github.com/browser-use/browser-use/releases/download/1.0.0/browser-use-1.0-vs-0.x-bench.png" alt="browser-use 1.0 vs 0.x: higher score, cheaper, faster, 80% less code" width="100%">
We started over. The agent now writes Python in one persistent REPL, straight on Chrome DevTools Protocol. No DOM serializer, no action schema, nothing in between. Simplicity wins: it is now the state-of-the-art web agent.
● task
│
● model writes a cell new_tab(url); bu.click(41); js(...)
│
● REPL runs it on Chrome (CDP) → output back to the model
│
✓ finish(result)
Same models: up to 33 points higher, 2.7–5.2x cheaper, 2–3x faster.
uv add 'browser-use==1.0.0'
export BROWSER_USE_API_KEY=bu_... # cloud.browser-use.com/new-api-key
from browser_use import Agent
print(Agent('Find the number of stars of the browser-use repo').run_sync().final_result())
Pin with ==. We ship fast from here.
<details> <summary><b>Upgrading from 0.x, benchmarks, everything that changed</b></summary>
Need 0.x? pip install 'browser-use<1'
| 0.x | 1.0 | |
|---|---|---|
Model (llm=None) |
required | bu-fast on the Browser Use gateway |
Browser (Browser()) |
local Chromium | a Browser Use Cloud browser; Browser('chromium') for a local one |
| Key | one per provider | one Browser Use key (BROWSER_USE_API_KEY, or browser-use login) |
| Mode | ultrafast (accessibility-tree ids) on bu-fast / bu-ultrafast; mode='default' opts out |
|
run() limits |
max_steps=500 |
max_steps=100, timeout=1700 s |
Your own model: ChatOpenAI, ChatAnthropic, ChatGoogle, or any Pydantic AI model string (llm='groq:llama-3.3-70b-versatile').
Read MIGRATING.md first. It lists every 0.x option and its 1.0 answer.
TypeError naming it before any browser starts or anything is billed. No silent no-ops. 0.x tuning knobs 1.0 has no use for (use_vision=True, flash_mode, max_failures, ...) warn once and are ignored.ImportError that names the replacement and links MIGRATING.md. except ImportError fallbacks still work.ActionResult(include_in_memory=...) warns and is ignored: every tool result stays in the conversation.run() does not raise: a rejected key or model, a browser that fails to start (cloud create 402/5xx/timeout) and failures inside the run all return status='error' with result.error, and a browser that did start is stopped. A missing key still raises before anything is created.Browser() is a cloud browser. Tasks on localhost or private IPs need Browser('chromium') (one warning says so).Browser(local=True) now means Browser('chromium'). Your own open Chrome is Browser('chrome').Browser('chrome') attaches only to a browser whose live DevToolsActivePort is in your own profile dirs. It no longer attaches to whatever answers on 127.0.0.1:9222.run(max_steps=...) defaults to 100 (was 500). One step is one model request.is_successful() is None when the agent gave no verdict. Use r.is_done() and r.success is not False for the 0.x default.errors() returns only the error that ended the run. Recovered cell errors are in result.messages.result.partial is set only by checkpoint(..., partial=True). To get one, name the flag in the task: "after each page, save the titles with checkpoint('titles.json', titles, partial=True)".result.cost_usd covers model and web search charges, not the cloud browser or its proxy.browser-use-*) is kept after the run. A workspace= path is used as given (made absolute, symlinks not resolved), so result.downloads[0].relative_to(workspace) works.prohibited_domains=['example.com'] also blocks its subdomains. With any domain rule set, every file: URL is blocked. BROWSER_USE_ALLOWED_DOMAINS is not read: Browser() raises while it is set.bu-2-0, bu-1-0, bu-latest, ...) are rejected. BROWSER_USE_LLM_URL is not read: use ChatBrowserUse(base_url=...).TypeError. Azure reads AZURE_OPENAI_API_KEY (not AZURE_OPENAI_KEY) and needs OPENAI_API_VERSION.browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools (browser_exec, browser_screenshot, browser_login). browser-use init was removed.bu and browser console scripts are gone (they broke uv tool install next to other tools). Use browser-use or browseruse.aws, oci, video, examples, eval, core, openai-tools, cli-oci are gone (pip warns and installs without them). anthropic, google, mcp, cli, all still resolve; Anthropic, Google and MCP are core dependencies.Agent(task, output_model_schema=MyModel) then result.structured_output.workspace=, downloads in workspace/downloads and result.downloads. The agent uploads the file you named and never creates or overwrites it.sensitive_data= typed with type_secret() and redacted from cell output, checkpoints, tool results and the result. Not a sandbox; see the trust model in MIGRATING.md.allowed_domains, prohibited_domains, block_ip_addresses) with the 0.x patterns, in every browser mode.web_search=True), billed per search to your Browser Use key and added to result.cost_usd.browser-use pipes Python to browser-harness; browser-use --cli-mcp runs the MCP server; browser-use skill install registers the agent skill.browseruse/browseruse:1.0.0 and ghcr.io/browser-use/browser-use:1.0.0. Every image build is smoke-tested on amd64 and arm64 in CI, on pull requests too.bu-smart turns no longer end at 300 s.server_error, overloaded, request_timeout, upstream_stream_failed) or a stream that ends before its final event is retried once per run, like a 429 or 5xx.Agent raises a low soft open-file limit to min(hard limit, 10240) (macOS defaults to 256). A connection error now names its cause, e.g. ModelConnectionError: Connection error. (OSError: [Errno 24] Too many open files).ChatOpenAI('gpt-6-luna', base_url=...) gets reasoning_effort='none' on any host, so tools work behind a proxy.wait_for_network_idle() no longer waits for Web Worker scripts and cross-site iframes, which ran it to its 10 s cap.beforeunload prompt, or a screenshot with a JS dialog open, names the dialog within seconds.about:blank.BU Bench v2, 60 tasks x 3 reps, cloud browsers, 2026-11-09.
