Sighttrue MCPサーバー

Sighttrue

sighttrue.com · your stack · compare · method · feed

Check your dependencies in CI

permissions:
  pull-requests: write   # only needed for the comment

steps:
  - uses: actions/checkout@v4
  - uses: sighttrue/sighttrue@v1
    with:
      manifest: package.json
      fail-on: withdrawn,archived,relicensed

Fails the build when a dependency's own publisher has withdrawn it — npm calls it deprecated, PyPI and crates.io call it yanked, RubyGems yanked, Packagist abandoned, NuGet deprecated — or when its repository has been archived or moved to a source-available licence, and reports advisories either way. Reads package.json, requirements.txt, pyproject.toml, Cargo.toml, composer.json and Gemfile. No key, no account, no service to sign up to — it reads published measurements and talks to OSV. A network problem is never a finding: if the readings cannot be fetched the step says so and passes, because a build that breaks when somebody else's site is down is a build that gets deleted.

On a pull request it also leaves a comment with the findings as a table, and rewrites that same comment on every push rather than adding another. A red build tells a reviewer that something is wrong; it does not tell them which dependency, or what changed about it, without opening the log. Set comment: false to report only in the step summary. A pull request from a fork gets a read-only token, so the comment is skipped there and the step still passes.

Also runs against requirements.txt and Cargo.toml.

Or install the App and change nothing

The Action needs a workflow file in every repository that wants it. The GitHub App needs one install and then watches every repository it is given, with no workflow, no key and no configuration: when a pull request adds a runtime dependency that is on the watchlist, it leaves one comment with what is on record — downloads, OpenSSF scorecard, advisory count, licence, last push — and edits that same comment on every later push.

It says nothing at all when a pull request changes no manifest, adds no dependency, or adds nothing that is tracked. That is the common case and it is the point: a bot that comments to say it found nothing gets uninstalled. It never calls a package safe, unsafe or recommended, because none of those are things this project measures.

The Action fails a build; the App only reports. Both read the same published bundle, so they cannot disagree.

An SBOM you are allowed to hand over

/stack exports the readout as a CycloneDX 1.6 document, with each dependency's advisory count, scorecard, licence, publish date and bus factor attached as properties. Built in the browser from the pasted manifest, like everything else on that page — nothing is uploaded, and there is no endpoint to send a manifest to.

It states no versions. A manifest declares ^1.6.0; it does not say what was installed, and it says nothing about the transitive tree underneath. The document says both of those things in its own metadata rather than leaving a reader to notice, because an SBOM that guessed would be worse than none: a compliance artefact asserting something nobody resolved.

Suggest something to watch

The watchlist is curated and partial, and it was chosen by hand with no method beyond judgement — its weakest point. Open an issue to argue for a repository. Every change to the list is a reviewed commit, so that issue is where the review happens.

For coding agents

An MCP server over the same readings, so an agent can answer "is this dependency healthy" from a measurement taken today rather than from training data a year old.

{ "mcpServers": { "readout": { "url": "https://sighttrue.com/api/mcp" } } }

given a package it returns only what a reviewer would be annoyed to discover afterwards — a publisher that withdrew it, scripts it runs on the installing machine, an archived repository, advisories, a source-available licence, a long silence — each with the address of the body that published it. It does not say whether to install. It says what is on record so the decision is made knowing it, and an empty result says plainly that it means nothing was found rather than that the package is safe.

Then check_package, check_stack, search_repositories, compare_repositories, find_model, check_eol, check_provider, list_readings. Read-only, no key, no account. Every result carries the limits of what it can support, because an agent will paste these figures into a code review and a scorecard quoted without "measures declared practices, not whether the project is safe" is a claim this project does not make.

Without MCP, one GET returns the same thing:

curl 'https://sighttrue.com/api/verdict?pkg=npm:axios'
curl 'https://sighttrue.com/api/verdict?pkg=pypi:django&version=4.2'

Advisories, licence, when the registry last shipped it, the bus factor, and whether the release is still supported — each with the address of the body that published it, so a figure quoted into a review can be checked in one click. Despite the name there is no verdict in it: no score, no rank, no recommendation. Those would be this project's judgement of somebody else's work wearing the costume of a measurement.

eleven readings, most of which never touch GitHub: 807 provider incidents kept after their own status pages dropped them, 522 release lines on the end-of-life clock, 382 model prices, 182 packages by real ship date, 416 commit histories for the bus factor. It runs every four hours on GitHub Actions, commits what it reads to this repository, and publishes a static site.

Every page and every bundle is a file. There are five dynamic routes and no page depends on any of them: /api/ask, /api/mcp, /api/chain, /api/github/webhook for the App, and the sign-in and watchlist endpoints under /api/auth and /api/watchlist. Signing in with GitHub saves a watchlist and nothing else — the readings on a signed-in page are the same published bundle anyone can download.

The commit history of data/ is the point. It is an audit trail: every reading is timestamped, append-only, and checkable against GitHub directly.

What it reports

Eleven readings, each named by the question it answers rather than by the collector that produces it. The same list drives the site's navigation, so the two cannot drift — they had, three times over, before it was written down once.

