Releases
TinySearch 0.7.0: What Changed Since 0.6.4, and Why It's More Stable
TinySearch 0.7.0 is on PyPI and Docker Hub. The last release post here covered 0.6.4 and the browser tools, in early September. Since then there have been six point releases and now 0.7.0, and most of that work went into making TinySearch behave the same way every time: one browser engine instead of two, limits you can see, a browser per client, and builds that install exactly what was tested.
The tools themselves haven’t changed. search, scrape_urls, browser_navigate, browser_act and get_current_datetime take the same arguments and return the same shapes, so an agent that worked with 0.6.4 works with 0.7.0.
One browser engine instead of two
Until 0.6.5, scraping went through Crawl4AI while the browser tools drove Playwright, so a TinySearch server carried two browser stacks. Scraping now drives Playwright directly, and the scraper and the browser tools share one runtime and one Chromium. That removed a large dependency and its moving parts, and it means a page is loaded the same way whether scrape_urls reads it or the browser tools open it.
A scraped page is read from its accessibility tree (an ARIA snapshot) and split into chunks along its headings, so a section stays with its heading. A few smaller changes came with the switch:
- Embeddings for every page in one
scrape_urlscall are batched and deduplicated, instead of computed page by page. - Images, fonts and media are blocked while scraping. The scraper only reads text, so it no longer downloads them.
- Playwright’s own log noise is filtered out of the server’s stderr.
- The bundled browser can go through a proxy.
Pages that load their content late
Many pages send a thin HTML shell and fill it in with scripts. Before 0.7.0, TinySearch read a page as soon as its HTML had loaded, so on those pages it could read a loading state instead of the content.
0.7.0 waits for the page’s network activity to settle before reading it, for up to scrape_settle_seconds (2.5 by default). A page that settles sooner is read at once, so ordinary pages aren’t slowed down. A page that keeps polling in the background costs at most the full wait. Set it to 0 to read pages the moment their HTML arrives, as before.
The page’s subject first
The text TinySearch hands the model is cut to a token budget (scrape_max_tokens, 2,000 per page by default), so what comes first matters. Two releases changed that.
0.6.11 started cleaning the accessibility snapshot before chunking it. On real pages, half or more of a raw snapshot is outline syntax rather than page text: link targets, role names, empty wrapper nodes, icon-font glyphs, and child nodes that repeat their parent’s name. All of that is now dropped, and every visible string is kept. Headings, list items and table rows keep a plain-text shape (#, -, | a | b |). If a snapshot ever has a shape the cleaner doesn’t recognise, it passes through unchanged, so a format change costs tokens rather than text.
0.7.0 puts the page’s main content first. The text of the page’s main landmark, or its top-level article if it has no main, now comes before the navigation, header, footer and sidebars. Nothing is removed, but a reader that stops at the budget spends it on what the page is about rather than its menus. A page with neither landmark keeps its original order.
0.7.0 also fixes a quiet failure in that cleaner. Text in the snapshot is quoted the way YAML quotes it, and the cleaner was decoding it with a JSON parser. A page containing an escape that YAML allows and JSON doesn’t made the cleaner give up and return the raw snapshot, so that page cost noticeably more tokens. The cleaner now decodes YAML’s escapes itself.
Limits you can see
0.6.11 made the size of a scrape_urls batch configurable with scrape_max_items (5 by default), and the tool’s description and schema now state that number, so the model knows the limit before it builds a call. A second setting, scrape_max_concurrent_items, caps how many of those pages load at the same time. The rest wait their turn. Lower it on a small host, or for sites that throttle parallel requests.
Each client gets its own browser
When several MCP clients share one TinySearch server, they used to share one interactive browser too, so one client’s browser_navigate could change the page another client was in the middle of reading. Since 0.6.7, every client connection gets its own browser, page, tabs and cookies.
browser_max_sessions (4 by default) caps how many are open at once. A browser holds its slot until it closes, either explicitly or after it has been idle for a while, and every browser is closed when the server shuts down cleanly.
Seeing what it did
0.6.7 added opt-in activity recording, and 0.6.8 added a mode that also keeps what the model was shown. It is off unless you turn it on with TINYSEARCH_ACTIVITY. With it on, an MCP client that supports TinySuite’s management views can show an overview of searches and page fetches, the domains TinySearch has read and which fetches failed, and each caller’s own searches and pages. With content mode, opening a fetch shows the exact passages that reached the model. TinyWebUI shows all of this on its MCP page, and the TinySearch demo ends on it.
OpenTelemetry export, when you turn it on, still never carries queries, URLs or page text.
Builds that install the same thing every time
0.7.0 adds a lockfile, uv.lock, and both CI and the Docker image now install from it:
- CI installs with
uv sync --locked, which fails if the lockfile is missing or out of date, so the tests run against the exact versions that ship. - The Docker image installs the same pinned versions, with
uvitself pinned by digest, so rebuilding a tag can’t quietly pull in a different dependency. - The image compiles Python bytecode at build time, so the server doesn’t compile every module each time it starts. Dependency layers are also cached separately from the source, so most rebuilds are quicker.
Installing from PyPI works as before: pip install and uvx resolve the package’s declared version ranges. CI still has a job that builds the wheel and installs it into a fresh environment the way pip would, so that path is tested too. Separately, 0.6.10 added a system library the browser runtime needs to the slim Docker image.
Upgrading
pip install --upgrade tinysuite-searchIf an MCP client launches TinySearch with uvx, run uvx --refresh --from "tinysuite-search[server]" tinysearch once so it picks up 0.7.0 instead of a cached version. The Docker image is marcellm01/tinysearch:v0.7.0. TinyWebUI picks up the new version the same way, the next time uvx refreshes its cache.
There is one new setting, scrape_settle_seconds, and no breaking changes. If a deployment needs scrapes to be as fast as before at the cost of script-loaded content, set it to 0. The configuration reference lists it with every other setting.