{"slug": "libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead", "title": "Libretto packages six Playwright tools to reduce browser-agent overhead", "summary": "Libretto, a two-person San Francisco startup founded by Michael Kronovet and Tanishq Kancharla, released Browser Tools SDK, a package of six Playwright-based tools designed to reduce token usage and runtime overhead for browser-controlling AI agents. The SDK, which emerged from the founders' earlier healthcare pivot Saffron Health, centers on a snapshot tool that returns a compact accessibility tree and an execution tool that runs model-written Playwright code against live pages. Libretto entered Y Combinator's Spring 2025 batch as Saffron Health before pivoting to browser tooling after finding existing agent frameworks too costly and unpredictable for healthcare workflows.", "body_md": "[Michael Kronovet](https://michaelkronovet.com/my_resume.pdf?ref=runtimewire) and Tanishq Kancharla, the founders of [Libretto](https://libretto.sh/?ref=runtimewire), built [Browser Tools SDK](https://libretto.sh/browser-tools?ref=runtimewire) around six tools for reading and controlling real browsers through Playwright. Their bet is that agents can complete web tasks with fewer tokens and less runtime machinery when the model receives a compact page representation and writes the Playwright code for each action.\n\nThe release is the latest product from a two-person, San Francisco-based company that began in healthcare. Kronovet previously led Palantir's work with the U.S. State Department and built the machine-learning backend at Curia.ai. Kancharla was a product engineer at Shortwave, where Y Combinator says he led development of its iOS and desktop apps. Both studied at Carnegie Mellon University, with Kronovet completing a statistics and machine learning degree in under three years and Kancharla studying physics and computer science.\n\nThe pair entered [Y Combinator's Spring 2025 batch](https://www.ycombinator.com/companies/libretto?ref=runtimewire) as Saffron Health, initially working on AI software for chronic care before moving into specialist-referral automation. Their browser tooling emerged from the least glamorous part of that product: connecting to electronic health records and insurance portals that lacked usable integrations.\n\n### A healthcare pivot produced the browser stack\n\nKronovet and Kancharla described the origin of Libretto in a [retrospective on Saffron Health](https://www.saffron.health/?ref=runtimewire). After meeting rural clinic operator Hamilton Health Box in June 2025, they spent roughly five months manually handling referrals while building software to automate insurance authorizations, appointment coordination and document retrieval.\n\nThe referral product entered use at the clinic network, and the founders said the pilot produced revenue. Sales outside that customer stalled. By February 2026, prospective buyers were telling Kronovet and Kancharla that referral automation was useful yet low on their spending priorities.\n\nAt the same time, other health-tech founders were asking about the internal browser-integration system Saffron had built. The pair had tried agent-based tools including Stagehand and Browser Use, according to their retrospective, then moved toward direct browser scripts after finding runtime agents too costly and unpredictable for workflows involving patient data. Maintaining those scripts as websites changed created another engineering burden.\n\nThat work became [Libretto](https://libretto.sh/?ref=runtimewire), an open-source toolkit spanning a workflow-recording CLI, agents that propose repairs for failed Playwright runs, and the new browser SDK. The repository remained under Saffron Health's GitHub namespace and had about 804 stars when checked on July 21st.\n\nThe pivot gives Kronovet and Kancharla a sharper technical customer and a product born from repeated use. It also moves Libretto into a crowded market where browser infrastructure providers, agent frameworks and automation vendors are competing to define how models interact with websites.\n\n### Two tools carry most of the workload\n\nBrowser Tools SDK's design centers on `browser_snapshot`\n\nand `browser_exec`\n\n. The snapshot tool returns a compact accessibility tree with stable references instead of sending raw HTML to the model. The execution tool runs model-written Playwright code against the live page, then returns a snapshot diff showing what changed.\n\nFour additional tools open or connect to browser sessions, report session state and close sessions. The package includes adapters for [Vercel's AI SDK](https://libretto.sh/docs/browser-tools/adapters/ai-sdk?ref=runtimewire) and [Pi](https://libretto.sh/docs/browser-tools/adapters/pi?ref=runtimewire), along with a [framework-neutral interface](https://libretto.sh/docs/browser-tools/adapters/custom?ref=runtimewire). Developers can run Chromium locally or connect the tools to Libretto Cloud, Browserbase, Kernel and Steel.