| agent | score [95% CI] | $/task | s/task |
|---|---|---|---|
browser-use 1.0 (mode='ultrafast', the mode bu-fast uses) |
62.5 [58.3, 66.5] | 0.070 | 650 |
| Browser Use Pi (e7007f2), same window | 56.3 [50.2, 62.4] | 0.082 | 743 |
An OpenAI outage hit this window. Pi lost 18 of 180 runs to provider errors, 1.0 lost 1 (its one retry recovered 13 of 14). So the raw gap overstates quality. With provider errors excluded, paired 1.0 minus Pi: score +1.1 [-3.2, +5.7] (no significant difference), cost -$0.013/task [-0.024, -0.004], time -114 s/task [-202, -31]. Same quality, cheaper and faster.
Both arms ran gpt-6-luna on api.openai.com, judged by gpt-5.6-luna (xhigh). The default bu-fast model on the Browser Use gateway was not part of this run. 95% task-clustered bootstrap intervals.
==), and the lockfile is committed. This is a supply-chain policy: you get exactly the versions CI tested. The cost: a project that needs a different version of pydantic, pydantic-ai-slim, mcp, google-genai, pillow or another pinned package cannot install 1.0 next to it. Tell us which pin blocks you.google-genai stays below 3: 3.0.0 calls the Gemini API on v1, which rejects tools. This goes away once Pydantic AI supports it.Browser('chromium') for another viewport.Report problems at https://github.com/browser-use/browser-use/issues
Full changelog: https://github.com/browser-use/browser-use/compare/0.13.14...1.0.0
</details>
browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools ( browser_exec , browser_screenshot ). browser-use init was removed.
Release candidate (second one published, after 1.0.0rc4). Install it explicitly. Plain
pip install browser-useanduv add browser-usestill give you 0.13.14 until 1.0.0 ships. Please report problems at https://github.com/browser-use/browser-use/issues and say you are on 1.0.0rc5.
browser-use 1.0 replaces the engine. The agent no longer picks one action per step from a DOM dump. It writes Python in one persistent REPL with the browser-harness helpers and raw CDP, on Pydantic AI. The library went from 68,815 to 13,154 lines of Python. The 0.x surface you use most (Agent, Browser, Tools, run(), final_result()) still works.
uv add 'browser-use==1.0.0rc5'
export BROWSER_USE_API_KEY=bu_... # get one at https://cloud.browser-use.com/new-api-keyor pip install 'browser-use==1.0.0rc5'. CLI only: uv tool install 'browser-use==1.0.0rc5'. Python 3.11 or newer.
from browser_use import Agent
result = Agent('Find the number of stars of the browser-use repo').run_sync()
print(result.final_result())| 0.x | 1.0 | |
|---|---|---|
Model (llm=None) |
required | bu-fast on the Browser Use gateway |
Browser (Browser()) |
local Chromium | a Browser Use Cloud browser; Browser('chromium') for a local one |
| Key | one per provider | one Browser Use key (BROWSER_USE_API_KEY, or browser-use login) |
| Mode | ultrafast (accessibility-tree ids) on bu-fast / bu-ultrafast; mode='default' opts out |
|
run() limits |
max_steps=500 |
max_steps=100, timeout=1700 s |
Your own model: ChatOpenAI, ChatAnthropic, ChatGoogle, or any Pydantic AI model string (llm='groq:llama-3.3-70b-versatile').
Read MIGRATING.md first. It lists every 0.x option and its 1.0 answer. Need 0.x? Pin it: pip install 'browser-use<1'.
