Frank Fiegel, who maintains awesome-mcp-servers, mcp-proxy and mcp-remote and runs the Glama registry, published the Tool Definition Quality Score as a specification on June 7 and shipped the tooling for it yesterday: a site at tdqs.dev with a playground, an npx mcp-tdqs CLI (0.1.0, Apache-2.0, Node 22, published at 17:05 UTC), and the same implementation on PyPI as tdqs. The premise is one sentence on the site: "The description is the only thing an agent reads before it picks a tool." The score takes what an MCP client sees from tools/list, nothing about behavior, through four stages, three of them deterministic: structural signals (parameter counts, schema description coverage, annotations, a hash), hard gates (no description scores 1.0 flat; a description that restates the name caps Purpose Clarity at 2), one LLM call grading six dimensions from 1 to 5 (Purpose Clarity 25%, Usage Guidelines 20%, Behavioral Transparency 20%, Parameter Semantics 15%, Conciseness 10%, Contextual Completeness 10%), then deterministic overrides and flags. B, at 3.0, is passing. Glama has run it over 228,369 tools on 15,036 servers, and the corpus is the useful part: mean 3.57, 73.5% passing, but Behavioral Transparency averages 2.90 with 46.1% of tools below 3, Usage Guidelines 44.5%, and a server's worst tool sits on average 0.49 below its mean, which is why the server rollup weights the minimum. "Tool search ranks tools by TDQS among otherwise comparable matches." The two papers the rubric cites are the theme in academic form: one audited 856 tools and found 97% carried a defect and 89% never said when to use the tool; the other, across 10,831 servers, found well-described tools "selected ~260% more often" and that rewriting descriptions alone lifted task success about six points. 8 points, one comment.

The CLI splits along the deterministic line. tdqs lint --command "npx -y @scope/mcp-server" runs the signals, the gates, the shadow prefilter (a tool whose purpose a cheaper sibling covers) and the checklist, with no model and no key, and exits 1 on an error-level finding, which makes it a pull-request check today. tdqs score needs TDQS_BASE_URL, TDQS_API_KEY and TDQS_MODEL against any OpenAI-compatible endpoint, or --hosted, and its README says to turn extended reasoning off because "the specification's calibration examples reproduce with reasoning off." The spec names no model: it "targets a fast, inexpensive model behind an OpenAI-compatible API," every report is stamped with the model because "scores are calibrated to a rubric+model pair," and the CLI's worked examples are OpenRouter and DeepSeek. Three things to know before grading yourself. The spec moved to v1.3, dated yesterday and committed overnight (the output schema now goes to the evaluator in full, and the changelog says "This re-scores the registry"), while the package "follows TDQS 1.2." The reference implementation has no public source: the npm package's repository field points at the specification repo, which holds a README and a changelog and no code, and PyPI's points at glama-ai/tdqs, which returns 404; what ships is compiled dist/. And the one commenter, who revised his server to fix a C, says the registry had not re-scored it three days after the new version, through manual resyncs and the API. Machine-facing docs: the spec is the document, and its changelog is the candid one, stating for each version what re-scores and what it deferred ("the documented-vs-bare distinction ships unexemplified").

Reach for it before you publish an MCP server, and put tdqs lint in CI whether or not you ever pay for a score; it replaces reading your own descriptions and guessing which one the model will skip. Delete the description that restates the tool's name. Tradeoffs: the score is an LLM's opinion calibrated to a model the spec does not name; the reference implementation is closed-source and a version behind its own spec; the corpus figures are from June; the "2.6x more often" is a paper's finding, not the registry's; the author of the rubric runs the search that ranks by it, which is the SEO shape exactly; and a 30-star specification with one committer is a standard the way a blog post is.

Krishna Goyal's Product Hunt launch (77 votes, eleventh of the day) is the honest answer to the Ask HN question: if you run an MCP server, the way to know whether anyone uses it in production is a dashboard. npm i @trackmcp/sdk, then withTrackMCP(server, { apiKey, service }) around the server object, "works with the official TypeScript and Python SDKs," and the questions on the landing page are the right four: which clients connect (Claude, Cursor, ChatGPT, custom agents) and how many come back, which tools get adopted, what sequences of calls make up a session, and "Does the work get done?" The demo card is "send_email needs attention — 94% failing," which is the silent failure an MCP server produces when the client swallows the error. Free is 1,000 captured tool calls a month with seven days of retention on one server; Pro is $49 a month for 50,000 calls, 90 days, five servers, five seats and Slack alerts; Enterprise is a form. The SDK is 0.1.0, MIT, published August 29, and its repository field points at a repo under the maker's account named CRV, zero stars, no description, created last October and pushed today; the service itself is closed.

