Data refreshed 18 November 2026, 13:49 UTC
WebMCP: AI agent tools for this site
Top100 is an ordinary static website — and a WebMCP client-side MCP server for the same content. An AI agent running in the browser can discover 16 tools here and call them directly instead of scraping the tables, with the results the page itself would have produced.
- 16 WebMCP tools
- 14 read-only
- 8 return GitHub text
- 3,800 repositories reachable
- Same-origin only, no cookies
Try the tools in this tab
The fastest way to see WebMCP working is the debug panel on this site: it lists every registered tool with its JSON Schema and lets you call it by hand, through exactly the same validation, rate limiting and output budget an agent's call goes through.
To drive the tools with a real agent, enable WebMCP in your browser:
- Chrome:
chrome://flags/#enable-webmcp-testing→ Enabled, then relaunch. Some builds instead run the origin trial; the site works with either. - Agent extensions: the Model Context Tool Inspector lists the tools and calls them with a prompt.
- CLI agents: agent-browser
announces WebMCP tools automatically and can invoke them with
agent-browser webmcp invoke. - No browser support? This site ships its own polyfill, so the tools register anyway — see the polyfill section below.
Security model
WebMCP tools are code the page runs on the agent's behalf, so the interesting questions are what a tool can reach, what it can change, and what a hostile string could do once it is inside a model's context. The answers here are deliberately boring.
| Risk | Mitigation on this site |
|---|---|
| Cross-origin reach | Tools are registered without exposedTo, so only
this origin can see or call them. Registration is skipped entirely inside a cross-origin
iframe unless the embedder granted allow="tools". |
| Malformed or hostile arguments | Every call is validated against the tool's JSON Schema before it runs: types, ranges, enum membership, 200-character string caps, array length caps, four levels of nesting. Unknown arguments are ignored and reported back. |
| Prototype pollution | __proto__, prototype and
constructor are rejected anywhere in an argument tree, and validated values are
copied into a fresh object. |
| Runaway agents | 90 calls per minute per document, 30 per tool, a 100 ms
minimum gap, one in-flight call per tool, and an 8-second timeout (15 s for the three
tools that read a dataset file) wired to AbortSignal. |
| Context flooding | Results are capped at 1500 characters. Array payloads are shrunk until the envelope fits and the result says exactly how many items were dropped. |
| Prompt injection from repository text | Descriptions, topics and names come
from GitHub, so any tool that returns them declares
untrustedContentHint: true and every result carries
"trust": "untrusted" with a note. Control characters are stripped and every
free-text field is length-capped. Nothing on this site asks an agent to run a command,
reveal a secret or trust a suggestion. |
| Data exfiltration | The only network calls are same-origin GETs of files under
data/, with credentials omitted, redirects refused and a 4 MB ceiling. No
cookies, no analytics, no third-party requests, no sendBeacon. |
| Injection into the page | Tool results and tool names are rendered with
textContent; innerHTML and eval are never used
anywhere in the WebMCP code, and the build fails a test if they appear. |
| Silent automation | Tools act on the visible page and report what they did:
a live region announces every call, and the tools that change the page
(filter_repository_table, open_page) are the only two that are not
read-only. Both are annotated accordingly. |
Every tool
Annotations follow the WebMCP annotations object:
readOnlyHint means the tool does not change anything,
untrustedContentHint means its result contains third-party text, and
consequentialHint — never set here, because nothing on this site books,
buys or deletes anything — would ask for user confirmation.
Discover the site
get_site_info
Describe the Top100 site: what it publishes, how fresh the data is, how many languages and repositories it ranks, where the raw JSONL lives, and which pages exist. Call this first when you need to know what this site can answer or how current its numbers are.
read-only · available on: all
No arguments.
get_page_context
Return where the user is right now: page type, canonical URL, the language this page covers (if any), how many repository rows are on the page, and the timestamp of the data being shown. Use it to resolve deictic prompts such as "this list" or "the table here".
read-only, untrusted content · available on: all
No arguments.
list_languages
List the languages Top100 ranks, with each one's most-starred repository, the star count to reach 100th place, the median inside the top 100 and how many repositories GitHub reports for it. Paged: five rows per call, use offset to page through the full language list. Every row's page URL is base_url + /languages/<slug>.html.