TypeError naming it before any browser starts or anything is billed. No silent no-ops. 0.x tuning knobs 1.0 has no use for (use_vision=True, flash_mode, max_failures, ...) warn once and are ignored.ImportError that names the replacement and links MIGRATING.md. except ImportError fallbacks still work.ActionResult(include_in_memory=...) warns and is ignored: every tool result stays in the conversation.run() does not raise: a rejected key or model, a browser that fails to start (cloud create 402/5xx/timeout) and failures inside the run all return status='error' with result.error, and a browser that did start is stopped. A missing key still raises before anything is created.Browser() is a cloud browser. Tasks on localhost or private IPs need Browser('chromium') (one warning says so).Browser(local=True) now means Browser('chromium'). Your own open Chrome is Browser('chrome').run(max_steps=...) defaults to 100 (was 500). One step is one model request.is_successful() is None when the agent gave no verdict. Use r.is_done() and r.success is not False for the 0.x default.errors() returns only the error that ended the run. Recovered cell errors are in result.messages.result.partial is set only by checkpoint(..., partial=True). To get one, name the flag in the task: "after each page, save the titles with checkpoint('titles.json', titles, partial=True)".result.cost_usd covers model and web search charges, not the cloud browser or its proxy.browser-use-*) is kept after the run. A workspace= path is used as given (made absolute, symlinks not resolved), so result.downloads[0].relative_to(workspace) works.prohibited_domains=['example.com'] also blocks its subdomains. BROWSER_USE_ALLOWED_DOMAINS is not read: Browser() raises while it is set.bu-2-0, bu-1-0, bu-latest, ...) are rejected. BROWSER_USE_LLM_URL is not read: use ChatBrowserUse(base_url=...).TypeError. Azure reads AZURE_OPENAI_API_KEY (not AZURE_OPENAI_KEY) and needs OPENAI_API_VERSION.browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools (browser_exec, browser_screenshot). browser-use init was removed.bu and browser console scripts are gone (they broke uv tool install next to other tools). Use browser-use or browseruse.aws, oci, video, examples, eval, core, openai-tools, cli-oci are gone (pip warns and installs without them). anthropic, google, mcp, cli, all still resolve; Anthropic, Google and MCP are core dependencies.Agent(task, output_model_schema=MyModel) then result.structured_output.workspace=, downloads in workspace/downloads and result.downloads. The agent uploads the file you named and never creates or overwrites it.sensitive_data= typed with type_secret() and redacted from cell output, checkpoints, tool results and the result. Not a sandbox; see the trust model in MIGRATING.md.allowed_domains, prohibited_domains, block_ip_addresses) with the 0.x patterns, in every browser mode.web_search=True), billed per search to your Browser Use key and added to result.cost_usd.browser-use pipes Python to browser-harness; browser-use --cli-mcp runs the MCP server; browser-use skill install registers the agent skill.git clone -b 1.0.0rc5 https://github.com/browser-use/browser-use && cd browser-use && docker build . -t browseruse. Every image build is smoke-tested on amd64 and arm64 in CI, on pull requests too.bu-smart turns no longer end at 300 s.server_error, overloaded, request_timeout, upstream_stream_failed) or a stream that ends before its final event is retried once per run, like a 429 or 5xx.Agent raises a low soft open-file limit to min(hard limit, 10240) (macOS defaults to 256). A connection error now names its cause, e.g. ModelConnectionError: Connection error. (OSError: [Errno 24] Too many open files).ChatOpenAI('gpt-6-luna', base_url=...) gets reasoning_effort='none' on any host, so tools work behind a proxy.wait_for_network_idle() no longer waits for Web Worker scripts and cross-site iframes, which ran it to its 10 s cap.beforeunload prompt, or a screenshot with a JS dialog open, names the dialog within seconds.about:blank.BU Bench v2, 60 tasks x 3 reps, cloud browsers, on this commit:
| agent | score [95% CI] | $/task | s/task |
|---|---|---|---|
browser-use 1.0 (default, mode='ultrafast') |
62.6 [58.2, 67.1] | 0.071 | 585 |
browser-use 1.0, mode='default' |
61.1 [56.5, 65.6] | 0.073 | 675 |
| Browser Use Pi (e7007f2), same window | 61.3 [57.0, 65.4] | 0.087 | 689 |
Paired, 1.0 default minus Pi: score +1.4 [-2.3, +5.0] (no significant difference), cost -$0.016/task [-0.027, -0.004] (18% cheaper), time -105 s/task [-167, -42] (15% faster). Every arm ran gpt-6-luna on api.openai.com, judged by gpt-5.6-luna (xhigh); 178 of 180 runs judged for the default arm, 180 for mode='default', 179 for Pi. The default bu-fast model on the Browser Use gateway was not part of this run.
95% task-clustered bootstrap intervals.
The library and the CLI send anonymized usage telemetry: counts, model and helper names, exception classes, hostnames of visited pages. Never the task, results, code, typed text, secrets or page content. No feature flags. Turn it off with ANONYMIZED_TELEMETRY=false: no event, no background thread, no file written.
Browser.stop() wait 300 s.page_info() right after new_tab(url) no longer fails on local Chromium (CLI and MCP).browser-use login --device-code exits cleanly; clearer errors when no browser is found or Chrome exits at start.Browser('chrome'), the CLI and MCP no longer attach to whatever Chrome answers on port 9222; file: URLs are blocked under every domain rule; MCP http_get checks redirects against domain rules; the daemon folder must be private; page text is stripped of terminal escape codes in logs; Docker base images are pinned.==, never >=: we move fast and releases can break things.Every dependency is pinned exactly (==), and the lockfile is committed. This is a supply-chain policy: you get exactly the versions CI tested. The cost: a project that needs a different version of pydantic, pydantic-ai-slim, mcp, google-genai, pillow or another pinned package cannot install 1.0 next to it. Tell us which pin blocks you.
The docs at docs.browser-use.com still show plain uv add browser-use / pip install browser-use, which install 0.13.14 during the release candidate. Use the install lines above.
google-genai is capped below 3: 3.0.0 calls the Gemini API on v1, which rejects tools. The cap goes away once Pydantic AI supports it.
Viewport emulation does not work on cloud browsers yet. Use Browser('chromium') for another viewport.
Full changelog: 0.13.14...1.0.0rc5
browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools ( browser_exec , browser_screenshot ). browser-use init was removed.