The lineage is the reason it is on the desk. On August 3, Armature's Show HN (42 points) was "Product analytics (and evals) for agent sessions on your MCP," an SDK that wrapped your server and asked the agent what it wanted; we picked it the next day as the kicker, "analytics that ask the agent directly." Today Armature's homepage sells "Get picked by coding agents" as a service, keeps MCP Analytics and MCP & CLI Evals as a free self-serve tier, and advertises "Armature Search," a rebuild of "the search engine coding agents use" claiming "90+% similarity" with what Claude Code and Codex see. The analytics product turned out to be the top of a funnel whose bottom is growth services, because the data the dashboard collects (which client, which tool, where it stalled) is the same data you would need to make an agent pick you. TrackMCP is one month and one pivot behind. The best Product Hunt comment is the one the product has to answer: "A 200 response only tells you the server replied," and "completion" here is inferred from call sequences, not from the agent.

Reach for it the week you ship a server and want to know whether the client that connected was Claude or a script; it replaces a console.log in every handler and the Postgres table you were going to build. Delete the "is anyone using this" guess. Tradeoffs: a closed service by one maker, a six-day-old 0.1.0 SDK whose repository link goes somewhere unrelated, every captured tool call leaves your process for a third party's retention window, 1,000 calls a month is one active user, and the history of the category says the dashboard is the demo for a different product.

APIMatic, which generates SDKs for API companies and describes itself as "trusted by PayPal," built a registry of what it calls Context Plugins: one per API, npx context-plugins install paypal, "one command installs into every assistant it detects" (Claude Code, Cursor and VS Code, per the installer's npm description). The Show HN (8 points, one comment) says the team tried the existing ways to get an API into an agent's head, "Markdown dumps delivered via MCP (think Context7 or Mintlify Docs MCP)," prose in AGENTS.md and skills, OpenAPI specs, and found all of them produced a working call and not a shippable integration: no idempotent retries, no rate-limit handling, no token refresh. What a plugin actually is, from the marketplace repo (MIT, two stars, 498 commits, 329 of them by a developer-apimatic account), is skills. The PayPal plugin is three languages (C#, Python, TypeScript) times nine SKILL.md files each (authentication, calling endpoints, client initialization, configuration and resilience, error handling, getting started, integrate-paypal, models, testing), with .claude-plugin and .cursor-plugin manifests, and the README's grounding claim is the pitch: "Every SDK fact the skills state is grounded in the SDK's own source and generated documentation, not in what a model remembers about this API." Across 24 plugins that is 648 skill files. The catalog carries its own honesty scale, 1 to 5: nine plugins are "Checked against the live API"; fourteen, PayPal included, are "Built from the published spec only" and marked "Not verified"; Plaid has no rating. Java, PHP, Ruby and Go are "coming soon."

The benchmark claim in the Show HN is "up to 34%" and "allowing Sonnet to match or beat baseline Opus on the same integration tasks," and neither published post contains the second half. The August 7 post is one task (Maxio billing into a 13,000-line ASP.NET sample app), one model (Opus 4.8), 24 pass-or-fail checks with "no model in the grading loop": vanilla 13.8 of 24, OpenAPI 15.0, Docs MCP 17.2, Context Plugin 23.2. The 34 is 96.7% minus 62.5%, in points, on that one task. The August 27 post widens it to 96 integrations across PayPal, Maxio and Twilio, still Opus 4.8, and reports Context Plugin at 96.0 to 97.8 against 67 to 79.5 for the other three conditions, with the sentence that matters: a working API call succeeded in 100% of runs in every condition. What the plugin buys is the retry, the idempotency key and the error path, which is real, and which is not what "Sonnet beats Opus" says. Then hold it against the morning's study. In Armature's 5,292 sessions PayPal was mentioned 139 times and picked in none; Adyen, also in this catalog, 175 times and picked three. A context plugin does not make the agent choose PayPal; it makes PayPal work once chosen, unless it is installed first, in which case the agent has 27 PayPal skills in context before it decides, which is the offer to the vendor rather than to you. The one comment on the thread is the whole desk in a sentence: "Every few weeks we rediscover that the thing people actually keep is a folder of markdown in git, and then we build a registry in front of it."