read-only, untrusted content · available on: all
sortstring one ofrank,leader_stars,entry_bar,namedefaultrankrank (site order), leader_stars, entry_bar or name.limitinteger default5Rows to return, 1-30. Default 5.offsetinteger default0Row to start from, 0-based, for paging.
get_language_stats
Statistics for one language's top 100: the leading repository and its stars, the entry bar for 100th place, the median star count, the most common licence and how many of the 100 use it, the JSONL download with its record count and SHA-256 checksum, and a one-line description of the list. Accepts a language key such as python, c++, c# or objective-c.
read-only, untrusted content · available on: all
languagestring requiredLanguage key, for example python, rust, c++, c#.
get_download_links
List every file this site publishes for machine use, with its URL: the JSONL dataset (one file per language), the Markdown lists, feed.xml, sitemap.xml, robots.txt, llms.txt, llms-full.txt, README.md, LISTS.md, the search index and the WebMCP manifest that carries every dataset checksum. Pass language=<key> for one file's size and SHA-256; pass family=dataset for the per-language URLs.
read-only · available on: all
languagestringLanguage key for one dataset file, for example python.familystring one ofall,site,datasetdefaultallWhich family of files to return.
Query the data
get_top_repositories
Return a ranked slice of the repository table on the current page: rank, name, stars, forks, licence, last push and a short description. On a language page this is that language's top 100; on the homepage it is the overall top 10. Five rows by default, up to 25 (larger requests are trimmed to the result budget and say so). Use get_repository for the full record of one row.
read-only, untrusted content · available on: repos
limitinteger default5Rows to return, 1-25.offsetinteger default0Rank to start from, 0-based.sortstring one ofstars,stars_asc,forks,issues,recent,namedefaultstarsstars (default), stars_asc, forks, issues, recent or name.licensestringKeep only this SPDX licence id, for example MIT or Apache-2.0.min_starsintegerKeep only repositories with at least this many stars.
search_repositories
Search the full repository dataset by keyword (matched against name, description and topics) with optional filters for language, star range, SPDX licence and topic. Returns ranked matches with stars, forks, licence and page links. Sorting is deterministic: relevance, stars, forks, issues or recency.
read-only, untrusted content · available on: repos, home, languages
querystring defaultKeywords, for example "design system" or "llm agents".languagestringRestrict to one language key, for example python or rust.min_starsintegerMinimum stars.max_starsintegerMaximum stars.licensestringExact SPDX licence id, for example MIT.topicstringExact topic tag, for example cli or machine-learning.sortstring one ofrelevance,stars,forks,issues,recent,namedefaultrelevancerelevance (default), stars, forks, issues, recent or name.limitinteger default5Matches to return, 1-20.
get_repository
Look up one repository by its full name (owner/name), or by its rank within a language's top 100. Returns stars, forks, open issues, licence, last push date, topics and description, plus links to its GitHub page and to the language list it appears in.
read-only, untrusted content · available on: repos, home
full_namestringowner/name, for example torvalds/linux.languagestringLanguage key, required together with rank.rankintegerPosition in that language's top 100, 1-100.
compare_repositories
Compare two to five repositories side by side: stars, forks, open issues, stars per fork, licence, last push and shared topics. Accepts owner/name strings and returns them in the order given, with a one-line verdict naming the leader by stars.
read-only, untrusted content · available on: repos, home
repositoriesarray requiredTwo to five repositories as owner/name.
Act on the page
filter_repository_table
Act on the visible page: filter the repository table by keyword, licence, topic or minimum stars, then scroll it into view. Use reset to show every row again. Returns how many rows matched, and the page visibly updates so the human sees exactly what was filtered.
untrusted content · available on: table
querystring defaultMatch repository name, description or topics.licensestringKeep only this SPDX licence id.topicstringKeep only rows carrying this topic.min_starsintegerKeep only rows with at least this many stars.resetboolean defaultFalseClear every filter and restore all rows.
open_page
Navigate the current tab to one of this site's own pages: home, the language index, a specific language list, methodology, dataset reference, comparison, FAQ or the WebMCP tools page. Only same-site pages can be opened; the user sees the navigation happen.
no annotations · available on: all
pagestring required one ofhome,languages,language,methodology,dataset,compare,faq,webmcpWhich page to open.languagestringLanguage key, required when page is "language".