Release candidate (first one published; the version numbers 1.0.0rc1 to rc3 were unusable on PyPI). Install it explicitly. Plain
pip install browser-useanduv add browser-usestill give you 0.13.14 until 1.0.0 ships. Please report problems at https://github.com/browser-use/browser-use/issues and say you are on 1.0.0rc4.
browser-use 1.0 replaces the engine. The agent no longer picks one action per step from a DOM dump. It writes Python in one persistent REPL with the browser-harness helpers and raw CDP, on Pydantic AI. The library went from 68,815 to 13,154 lines of Python. The 0.x surface you use most (Agent, Browser, Tools, run(), final_result()) still works.
uv add --prerelease allow 'browser-use>=1.0.0rc4'
export BROWSER_USE_API_KEY=bu_... # get one at https://cloud.browser-use.com/new-api-keyor pip install --pre 'browser-use>=1.0.0rc4'. CLI only: uv tool install --prerelease allow 'browser-use>=1.0.0rc4'. Python 3.11 or newer.
from browser_use import Agent
result = Agent('Find the number of stars of the browser-use repo').run_sync()
print(result.final_result())| 0.x | 1.0 | |
|---|---|---|
Model (llm=None) |
required | bu-fast on the Browser Use gateway |
Browser (Browser()) |
local Chromium | a Browser Use Cloud browser; Browser('chromium') for a local one |
| Key | one per provider | one Browser Use key (BROWSER_USE_API_KEY, or browser-use login) |
| Mode | ultrafast (accessibility-tree ids) on bu-fast / bu-ultrafast; mode='default' opts out |
|
run() limits |
max_steps=500 |
max_steps=100, timeout=1700 s |
Your own model: ChatOpenAI, ChatAnthropic, ChatGoogle, or any Pydantic AI model string (llm='groq:llama-3.3-70b-versatile').
Read MIGRATING.md first. It lists every 0.x option and its 1.0 answer. Need 0.x? Pin it: pip install 'browser-use<1'.
TypeError naming it before any browser starts or anything is billed. No silent no-ops. 0.x tuning knobs 1.0 has no use for (use_vision=True, flash_mode, max_failures, ...) warn once and are ignored.ImportError that names the replacement and links MIGRATING.md. except ImportError fallbacks still work.ActionResult(include_in_memory=...) warns and is ignored: every tool result stays in the conversation.run() does not raise: a rejected key or model, a browser that fails to start (cloud create 402/5xx/timeout) and failures inside the run all return status='error' with result.error, and a browser that did start is stopped. A missing key still raises before anything is created.Browser() is a cloud browser. Tasks on localhost or private IPs need Browser('chromium') (one warning says so).Browser(local=True) now means Browser('chromium'). Your own open Chrome is Browser('chrome').run(max_steps=...) defaults to 100 (was 500). One step is one model request.is_successful() is None when the agent gave no verdict. Use r.is_done() and r.success is not False for the 0.x default.errors() returns only the error that ended the run. Recovered cell errors are in result.messages.result.partial is set only by checkpoint(..., partial=True). To get one, name the flag in the task: "after each page, save the titles with checkpoint('titles.json', titles, partial=True)".result.cost_usd covers model and web search charges, not the cloud browser or its proxy.browser-use-*) is kept after the run. A workspace= path is used as given (made absolute, symlinks not resolved), so result.downloads[0].relative_to(workspace) works.prohibited_domains=['example.com'] also blocks its subdomains. BROWSER_USE_ALLOWED_DOMAINS is not read: Browser() raises while it is set.bu-2-0, bu-1-0, bu-latest, ...) are rejected. BROWSER_USE_LLM_URL is not read: use ChatBrowserUse(base_url=...).TypeError. Azure reads AZURE_OPENAI_API_KEY (not AZURE_OPENAI_KEY) and needs OPENAI_API_VERSION.browser-use --mcp prints a deprecation notice and serves the --cli-mcp tools (browser_exec, browser_screenshot). browser-use init was removed.bu and browser console scripts are gone (they broke uv tool install next to other tools). Use browser-use or browseruse.aws, oci, video, examples, eval, core, openai-tools, cli-oci are gone (pip warns and installs without them). anthropic, google, mcp, cli, all still resolve; Anthropic, Google and MCP are core dependencies.Agent(task, output_model_schema=MyModel) then result.structured_output.workspace=, downloads in workspace/downloads and result.downloads. The agent uploads the file you named and never creates or overwrites it.sensitive_data= typed with type_secret() and redacted from cell output, checkpoints, tool results and the result. Not a sandbox; see the trust model in MIGRATING.md.allowed_domains, prohibited_domains, block_ip_addresses) with the 0.x patterns, in every browser mode.web_search=True), billed per search to your Browser Use key and added to result.cost_usd.browser-use pipes Python to browser-harness; browser-use --cli-mcp runs the MCP server; browser-use skill install registers the agent skill.git clone -b 1.0.0rc4 https://github.com/browser-use/browser-use && cd browser-use && docker build . -t browseruse. Every image build is smoke-tested on amd64 and arm64 in CI, on pull requests too.bu-smart turns no longer end at 300 s.server_error, overloaded, request_timeout, upstream_stream_failed) or a stream that ends before its final event is retried once per run, like a 429 or 5xx.Agent raises a low soft open-file limit to min(hard limit, 10240) (macOS defaults to 256). A connection error now names its cause, e.g. ModelConnectionError: Connection error. (OSError: [Errno 24] Too many open files).ChatOpenAI('gpt-6-luna', base_url=...) gets reasoning_effort='none' on any host, so tools work behind a proxy.wait_for_network_idle() no longer waits for Web Worker scripts and cross-site iframes, which ran it to its 10 s cap.beforeunload prompt, or a screenshot with a JS dialog open, names the dialog within seconds.about:blank.BU Bench v2, 60 tasks x 3 reps, cloud browsers, on this commit:
| agent | score [95% CI] | $/task | s/task |
|---|---|---|---|
browser-use 1.0 (default, mode='ultrafast') |
62.6 [58.2, 67.1] | 0.071 | 585 |
browser-use 1.0, mode='default' |
61.1 [56.5, 65.6] | 0.073 | 675 |
| Browser Use Pi (e7007f2), same window | 61.3 [57.0, 65.4] | 0.087 | 689 |
Paired, 1.0 default minus Pi: score +1.4 [-2.3, +5.0] (no significant difference), cost -$0.016/task [-0.027, -0.004] (18% cheaper), time -105 s/task [-167, -42] (15% faster). Every arm ran gpt-6-luna on api.openai.com, judged by gpt-5.6-luna (xhigh); 178 of 180 runs judged for the default arm, 180 for mode='default', 179 for Pi. The default bu-fast model on the Browser Use gateway was not part of this run.