Reading	Question it answers	Touches GitHub
Live	What changed in the last few hours?	yes
Ships	What released a new version?	yes
Forks	What is being copied faster than its own baseline?	yes
Demand	What are developers asking for, across more than one project?	yes
Dependencies	What is being added, dropped, or jumped a major version?	yes
Lineage	Which models say they were built on which?	no
Model prices	What does a million tokens cost, and when did that change?	no
Outages	Does the thing I depend on go down, and how often?	no
Ecosystem	What do the registries, advisories and forums say?	no
Depended on	What does everything else quietly rely on?	no
This week	What would I have missed looking once a week?	—

The six that never touch GitHub are the answer to the fair complaint that a GitHub summariser is worth only what GitHub already shows you.

Whether each detector's threshold is reachable at all is published too, on the index under Our record. A detector nothing has ever come close to is not a quiet detector, it is a misconfigured one, and the site says so about itself.

How to read it

Every comparison is against a repository's own history, not against other repositories. "27× baseline" means 27 times what that project normally does, and the baseline is drawn on the chart so you can see what normal means.

Confidence is stated, never implied.

State	Meaning
forming	Under 14 days of history. Raw counts only. No multiplier is computed and none is implied.
detected	Crossed the threshold once. Neutral treatment.
confirmed	Persisted across two consecutive daily snapshots.

Only confirmed signals are treated as findings. That costs up to a day of speed and buys the thing that cannot be bought back.

Generated prose is set in a different typeface from measured values. Where a sentence explains a number, every figure in that sentence is checked against the source record before it is published; if any of them is not in the record, the sentence is discarded and a templated one is used instead.

What this does not claim

• The watchlist is curated and partial. It is chosen by hand and is not a survey of open source. Dependency and demand findings describe the repositories being watched, not the ecosystem.
• The data is not real-time. Four hours is the floor, and scheduled runs are routinely delayed.
• Nothing here predicts anything, and nothing here says a repository is good, bad, safe, or unsafe.
• Appearing on the watchlist is an observation, not a relationship. No project listed here has endorsed this.

Published findings that turn out to be wrong are superseded by a correction event carrying the same prominence. Events are never deleted and history is never rewritten.

The data

Every bundle the site reads is served as JSON, so any claim can be checked against the same file the page used.

data/
├── live/state.jsonl          Latest reading per repository. Sorted, rewritten each pulse.
├── live/window.jsonl         Timestamped fork samples, for the rolling 24h delta.
├── live/manifests.jsonl      Last-seen dependency set, diffed daily.
├── history/YYYY-MM-DD.jsonl  One immutable snapshot per day.
├── events/YYYY-MM.jsonl      Append-only. Never rewritten, never pruned.
├── summaries.jsonl           Generated prose, keyed by event id.
├── watchlist.jsonl           What is watched, and since when.
└── meta.json                 Last run status.

Files are sorted with a fixed key order so a repository that did not change produces no diff. That is what keeps a repository committed to six times a day from growing without bound.

Running it

Node 24. Types are stripped natively, so there is no build step and no runtime dependencies.

It said 22.18+ until the D1 schema arrived. The tests load migrations/ into node:sqlite, which is stable in 24 and behind a flag in 22 — so on 22 the suite fails on an import rather than on an assertion, which is a confusing way to learn a version requirement.

npm install          # dev dependencies only: typescript, vitest, fonts
npm test
npm run typecheck

node scripts/pulse.ts --limit=20   # needs GITHUB_PAT
node scripts/daily.ts
node scripts/build.ts              # emits dist/

Watchlist maintenance

The watchlist is curated. Deriving it from a search query was tried and produced a worse list — topic search only finds repositories that tagged themselves, and sorting by stars ranks tutorials and link collections above infrastructure. derive-watchlist.ts is kept for proposing candidates a curator might have missed, and because the negative result is worth being able to reproduce.

node scripts/verify-watchlist.ts   # check every entry against the API
node scripts/retire-watchlist.ts   # mark archived and deleted ones inactive
node scripts/derive-watchlist.ts   # propose candidates, writes nothing
node scripts/derive-packages.ts    # map repositories to packages, writes nothing

derive-packages.ts fills the packages field, which is what lets the agent ask what a project's downloads are doing rather than only what its repository is doing. It never maps on a name match: a candidate is proposed, the registry's own record for that name is fetched, and the mapping is kept only when the registry points back at this repository. That rule is not decoration — the first version matched on substring and mapped angular/angular to the npm package angular, which belongs to the archived angular/angular.js, and would have credited a dead project's downloads to a live one.

Nothing is ever removed. Retired entries stay on the list with active: false, because deleting one would erase the record that it was watched at all, and the findings collected while it was still link to it.

Two scheduled workflows do the rest. pulse.yml runs every four hours for repository base and releases; daily.yml writes the canonical snapshot, classifies spikes, and collects issues and manifests. Both commit every run, including when nothing changed — scheduled workflows are disabled after 60 days of repository inactivity, and that commit is what prevents it.

Configuration

Secret	Purpose
SIGNAL_GITHUB_PAT	Fine-grained token, public repository read only. Secret names cannot begin with GITHUB_.
CLOUDFLARE_API_TOKEN	Cloudflare Pages: Edit
CLOUDFLARE_ACCOUNT_ID
GROQ_API_KEY	Optional. Without it every reading still publishes, with numbers and no prose.

Both optional groups are genuinely optional: unset, their steps skip and say so. Nothing is marked as a failed run, because choosing not to configure something is not a fault, and a warning that is always on is a warning nobody reads.

License

MIT. See LICENSE.

Cost

Cloudflare Pages, GitHub Actions on a public repository, and Groq's free tier. Static assets and bandwidth on Pages are unlimited, so visitor traffic has no ceiling. Every remaining limit applies to the build side, which is bounded and predictable.