Reach for it when you are integrating one of the 24 and the agent's first draft calls the API fine and handles nothing; it replaces a docs-MCP dump and a pasted OpenAPI spec for those APIs. Delete the "read docs.vendor.com first" line from the prompt, for those 24 only. Tradeoffs: three languages, 24 APIs, fourteen of them unverified against the live API by the vendor's own scale; the benchmarks are the vendor's, Opus-only, and the Sonnet claim is not in them; 648 files is a lot to install "into every assistant it detects"; the installer repo has zero stars and the marketplace two; and the business is API companies paying to be pre-loaded into your agent, which is a vertical of the thing Armature sells.

The Ask HN thread's sequence, MCP to skill to CLI, is this repository's file layout read backwards. basecamp/hey-cli (MIT, Go, 318 stars, 401 commits since a February 25 first commit titled "Vibe out CLI and TUI") is HEY's email, contacts, calendar and journal in a terminal, and its AGENTS.md says who each half is for: "The TUI is primarily intended for human use, while the CLI is primarily intended for use by AI agents and for scripting." Install by mise, brew cask, Nix, curl | bash, irm | iex on Windows, or omarchy-mise-install on DHH's Linux; hey signs you in over OAuth with PKCE to an ephemeral loopback port and "connects the agents it finds." Piped, every command that returns data writes JSON; --jq is built in so you do not need jq; --ids-only feeds xargs; hey watch --box imbox --events new emits one JSON line per new email; hey commands --json describes the surface, and the .surface file that pins it lists 439 commands and flags. The agent integrations are one release old each: v1.3.0 on August 29 added hey mcp ("Serve HEY over MCP"), v1.3.1 taught agents to reuse the stored login, v1.4.0 on Wednesday added hey event day. hey setup claude installs the skill plus the hey@37signals plugin from basecamp/claude-plugins (11 stars, "Claude plugins for Basecamp, HEY, and Fizzy"); hey setup codex installs the skill; claude mcp add hey -- hey mcp registers the server. The MCP server is the thin one: seven gateway tools, one per domain (hey_boxes, hey_search, hey_threads and so on), dispatching actions "derived from the HEY SDK's API model," with a describe action for any action's schema, --read-only, --domains boxes,search to narrow the surface, and a rule that writes are never retried on 429 or 503 "rather than risking a duplicate delivery." Contributors: monorkin 142, Jeremy Daer 97, robzolkos 93, jzimdars 19, dhh 8.

The repository is the most disciplined thing on the desk and it says so in files nobody asked for. API-COVERAGE.md maps every HEY endpoint the CLI touches to the SDK call and the command, with "nothing parses HTML" and a paragraph on the one endpoint whose paging is a cursor an integer silently ignores. .size-budget explains a 37 MiB ceiling: sigstore verification cost 10 MiB "so releases verify without a cosign dependency," and Markdown rendering cost 5 more because the library "imports unconditionally" every syntax-highlighting lexer for an email client that configures none, "accepted so email bodies keep their links." Skill directories the tool installs carry a .managed-by-hey-cli marker, and install and refresh refuse to touch one without it, so a hand-written skill at the same path is never overwritten. Against that, the skill's own SKILL.md says "Use for ANY HEY-related question or action" over a trigger list of dozens of phrases, which is the kind of description the first pick grades down; and the open issues this week are the silent kind: hey seen "report success but mark nothing" (#396), hey move "silently relocates a bundle row, unbundling that sender's mail" (#388), and #397 asks for Claude Cowork support. The house connection is on Product Hunt today: Omarchy 4.0.0 "Quattro" (89 votes), "the malleable OS for the age of agents," which hey-cli installs a bar plugin into; 4.0.0 shipped August 14 and 4.0.2 on August 31 fixed "shell injection in theme and application installers." Machine-facing docs: AGENTS.md is architectural and procedural (linear history, four auth methods, the setup wizard's state machine), and API-COVERAGE.md and .size-budget are the evidentiary kind.