95% task-clustered bootstrap intervals.
The library and the CLI send anonymized usage telemetry: counts, model and helper names, exception classes, hostnames of visited pages. Never the task, results, code, typed text, secrets or page content. No feature flags. Turn it off with ANONYMIZED_TELEMETRY=false: no event, no background thread, no file written. Every field is listed in TELEMETRY.md
Every dependency is pinned exactly (==), and the lockfile is committed. This is a supply-chain policy: you get exactly the versions CI tested. The cost: a project that needs a different version of pydantic, pydantic-ai-slim, mcp, google-genai, pillow or another pinned package cannot install 1.0 next to it. Tell us which pin blocks you.
The docs at docs.browser-use.com still show plain uv add browser-use / pip install browser-use, which install 0.13.14 during the release candidate. Use the install lines above.
A file downloaded in a second tab the agent opened makes Browser.stop() wait 300 s before it returns. The file itself is complete. Fixed in the next rc.
CLI and MCP on a local Chromium: print(page_info()) right after new_tab(url) can fail with Cannot read properties of null. Call wait_for_load() in between. Agent runs recover on their own. Fixed in the next rc.
google-genai is capped below 3: 3.0.0 calls the Gemini API on v1, which rejects tools. The cap goes away once Pydantic AI supports it.
Viewport emulation does not work on cloud browsers yet. Use Browser('chromium') for another viewport.
Full changelog: 0.13.14...1.0.0rc4
This release keeps the Python agent and its existing models. The Pi migration remains separate.
This release keeps the Python agent and its existing models. The Pi migration remains separate.
Changes since the last PyPI release:
Validation:
browser-use --help.Install with uv add browser-use==0.13.14 or pip install browser-use==0.13.14.
The protected publishing workflow requires another release reviewer. PyPI publication follows that approval.
This release includes Browser Use toolsets for Claude, available at browser_use.integrations.toolsets_for_claude. The integration requires an Anthropi
This release includes Browser Use toolsets for Claude, available at browser_use.integrations.toolsets_for_claude. The integration requires an Anthropic SDK release containing anthropic.tools.browser; the current public Anthropic SDK does not yet include it.
Documentation: https://docs.browser-use.com/open-source/customize/integrations/toolsets-for-claude
PyPI publication runs through the protected release workflow and requires a separate reviewer approval.
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.10...0.13.11
…CodeQL, GitGuardian, and an installed-runtime vulnerability audit.
Validation includes the full GitHub test matrix, model adapters, MCP stdio handshakes, cross-platform wheel installation, focused PDF tests, CodeQL, GitGuardian, and an installed-runtime vulnerability audit.
fix(deps): update vulnerable Click, MCP, and pypdf pins by @green3sf in https://github.com/browser-use/browser-use/pull/5526
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.8...0.13.9
fix(mcp): update deprecated Bedrock Claude model identifier by @scemii in https://github.com/browser-use/browser-use/pull/5448
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.7...0.13.8
Ships the latest version of the CLI, which includes bug fixes and better browser management.
Ships the latest version of the CLI, which includes bug fixes and better browser management.
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.6...0.13.7
Release 0.13.6 with Browser Harness 0.1.6 by @gregpr07 in https://github.com/browser-use/browser-use/pull/5245
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.5...0.13.6
Update README.md by @ShawnPana in https://github.com/browser-use/browser-use/pull/5197
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.4...0.13.5
MCP server for CLI 3.0: shared MCP runner, max_dim guard.
max_dim guard.0.1.5 (Windows reliability fixes, cloud-browser suggestion).Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.3...0.13.4
Launch Browser Use CLI 3.0 powered by Browser Harness.
browser-use skill so coding agents can install the Browser Use skill directly.0.1.4 and ship the Browser Use skill with the package.Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.2...0.13.3
support claude fable 5 by @laithrw in https://github.com/browser-use/browser-use/pull/5008
Full Changelog: https://github.com/browser-use/browser-use/compare/0.13.0...0.13.1
Browser Use 0.13.0 introduces a new Rust-backed beta agent.
<img width="2912" height="1632" alt="browser-use_a_hooded_traveler_in_saffron_robes_with_a_long_staf_99dc26d6-efe0-4f9b-b2ea-7329f26cdb2a (1)-2" src="https://github.com/user-attachments/assets/d6f49f19-ccb7-4284-afe1-4bad0fa228ae" />
<br/>
Browser Use 0.13.0 introduces a new Rust-backed beta agent.
It gives modern models a more direct browser control loop, guided by robust helpers instead of brittle browser abstractions.
uv add "browser-use[core]"
from browser_use.beta import Agent
The existing Python agent remains unchanged.
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.9...0.13.0
Browser Use 0.13.0 introduces a new Rust-backed beta agent.