Read the documentation
get_methodology
Explain how the data is produced: the GitHub Search API query, the hourly schedule, rate-limit handling, the fields the Search API omits, and the known limitations. Called without arguments it returns the section ids, their titles and the overview; call it again with one of those ids to read that section in full.
read-only · available on: all
topicstringSection id from a previous call, for example limitations.
get_dataset_schema
Document the JSONL dataset: one file per language, one complete GitHub Search API repository object per line. Called without arguments it lists the field names and types; call it with field=<name> to read what one field means, or with language=<key> to get that file's download URL, size and SHA-256 checksum.
read-only · available on: all
fieldstringA field name such as stargazers_count or license.languagestringReturn the download details for this language only.
get_comparison
Explain how Top100 differs from other GitHub ranking sources — GitHub Trending, gitstar-ranking.com, EvanLi/Github-Ranking, OSS Insight, Star History, GH Archive and BigQuery — including where each one is the better tool. Called without arguments it lists the sources; pass one to read its full comparison.
read-only · available on: all
sourcestringOne source name to compare against, for example GitHub Trending.
answer_question
Answer a question about this project from the site's own FAQ: refresh cadence, licensing, whether it needs an API key, how to download the data, how to add a language, why some fields are missing. Returns the closest matches with their answers, ranked deterministically.
read-only · available on: all
questionstring requiredThe user's question in natural language.max_resultsinteger default2Matches to return, 1-5.
Diagnostics
run_webmcp_self_test
Run this page's own WebMCP checks: page data present, tools registered, character budgets respected, prototype pollution rejected, output budget enforced, and every registered tool answering a call with a structured envelope. Intended for developers and test harnesses.
read-only, debugging only · available on: all
No arguments.
How the runtime works
assets/webmcp-core.js does three things, in this order:
- Use what the browser has. If
document.modelContextexists (or the oldernavigator.modelContextfrom the origin-trial builds), the tools are registered there, unchanged. - Otherwise install a polyfill. A dependency-free implementation of the same
surface —
registerTool,getTools,executeTool,provideContext, thetoolchange,toolactivatedandtoolcancelevents, and unregistration through anAbortSignal— is installed ondocument.modelContext. It never replaces a real implementation, and it refuses to install on an insecure origin or in a cross-origin frame without permission. - Wrap every tool in the same guard. The validation, rate limiting, timeouts,
output budget and envelope formatting live inside each tool's
execute(), so an agent that calls the nativeexecuteToolgets exactly the same guarantees as one that calls the polyfill.
Two URL parameters exist for testing, and neither is read by anything else:
?webmcp_debug=1 opens the debug panel, and
?webmcp_polyfill=1 forces the polyfill even where the browser has native
support.
Tools are registered per page. A language page registers the tools that make sense for its 100 rows; the FAQ page registers the documentation tools; nothing registers a tool it cannot answer. The manifest of what is available where is published as data/webmcp-manifest.json.
For agents that never open the page
WebMCP's own discoverability limit is that a client has to visit a page to know what it offers. Two static artifacts cover the gap:
data/webmcp-manifest.json— every tool with its full JSON Schema and annotations, the runtime facts, the security model, and SHA-256 checksums for all 38 JSONL files.llms.txtandllms-full.txt— the same site described for agents that read text instead of calling tools.