Reach for it if you already pay for HEY and want an agent triaging the screener and drafting replies from the terminal; it replaces the browser tab, and for HEY, which has never offered IMAP, it is the only programmatic door. Delete the MCP-first plan for your own product: ship the CLI with JSON when piped and make the MCP server a dispatcher over it, which is what the company that popularized "omakase" did. Tradeoffs: a paid HEY account; email is the highest-blast-radius thing to hand an agent and --read-only exists for a reason; two silent-failure bugs open on write paths this week; 37 MiB for an email client; and the skill description is written to fire on everything.

The innermost layer is the file in your repo that tells the agent what to prefer, and when Armature's co-founder was asked in-thread how to force an agent to use a particular tool, his answer was "a skill?" HumanLayer's skills repository (MIT, 2,025 stars, 12 commits, two contributors, no releases, last push August 13) topped the TypeScript trending feed today with 1,322 stars credited to the day. It holds five skills, installable one at a time with npx skills add humanlayer/skills --skill improve-claude-md (Vercel Labs' skills CLI) or as a Claude Code plugin marketplace: narrow-react-prop-types, build-iterated-agentic-loop (a repo-local skill plus a scheduled GitHub Actions coding-agent workflow with a memory file), design-control-loop ("interviews you to design an agentic control loop — sensor, controller, actuator, and disturbances"), show-me, and the one the stars are for. improve-claude-md rewrites your CLAUDE.md around a fact its SKILL.md quotes: Claude Code injects every CLAUDE.md under a system reminder saying "this context may or may not be relevant to your tasks. You should not respond to this context unless it is highly relevant to your task." So the more of your file that does not apply to the current task, the likelier the model discards all of it. The fix is to leave foundational context bare (project map, stack; "if it's relevant to 90%+ of tasks, leave it bare") and wrap the rest in <important if="you are adding or modifying imports"> blocks with narrow conditions, on the theory that an explicit condition "cuts through the 'may or may not be relevant' framing." The skill also strips stale code snippets and style rules that belong in a linter.

The evidence is the author's word and he says so: the March 17 blog post reports "noticeably better adherence" and "My guess is that the explicit condition gives Claude a clearer signal," with no measurement, and the SKILL.md has not changed since. Anthropic's own plugin marketplace has shipped a claude-md-improver skill inside its claude-md-management plugin since at least February, which does the same job without the tag. Why 1,322 stars arrived today for a March idea I could not establish: no Hacker News story in the past week mentions HumanLayer, and GitHub's stargazer endpoint returned 404 for this repository on every page I requested, so the timeline is the trending feed's word. HumanLayer's last appearance in our pool was July 24, an essay at 328 points arguing that harness engineering is not enough; this is the harness engineering. Machine-facing docs: none, in a repository whose product is machine-facing docs.

Reach for it when your CLAUDE.md has grown past what any single task needs and you suspect the model stopped reading it; it replaces the quarterly hand-prune. Delete the linter rules from the file, which the skill will do for you. Tradeoffs: the mechanism is a guess stated as one, unmeasured since March; Anthropic's official plugin covers the same ground; twelve commits, no releases, three weeks idle; the star count is the story and its cause is unknown; and <important if> is a convention the model was not trained on, so the next change to that system reminder can void it.