It gives modern models a more direct browser control loop, guided by robust helpers instead of brittle browser abstractions.
uv add "browser-use[core]"from browser_use.beta import AgentThe existing Python agent remains unchanged.
Full Changelog: 0.12.9...0.13.0
bump browser-use package version to 0.13.0rc4
Full uv sync --all-extras / install validation must wait until
browser-use-core==0.13.0rc4 finishes publishing from terminal run
https://github.com/browser-use/terminal/actions/runs/27166198498.
Prepare rc4 release by bumping browser-use to 0.13.0rc4 and aligning
the [core] extra to browser-use-core 0.13.0rc4 across all platforms.
This keeps the wrapper and core in sync for clean installs.
Dependencies
browser-use version to 0.13.0rc4.Pin browser-use-core==0.13.0rc4 for macOS (arm64, x86_64), Linux
(x86_64, aarch64), and Windows (x86_64).
Migration
No code changes. Install [core] after browser-use-core 0.13.0rc4
is published.
Written for commit 5fea46d.
Summary will update on new commits.
Nothing published for this version
Nothing published for this version
Pass session id to judge LLM calls by @MagMueller in https://github.com/browser-use/browser-use/pull/4918
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.8...0.12.9
fix(daemon): restrict unix socket file to owner-only access by @sauravpanda in https://github.com/browser-use/browser-use/pull/4870
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.7...0.12.8
fix: upgrade aiohttp to 3.13.4 to patch memory exhaustion vulnerability by @sauravpanda in https://github.com/browser-use/browser-use/pull/4596
record start/stop CLI command for session video capture by @sauravpanda in https://github.com/browser-use/browser-use/pull/4710'type' from second elif branch in optimize_schema() by @kuishou68 in https://github.com/browser-use/browser-use/pull/4704Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.6...0.12.7
fix: set default temperature to 1.0 for gemini-3 models by @markmcd in https://github.com/browser-use/browser-use/pull/4489
<img width="1661" height="278" alt="image" src="https://github.com/user-attachments/assets/151c26e2-b7d8-47e6-b628-1121677527ee" />
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.5...0.12.6
Remove litellm from core dependencies in response to the litellm supply chain attack (versions 1.82.7 & 1.82.8 backdoored on March 24, 2026).
Remove litellm from core dependencies in response to the litellm supply chain attack (versions 1.82.7 & 1.82.8 backdoored on March 24, 2026).
pip install browser-use no longer installs litellmChatLiteLLM wrapper is preserved — install litellm separately if needed: pip install litellmSee: https://github.com/browser-use/browser-use/pull/4515
fix: detect uv installed via curl script by @ishan0102 in https://github.com/browser-use/browser-use/pull/4478
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.3...0.12.4
fix open security advisories: restrict workflow permissions, bump vulnerable deps, remove unused authlib by @laithrw in https://github.com/browser-use…
The fastest browser automation for AI coding agents. 2x faster, 50% fewer tokens, and works with any CLI agent — Claude Code, Codex, and more.
Built on direct CDP (Chrome DevTools Protocol) instead of Playwright, giving ~50ms command latency via a persistent background daemon.
https://github.com/user-attachments/assets/a604895b-4ff9-47be-8161-56c48773f93b
# Install the CLI
curl -fsSL https://browser-use.com/cli/install.sh | bash
# Install the skill (for Claude Code / Codex)
mkdir -p ~/.claude/skills/browser-use
curl -o ~/.claude/skills/browser-use/SKILL.md \
https://raw.githubusercontent.com/browser-use/browser-use/main/skills/browser-use/SKILL.md
Works with any CLI coding agent — Claude Code, Codex, OpenClaw, and more. Just install the CLI and the skill.
browser-use open https://example.com # Navigate (starts browser if needed)
browser-use state # Returns clickable elements: [0] button "Submit", [1] input "Email"...
browser-use click 0 # Click by index
browser-use input 3 "hello@example.com" # Click element, then type
browser-use type "search query" # Type into focused element
browser-use screenshot page.png # See the page
browser-use close # Done
The CLI uses element indices from state — no selectors to maintain. browser-use state is optimized for agents: fast, token-efficient, and returns every interactable element on the page.
| Mode | Command | Description |
|---|---|---|
| Headless | browser-use open <url> |
Default, fast, invisible |
| Headed | browser-use --headed open <url> |
Visible window for debugging |
| Real Chrome | browser-use --profile "Default" open <url> |
Your Chrome with existing logins & cookies |
| Connect | browser-use --connect open <url> |
Auto-discover and attach to running Chrome via CDP |
| CDP URL | browser-use --cdp-url ws://localhost:9222/... open <url> |
Connect to any browser via CDP URL |
| ☁️ Cloud | browser-use cloud connect |
Stealth cloud browser + proxies included |
Chrome only. The CLI uses Chrome DevTools Protocol (CDP), which is Chrome/Chromium-specific. Safari and Firefox are not supported.
Cloud browser requires
BROWSER_USE_API_KEYfrom cloud.browser-use.com. Cloud gives you agent-ready browsers with anti-detection, proxies, and massive parallelization.