| Language | File | Records | Size | SHA-256 |
|---|---|---|---|---|
| TypeScript | data/JSONL/typescript.jsonl | 100 | 560.1 KB | 1db48b7dd8ce7380… |
| Python | data/JSONL/python.jsonl | 100 | 589.0 KB | a5525902c812f944… |
| JavaScript | data/JSONL/javascript.jsonl | 100 | 564.9 KB | 2061e567139461fc… |
| Java | data/JSONL/java.jsonl | 100 | 563.1 KB | 734263fb7258c68e… |
| C# | data/JSONL/c#.jsonl | 100 | 567.6 KB | 60821a91b0d6f268… |
| C++ | data/JSONL/c++.jsonl | 100 | 556.3 KB | 26a84586a623a8ab… |
| PHP | data/JSONL/php.jsonl | 100 | 554.1 KB | 519959709aeca4f5… |
| Shell | data/JSONL/shell.jsonl | 100 | 562.6 KB | a3e5fe36df37b632… |
| C | data/JSONL/c.jsonl | 100 | 546.3 KB | f1b2e4723cced173… |
| Go | data/JSONL/go.jsonl | 100 | 548.6 KB | 9d7fcfd04e60ea4a… |
| Rust | data/JSONL/rust.jsonl | 100 | 547.6 KB | 2da87b898fc23f09… |
| Ruby | data/JSONL/ruby.jsonl | 100 | 546.9 KB | 19568671a4b242ba… |
| Kotlin | data/JSONL/kotlin.jsonl | 100 | 556.3 KB | edf2ba93cca36a29… |
| Swift | data/JSONL/swift.jsonl | 100 | 557.3 KB | 46b67e088559552a… |
| Dart | data/JSONL/dart.jsonl | 100 | 568.6 KB | 61bb9ca140b0895e… |
| HTML | data/JSONL/html.jsonl | 100 | 574.6 KB | c75c4e33f23bf6da… |
| CSS | data/JSONL/css.jsonl | 100 | 562.9 KB | fe5b010273481340… |
| SQL | data/JSONL/sql.jsonl | 100 | 588.0 KB | af891d9e070310d5… |
| Scala | data/JSONL/scala.jsonl | 100 | 549.9 KB | bb6bd17e91b2d4db… |
| Lua | data/JSONL/lua.jsonl | 100 | 555.7 KB | b4dcf5d811b7496a… |
| Groovy | data/JSONL/groovy.jsonl | 100 | 577.3 KB | 00159e3d31ed465d… |
| Objective-C | data/JSONL/objective-c.jsonl | 100 | 559.8 KB | e4df8fa3a7a64814… |
| Perl | data/JSONL/perl.jsonl | 100 | 549.9 KB | b818738e9a9659ab… |
| Haskell | data/JSONL/haskell.jsonl | 100 | 546.1 KB | 55682a7cf683b279… |
| Assembly | data/JSONL/assembly.jsonl | 100 | 709.7 KB | c8aadad0e43ed7b5… |
| R | data/JSONL/r.jsonl | 100 | 546.0 KB | c1bc773e4603b2b4… |
| Julia | data/JSONL/julia.jsonl | 100 | 567.3 KB | 2ecb6130d389483f… |
| Elixir | data/JSONL/elixir.jsonl | 100 | 550.1 KB | f52dfca453e6e8e6… |
| PowerShell | data/JSONL/powershell.jsonl | 100 | 578.1 KB | cf833bc6c261ee79… |
| Nix | data/JSONL/nix.jsonl | 100 | 552.6 KB | 235281400af542dd… |
| Zig | data/JSONL/zig.jsonl | 100 | 536.0 KB | 08a32e0555cd1a20… |
| Clojure | data/JSONL/clojure.jsonl | 100 | 542.4 KB | f4a69f59d78bbc23… |
| Erlang | data/JSONL/erlang.jsonl | 100 | 532.1 KB | ee23b4ce91447d35… |
| OCaml | data/JSONL/ocaml.jsonl | 100 | 546.4 KB | e587ef7202a3bebd… |
| Solidity | data/JSONL/solidity.jsonl | 100 | 567.4 KB | cc8929103d70eea6… |
| Fortran | data/JSONL/fortran.jsonl | 100 | 550.9 KB | 2eea33c565924f88… |
| Vim Script | data/JSONL/vim script.jsonl | 100 | 545.0 KB | 8f6724cf50ed3d1e… |
| F# | data/JSONL/f#.jsonl | 100 | 558.1 KB | e01913668bf5058c… |
Limitations, stated plainly
- WebMCP is an early standard. The API shape is still moving; this site supports both
document.modelContextand the earliernavigator.modelContext, and degrades to its own polyfill when neither exists. - A page must be open for its tools to exist. Nothing here works headless unless the HTML is actually loaded in a browser.
- Tool results are for one page's worth of context. Deep analysis across all 3,800 records belongs in the JSONL files, not in a tool result.
readOnlyHintand friends are declarations, not enforcement. They are accurate here, but a client should treat every page's claims as untrusted — as OpenAI's and agent-browser's own documentation says.- Tools read the dataset snapshot embedded in the page or committed to the repository. Between refreshes, they are as stale as the timestamp at the top of every page.
Questions about the tools
Do I need a browser extension?
No. Any browser with WebMCP support exposes these tools to whatever agent is running there; the extension is only a convenient way to test.
Does this change how the site works for people?
No. Every page is complete HTML before any script runs, the WebMCP code adds a small collapsible panel and nothing else, and with JavaScript off the site is byte-for-byte the site it was.
Is anything tracked?
No cookies, no analytics, no third-party requests. The only network traffic the tools generate is this site fetching its own data files.