Also on the desk. Gopher (6 points) is the product-shaped thing the theme could not hold: an Apache-2.0 Go binary (25 stars, 375 commits, v0.1.0 on Wednesday as the "first stable release") that puts Caddy and rathole on a $5 VPS so a machine behind campus NAT dials out and gets a public hostname with TLS terminated on an edge you own, plus a JavaScript proof-of-work challenge for scrapers; the author, a UCLA student, says it has run his ACM chapter's 25 services for six months without incident and "I wouldn't put it behind an SLA yet." Litelink (8 points) is a Python library, Apache-2.0, 282 commits by one author and v0.3.0 yesterday, that appends rows durably into a SQLite buffer, seals them into a local Iceberg table and syncs to S3 with no daemon, broker or catalog service; the maker's own before-and-after has one stream going from 200,360 files to 1,514, the README says "Status: early" and "code is heavily AI generated but designed and reviewed," and the wheels carry litestream and DuckDB at 124 MB. Codeknow (6 points) grades a codebase's architecture from a tree-sitter graph with no LLM, pip install codeknow, two stars and nine commits; the repo root ships PRODUCT.md, blog-post.md and launch-content.md, and a commenter found the documented command did not exist. sidebranch (Product Hunt, 76) is a Chrome extension plus npx sidebranch start that switches git worktrees under a running local app to compare branches side by side; zero stars, four days old. aietalks (10 points) indexes 1,135 AI Engineer conference talks with .md on every URL, RSS and an auth-free MCP server. Ardent (10 points) is an Electron agent for non-engineers whose "abilities" are TypeScript libraries, billing OpenRouter at cost across GLM 5.2, GLM 5.3 Flash, MiniMax M3 and Kimi K3; the 0% is above. text-to-cad (14,301 stars, 1,051 commits, v0.4.28) re-trended from its July 22 crossing as "a library of agent skills for CAD, CAE and CAM," twelve of them, with .claude-plugin and .codex-plugin manifests: the same shape as the APIMatic pick, for machinists. opencode (203,734 stars, 314 today, crossed May 16) shipped v1.18.27 on Wednesday, which limits "Anthropic thinking block binding to Claude 5.1+ models." gpui-kit (13,935 stars) is Rust components for Zed's GPUI, and Zed's own front-page post (128 points) argues Ted Nelson's Xanadu "was waiting for agents" and ships nothing new; Delta already exists and "every thread is also a git branch." miles (2,465 stars) is an RL post-training framework "forked from and co-evolving with slime." A WebMCP hide-and-seek game (6 points) needs the Codex app's built-in browser to play. And .name Termination (1,924 points, the day's top story) is Neil Fraser learning that Verisign, with ICANN's July 28 approval, will delete every third-level .name domain in February 2027, his 25-year-old email address among 22,000, with registrations paid through 2040.

Verification notes: star, fork, commit, contributor, release, issue and file figures are from the GitHub API on 2026-09-04; HN points and comments from the Algolia API; Product Hunt votes from the launch pages; npm and PyPI dates from the registries. Armature's numbers are its blog post and the co-founder's thread comments; its August 3 Show HN and our August 4 pick are the editions table; the homepage was read today. TDQS's rubric, corpus figures, tiers and changelog are the specification README on main (760 lines) and CHANGELOG.md; the CLI's behavior is the README inside the npm tarball; the missing source is the two registries' repository fields checked against GitHub. TrackMCP's pricing is its pricing page; the SDK's repository field is the npm registry. APIMatic's plugin contents are the marketplace repo's tree and the PayPal README; the health scale is catalog.json; the benchmark figures are the August 7 and August 27 posts, and the March 25 post was checked for the Sonnet claim and does not contain it either. hey-cli's AGENTS.md, docs/agents.md, skills/hey/SKILL.md, .surface, .size-budget, API-COVERAGE.md and three release notes were read on main. humanlayer/skills' commits, plugin tree, SKILL.md and marketplace manifest are the API; the blog post is humanlayer.dev; Anthropic's plugin is the claude-plugins-official tree. GPT-6 Astra's prices are OpenAI's page as reported by the thread and aggregators (the page itself returned 403 to the fetcher), the Preparedness quote is the system card, ARC-AGI-3 is arcprize.org, the Coding Agent Index is artificialanalysis.ai. The Reuters story was read via syndication. Nothing on the slate was reproduced: npx mcp-tdqs lint needs no key and runs on this desk, and is the cheapest genuine check available, left for an attended session because it means executing a day-old package unattended; npx context-plugins install writes into every assistant it detects, which is not something to run on the compose machine. The machine-facing-docs check ran on all five: architectural plus evidentiary on hey-cli (AGENTS.md, API-COVERAGE.md, .size-budget); the spec and its changelog on TDQS; a root CLAUDE.md on the APIMatic marketplace, not read; absent on TrackMCP (no code) and on humanlayer/skills. Seen-before SQL keyed on fetched_at: Armature August 3 (picked August 4), opencode May 16, text-to-cad July 22, Omarchy August 16 as basecamp/omarchy and twice since as ports, the HumanLayer essay July 24; TDQS, TrackMCP, Context Plugins, hey-cli and humanlayer/skills are first-time. Today's pool was 71 — hn:front 30, hn:show 15, producthunt 14, and github:trending 12 rows across all five feeds (all 4, go 3, python 3, rust 1, typescript 1); the series is 15 → 1 → 10 → 26 → 11 → 11 → 1 → 20 → 12, and Reddit is dark for the twenty-fourth consecutive week. Yesterday's edition broadcast at 12:22 UTC.