One of the most requested features — connect to your already-running Chrome with all your logins, cookies, and extensions:
# Auto-discover running Chrome's CDP endpoint
browser-use --connect open https://github.com
# Or use an existing Chrome profile directly
browser-use --profile "Default" open https://github.com # Already logged in!
browser-use --profile "Profile 1" open https://gmail.com # Named profile
# List available profiles
browser-use profile list
Tip for WSL2 users: Add
browser-use --cdp-urlto your CLAUDE.md to control the host browser from WSL2.
browser-use -s work open https://work.example.com
browser-use -s personal open https://gmail.com
browser-use sessions # List all
browser-use close --all # Close all
The CLI excels at form filling and complex interactions:
browser-use input 5 "user@example.com" # Fill email field
browser-use input 6 "password123" # Fill password
browser-use click 7 # Submit button
browser-use select 3 "United States" # Dropdown selection
browser-use upload 4 ./resume.pdf # Upload a file
browser-use hover 2 # Hover over element
browser-use keys "Enter" # Send keyboard keys
browser-use get title # Page title
browser-use get html --selector "table" # Scoped HTML
browser-use get text 5 # Element text
browser-use eval "document.title" # Run JavaScript
browser-use screenshot --full page.png # Full-page screenshot
browser-use python "browser.goto('https://example.com')"
browser-use python "items = browser.html"
browser-use python "print(len(items))" # Variables persist across calls
browser-use python --vars # Show defined variables
browser-use tunnel 3000 # → https://abc.trycloudflare.com
browser-use open https://abc.trycloudflare.com
browser-use tunnel stop --all
The CLI doesn't use screenshots or image detection. Instead, it directly reads the browser's DOM tree via CDP and returns structured, token-efficient element data. Your coding agent processes the content; the CLI handles browser context and command execution. Every command shares the same battle-tested implementation as the browser-use Python library.
Architecture: A background daemon keeps the browser alive between commands, so there's no startup cost per command — just ~50ms latency.
Does it work with Codex / other CLI agents? Yes — it works with any CLI coding agent. Install the CLI and the skill file.
Can it run headless?
Yes, headless is the default. Use --headed when you want to see the browser.
How does it compare to the Claude Chrome extension? The Browser Use CLI is also free, faster and more token efficient, doesn't require a Chrome extension, and works with any CLI agent (not just Claude).
How does it compare to Vercel Agent Browser? 2x the speed for half the cost with half the commands.
Does it work with Safari/Firefox? No — it uses Chrome DevTools Protocol (CDP), which is Chrome/Chromium only.
Can it upload files?
Yes! Use browser-use upload <index> <path>.
What about security/pentesting? Yes — popular use cases include form filling, QA/pentesting, and general web automation.
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.2...0.12.3
don't explicitly pass newWindow: false to CDP createTarget by @laithrw in https://github.com/browser-use/browser-use/pull/4283
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.1...0.12.2
Fix proxy auth attach handler overwriting tab tracking by @laithrw in https://github.com/browser-use/browser-use/pull/4220
Full Changelog: https://github.com/browser-use/browser-use/compare/0.12.0...0.12.1
improved csv file generation by @sauravpanda in https://github.com/browser-use/browser-use/pull/4213
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.13...0.12.0
updated laminar version by @sauravpanda in https://github.com/browser-use/browser-use/pull/4154
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.12...0.11.13
Fix CDP connectivity issues by @LarsenCundric in https://github.com/browser-use/browser-use/pull/4135
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.11...0.11.12
fix(#3895): stop agent if browser closed by @laithrw in https://github.com/browser-use/browser-use/pull/4056
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.9...0.11.11
Fix download monitoring issues with fetch requests and blob downloads by @sandeepsalwan1 in https://github.com/browser-use/browser-use/pull/3518
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.10a1...0.11.10a2
fix(#3895): stop agent if browser closed by @laithrw in https://github.com/browser-use/browser-use/pull/4056
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.9...0.11.10a1
added basic planing to agents by @sauravpanda in https://github.com/browser-use/browser-use/pull/4008
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.8...0.11.9
docs: add quotes around browser-use[cli] for shell compatibility by @ShawnPana in https://github.com/browser-use/browser-use/pull/4002
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.7...0.11.8
add grep tool to agent by @sauravpanda in https://github.com/browser-use/browser-use/pull/3997
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.6...0.11.7
fix(#3264): clear agent eventbus, not just browser session's by @laithrw in https://github.com/browser-use/browser-use/pull/3964
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.5...0.11.6
Release: 0.11.5 - Browser Use Model bu-2-0 🚀
<html><head></head><body><h1>Release: 0.11.5 - Browser Use Model bu-2-0 🚀</h1> <p>We're excited to announce <strong>bu-2-0</strong>, our latest premium model for browser automation.</p> <h2>What's New</h2> <p><strong>bu-2-0</strong> delivers a significant leap in accuracy while maintaining the speed that made bu-1-0 the fastest web agent:</p>
| Metric | bu-1-0 | bu-2-0 |
|---|---|---|
| Accuracy | 74.7% | 83.3% (+12%) |
| Speed | 58s/task | 60s/task (≈ same) |
<h2>Quick Start</h2> <pre><code class="language-python">from browser_use import Agent, ChatBrowserUse <br> llm = ChatBrowserUse(model='bu-2-0') <br> agent = Agent( task="Your task here", llm=llm )
await agent.run() </code></pre> <h2>Environment Setup</h2> <pre><code class="language-bash">export BROWSER_USE_API_KEY=your_api_key_here </code></pre> <p>Get your API key at <a href="https://cloud.browser-use.com/new-api-key">cloud.browser-use.com</a></p></body></html>
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.4...0.11.5
Give your Claude Code/Codex agent access to local and remote browsers. Really helpful for local dev.
Give your Claude Code/Codex agent access to local and remote browsers. Really helpful for local dev.
