A local browser API
for your agents.
Let your agent search, read pages, and use signed-in sites through your browser. You approve access.
Built with Rust Compact JSON Your browser, your choice
Search for “rust browser automation libraries” and return three results.
I’ll search with your local browser.
Local Chrome
duckduckgo.com/?q=rust%20browser%20automation%20libraries
rust browser automation libraries
Your everyday ChromeSigned in
Chrome asks for your approval
Use your current logins.
Connect to Chrome and use the sites you’re already signed into.
How to connect
- In Chrome 144+, open
chrome://inspect/#remote-debugging and enable remote debugging. - Run the command above, then approve Chrome’s connection prompt.
On macOS and Linux, approve once per connection, not per search. Run lsearch disconnect to end access without closing Chrome. Reconnecting asks for approval again.
Keep agent work separate.
Sign in once in a browser profile just for your agent.
How to set up
- Run the command above to open the local-search browser.
- Sign into the sites your agent needs in that browser.
Your everyday Chrome is untouched. Logins last until you or the site sign out.
Your choice is saved. If it disconnects, lsearch tells you. It never switches browsers.
Less context per result.
Normalized tokens per result. Lower is better.
12 queries · 24 requests / provider
- local-search
53.4 tokensBaseline
- Exa
881.2 tokens16.5× baseline
- Brave Search
108.6 tokens2.0× baseline
- Tavily
259.5 tokens4.9× baseline
- Firecrawl
74.8 tokens1.4× baseline
Recorded July 21, 2026. Local timing includes fresh and cached searches: 384.5 ms cold · 6.5 ms cached.
96.5%
less agent context.
Three results instead of a whole search page.
local-search JSON309 tokens
Browser page snapshot8,760.5 tokens
Command + JSON vs. text snapshots of the page. Not images.
How we measured
The hosted comparison uses 12 identical queries at 3- and 10-result depths. Responses are normalized to a shared result schema. Token counts use o200k_base. Pricing is recorded as of July 21, 2026.
Local-search uses its normal five-minute cache: 384.5 ms cold and 6.5 ms cached, with a combined 148.7 ms median. These are recorded results for this workload, not a guarantee for every search.
The context comparison covers 36 searches at three-result depth. It compares visible command text plus JSON output with compact interactive browser snapshots. It does not measure screenshot image tokens.
Local-search fulfilled 24/24 requested depths in the hosted comparison. Tavily fulfilled 17/24; the other providers fulfilled 24/24.
Recorded July 21, 2026| Provider | Tokens / result | Median (ms) | 24-request usage |
|---|
| local-search | 53.4 | 148.7 | $0 |
|---|
| Exa | 881.2 | 501.9 | $0.324 |
|---|
| Brave Search | 108.6 | 322.0 | $0.120 |
|---|
| Tavily | 259.5 | 1,184.4 | $0.192† |
|---|
| Firecrawl | 74.8 | 1,520.8 | ≈$0.154‡ |
|---|
† Tavily: 24 credits at $0.008 per credit. ‡ Firecrawl: Hobby-plan equivalent, 48 × ($16 / 5,000). No pay-as-you-go plan.
Read the methodology Original graphic Get the data
What is local-search?
A command-line tool that lets agents search the web, read pages, extract data, and use websites through Chrome or Chromium on your machine.
Can it use my logins?
Yes. On macOS/Linux, approve one existing-Chrome connection and reuse it across commands. Run lsearch disconnect to end access. Reconnecting requires approval again. Or sign in once in a separate persistent profile.
Which agents work with it?
Claude Code, Codex, Cursor, OpenClaw, or any agent that can run a shell command. Results come back as compact JSON.
Which search engines can I use?
Google, Bing, Brave, or DuckDuckGo. Choose with --engine. All return the same JSON fields: rank, title, URL, domain, and snippet.
Does it bypass logins or CAPTCHAs?
No. Your account still needs access. If a site blocks a search or asks for verification, local-search reports it.
Where does my data go?
Your browser sends cookies to the sites you visit. local-search has no hosted service. Page content goes to your agent and may reach its model provider. Only grant access to accounts you trust it with.
How small and fast is it?
Version 0.2.0 measured 1.26 MB on arm64 macOS on September 25, 2026. An earlier July build measured 4.61 ms median warm startup over 200 launches. Startup has not been remeasured for 0.2.0. Results vary by build and machine.
Is it free and open source?
Yes. MIT licensed, with no local-search subscription or per-search fee. You need a local browser and internet access. Each site’s terms still apply.