\n\nThe SDK does not replace Playwright. It limits the interface an agent must understand while preserving access to Playwright's selectors and browser APIs. That approach gives developers inspectable code and a smaller stream of page context, two practical controls for agents that otherwise consume large DOM dumps or choose among many narrowly defined actions.\n\nLibretto distributes the package under the MIT license through its [public GitHub repository](https://github.com/saffron-health/libretto/tree/main/packages/browser-tools?ref=runtimewire). It can be installed with `npm i libretto-browser-tools`\n\n.\n\n### Libretto's benchmark needs a careful reading\n\nLibretto promotes Browser Tools SDK as costing 55% less than the alternatives in its test. The underlying [benchmark results](https://github.com/saffron-health/libretto/blob/main/packages/browser-tools/benchmarks/RESULTS.md?ref=runtimewire), completed between July 17th and July 20th, cover 26 tasks on live search, commerce, travel, delivery, real-estate and documentation websites.\n\nAcross three runs, Libretto tested Browser Tools against `agent-browser`\n\n, `playwright-cli`\n\nand `dev-browser`\n\n. Each full run scheduled 104 attempts using the model identified in the methodology as GPT-5.6 Sol through the Pi agent and Browser Use Cloud.\n\nBrowser Tools' selected result passed 24 of 26 tasks at an average agent cost of $0.106 per successful task. Dev-browser also passed 24 tasks in its selected result, at $0.257 per pass. Agent-browser passed 23 at $0.235, while playwright-cli passed 22 at $0.293.\n\nThose rows were selected separately for each harness from whichever run produced its best result. They were not recorded in one simultaneous head-to-head run. Libretto also cautioned that the results were exploratory rather than causal rankings, with live content, proxy reputation and agent behavior changing between runs. Anti-bot systems accounted for every failure in the selected Browser Tools result.\n\nThe benchmark supports the narrower claim that Libretto's compact interface used fewer tokens in this test. Browser Tools consumed 1.45 million agent tokens in its selected run, compared with 3.51 million for dev-browser, while reaching the same pass count. Independent tests across longer workflows and authenticated applications will be needed to establish whether that advantage holds outside Libretto's methodology.\n\n### A code-first bet in a managed-agent market\n\nLibretto is arriving as larger browser platforms push toward managed, natural-language agents. On June 30th, [Browserbase introduced Browserbase Agents](https://www.browserbase.com/blog/introducing-browserbase-agents?ref=runtimewire), which lets customers describe a task and run it through a hosted agent with one API call.\n\nKronovet and Kancharla are pursuing a different layer. Browser Tools can run on Browserbase itself, while Libretto owns the harness, workflow code and repair tooling around the session. That makes browser providers potential infrastructure partners as well as competitors for developer attention.\n\nThe founders have not published current revenue, customer counts or dedicated pricing for Browser Tools SDK. Libretto's disclosed backing is its place in Y Combinator's Spring 2025 batch. Its immediate distribution strategy rests on open source and a low-friction npm package, with the CLI, cloud browser support and debugging products providing the surrounding product surface.\n\nThe healthcare detour supplied Libretto with a demanding proving ground: old portals, unstable interfaces, sensitive data and workflows where an incorrect click carries real consequences. Browser Tools packages the founders' response into six primitives. The open question is whether developers prefer that visible, code-first control as managed browser agents become easier to buy.", "url": "https://wpnews.pro/news/libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead", "canonical_source": "https://runtimewire.com/article/libretto-browser-tools-sdk-ai-agents-playwright", "published_at": "2026-07-21 23:22:18+00:00", "updated_at": "2026-07-21 23:37:30.415373+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "ai-tools", "ai-startups"], "entities": ["Libretto", "Michael Kronovet", "Tanishq Kancharla", "Playwright", "Browser Tools SDK", "Y Combinator", "Saffron Health", "Carnegie Mellon University"], "alternates": {"html": "https://wpnews.pro/news/libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead", "markdown": "https://wpnews.pro/news/libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead.md", "text": "https://wpnews.pro/news/libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead.txt", "jsonld": "https://wpnews.pro/news/libretto-packages-six-playwright-tools-to-reduce-browser-agent-overhead.jsonld"}}