<img width="669" height="523" alt="local-qa" src="https://github.com/user-attachments/assets/0fa045c8-bcd2-4ff2-bd21-57300ab36c93" />
<br />
browser-use open https://example.com # Navigate
browser-use state # Returns clickable elements: [0] button "Submit", [1] input "Email"...
browser-use click 0 # Click by index
browser-use type "hello@example.com" # Type text
browser-use screenshot page.png # See the page
browser-use close # Done
Just install the skill and ask Claude to browse:
mkdir -p ~/.claude/skills/browser-use
curl -o ~/.claude/skills/browser-use/SKILL.md \
https://raw.githubusercontent.com/browser-use/browser-use/main/skills/browser-use/SKILL.md
| Mode | Command | Description |
|---|---|---|
| Headless | browser-use open <url> |
Default, fast |
| Headed | browser-use open --headed <url> |
Visible window |
| Real Chrome | browser-use --browser real open <url> |
Your Chrome with existing logins |
| ☁️ Cloud | browser-use --browser remote open <url> |
Stealth browser + proxies included |
Cloud browser requires
BROWSER_USE_API_KEYfrom cloud.browser-use.com. Proxies and anti-detection are automatic - just use--browser remote. Returns a live preview URL.
browser-use -s work open https://work.example.com
browser-use -s personal open https://gmail.com
browser-use sessions # List all
browser-use close --all # Close all
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.3...0.11.4
Bu_oss example flash mode by @mertunsall in https://github.com/browser-use/browser-use/pull/3774
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.2...0.11.3
We released an open source model specialized in Browser Use https://huggingface.co/browser-use/bu-30b-a3b-preview.
We released an open source model specialized in Browser Use https://huggingface.co/browser-use/bu-30b-a3b-preview.
Check it out 🚀
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.1...0.11.2
docs: fix default action count in docs by @sauravpanda in https://github.com/browser-use/browser-use/pull/3741
Full Changelog: https://github.com/browser-use/browser-use/compare/0.11.0...0.11.1
```python from browser_use import Agent, ChatBrowserUse
from browser_use import Agent, ChatBrowserUse
agent = Agent(
task='Your task',
skills=['skill-uuid-1', 'skill-uuid-2'], # Specific skills (recommended)
# or
# skills=['*'], # All skills
llm=ChatBrowserUse()
)
await agent.run()
Full Changelog: https://github.com/browser-use/browser-use/compare/0.10.1...0.11.0
Release again because forgot to bump pyproject :)
Release again because forgot to bump pyproject :)
Full Changelog: https://github.com/browser-use/browser-use/compare/0.10.0...0.10.1
Nothing published for this version
Fix: Ignore Inert Carousel Elements by @PredictiveManish in https://github.com/browser-use/browser-use/pull/3564
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.6...0.9.7
Remove deprecated agent.context documentation by @mertunsall in https://github.com/browser-use/browser-use/pull/3561
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.5...0.9.6
small-fixes2 by @MagMueller in https://github.com/browser-use/browser-use/pull/3457
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.4...0.9.5
make pip workflow funciton by installing uv from pip by @Alezander9 in https://github.com/browser-use/browser-use/pull/3454
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.3...0.9.4
Add DOM serializer test with shadow DOM, iframes, and click functionality by @MagMueller in https://github.com/browser-use/browser-use/pull/3435
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.2...0.9.3
bug-template by @MagMueller in https://github.com/browser-use/browser-use/pull/3416
uvx browser-use install by @ShawnPana in https://github.com/browser-use/browser-use/pull/3429Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.1...0.9.2
readme faq by @MagMueller in https://github.com/browser-use/browser-use/pull/3390
Full Changelog: https://github.com/browser-use/browser-use/compare/0.9.0...0.9.1
Just import the CodeAgent. The APIs are compatible.
<img width="830" height="274" alt="code-use" src="https://github.com/user-attachments/assets/ea7e1a0d-c6e6-46e6-9f40-751ee880ac4c" />
Just import the CodeAgent. The APIs are compatible.
from browser_use import CodeAgent, ChatBrowserUse
agent = CodeAgent(
task=task,
llm=ChatBrowserUse(), # requires our special LLM
)
await agent.run()
We have new highlights which show where the agent clicked:
We have new highlights which show where the agent clicked: <img width="818" height="350" alt="image" src="https://github.com/user-attachments/assets/8bae4e2a-856a-48a6-8a35-03c83946d805" />
Full Changelog: https://github.com/browser-use/browser-use/compare/0.8.0...0.8.1
## New light speed LLMs available Full Changelog: https://github.com/browser-use/browser-use/compare/0.7.12...0.8.0
<img width="651" height="223" alt="image" src="https://github.com/user-attachments/assets/7667fbb3-629c-4de6-87f3-e3eed1a689ed" />
Full Changelog: https://github.com/browser-use/browser-use/compare/0.7.12...0.8.0
0.7.12 by @gregpr07 in https://github.com/browser-use/browser-use/pull/3274
Full Changelog: https://github.com/browser-use/browser-use/compare/0.7.11...0.7.12
fix the spelling mistakes in the using_cdp example by @Cu1111 in https://github.com/browser-use/browser-use/pull/3190
Full Changelog: https://github.com/browser-use/browser-use/compare/0.7.10...0.7.11
speed-ups by @MagMueller in https://github.com/browser-use/browser-use/pull/3172
Full Changelog: https://github.com/browser-use/browser-use/compare/0.7.9...0.7.10
One column per month.