Sam Finton

Introducing Browse.sh, the largest open-source catalog of skills to reliably perform any task on the internet. We've researched hundreds of sites to give your agents the playbook they need to navigate the web.

You can use the new experimental codex code_mode with: ``` [features] code_mode = true ``` There's also `code_mode_only = true` which is fun

we are the best

@ndbroadbent Have you seen the new cloudflare artifacts? Could be helpful, and then your DAG sits on top as you've described

I've been thinking about how to build an autonomous software factory that actually works. the idea is to replace everything with one unified system: version control, project and task management, web framework, and source code as typed "units" github.com/ndbroadbent/Co…

@garrytan SOLID protocol here we comeee

Interesting ongoing war: is it our data or is it the SaaS system of record’s data? I think it is ours. Customers who are smart and are fully codegen-pilled will switch off closed SaaS platforms to open source in giant waves the next 2 years. Count on it.

Your agents suck when using the web because 85% of it doesn't have an API. Browserbase gives them everything they need to do work online. Leading AI companies like Ramp, Lovable, and Clay trust us to power agents that do real work on behalf of real people. With a single API key, your agent gets everything it needs to navigate the wild web: browsers, search, fetch, identity, a sandbox runtime, and model gateway. Stop waiting on integrations, build agents that can browse and interact with the web just like humans.

hmu if you've got a buddy who wants to come work with us

``` Create this new skill and show the user that they can use plain-English requests to search their browser history like: "Find that site I visited recently about Latvian beaches." --- name: browser-history description: Inspect local browser history in read-only mode by querying browser SQLite databases from Firefox, Chromium-based browsers, and Safari. Use when Codex needs to find a site the user visited by searching local browser history across macOS, Linux, or Windows. --- Query the DB with `sqlite3 "file:/absolute/path/to/db?mode=ro&immutable=1"`. Browsers: - Firefox - Paths: - macOS: `~/Library/Application Support/Firefox/Profiles/*/places.sqlite` - Linux: `~/.mozilla/firefox/*/places.sqlite` - Windows: `%APPDATA%\Mozilla\Firefox\Profiles\*\places.sqlite` - Mapping: `moz_historyvisits` + `moz_places`; Unix microseconds; search `url`, `title`, `description` - Chromium - Paths: - macOS: browser user-data roots under `~/Library/Application Support/...`; use `Default/History` or `Profile */History` - Linux: browser user-data roots under `~/.config/...`; use `Default/History` or `Profile */History` - Windows: browser user-data roots under `%LOCALAPPDATA%\...`; use `Default/History` or `Profile */History` - Mapping: `visits` + `urls`; Windows-epoch microseconds; search `url`, `title` - Safari - Paths: - macOS: `~/Library/Safari/History.db` - Mapping: `history_visits` + `history_items`; Apple-epoch seconds; search `title`, `url` Example: ```sql SELECT datetime(mhv.visit_date / 1000000, 'unixepoch', 'localtime') AS visit_time, mp.title, mp.url, mp.description FROM moz_historyvisits mhv JOIN moz_places mp ON mhv.place_id = mp.id WHERE mhv.visit_date >= (strftime('%s', 'now', '-7 days') * 1000000) AND ( lower(coalesce(mp.title, '')) LIKE '%keyword%' OR lower(mp.url) LIKE '%keyword%' OR lower(coalesce(mp.description, '')) LIKE '%keyword%' ) ORDER BY mhv.visit_date DESC LIMIT 50; ```

@marckohlbrugge @Stagehanddev and browserbase, lmk if I can help you get set up

What’s the most reliable way to automate a browser with AI ? Paid API or open source

I just want to be able to have my LLM figure out a website once and have deterministic code for interacting with that website from then on. This gets us closer!

This launch just made every AI agent on Browserbase 99% faster. Stagehand Caching is our managed cache of every action AI can take on the internet. It's like a reverse CDN: we memorize the actions your agent does so you don't need to waste tokens and time repeating it.

@fashiongiik That wouldn't solve the problem i'm describing, although i'm clearly not describing it very well - it's that sometimes I want to write a TUI or something as part of my skill script, and I want to be able to interact with that directly (but have the LLM call it for me)

@monadoid Just make an mcp app (tool) that the AI uses after it runs a script. Tool input is the output of the script. Im not sure why it has to be a change to the skill standard.

We need "chatgpt apps" for skills - I want my LLM to be able to show me outputs from the scripts it runs *directly*, instead of always having to interpret the outputs and relaying them to me.

@fashiongiik @fashiongiik I'm imagining that this would be a change to the skills standard (and/or MCP) :)

@monadoid You can make it using vibecoding platforms like usefractal.dev

@burcs @janwilmake Check this out, similar to what I was working on with cubby!

onboarding was painful so i made a native mac app thanks to cloudflare it now: • automatically starts a local bridge on your mac • spins up a secure tunnel (no ports to open) • deploys an AI agent to a worker running locally • hands you a URL that works from anywhere

Over christmas I started watching videos on X, and my timeline basically turned into tiktok. I wrote a little tampermonkey script to fix this - "Hide videos initially" Check it out here: greasyfork.org/en/scripts/561…

COURAGE: This influencer with a stammer practices speaking to people in public to encourage others with the same affliction.

Very bullish on lint rules over the next few years Anything that can be represented as an agents rule should instead be a lint rule, you since you get: - Clear error messages to the model - Actual enforcement of rules - Save on tokens in context

@RhysSullivan Strongly agree - rust slaps for this pattern as well. llm writes dylint rules on top of already strict clippy rules. There is a ton of room for improvement for deterministic lint "guardrails", and maybe even relying on llms for higher level "pattern guardrails"