Build / Cursor · engineering-technology (AI / developer-tools ecosystem: changelogs, API references, release notes, model SKUs, plus arXiv AI-infrastructure research)

DepWatch

A cited dependency & AI-ecosystem watchdog that lives in your IDE and your CI — every breaking change comes with a real source link. A working watchdog that, on first run, shows you the real, source-linked changes and AI-infra research affecting the exact libraries and AI vendors in your codebase — and keeps watching on a schedule via a saved Synorb Beacon.

A cited dependency watchdog for your repo.

A real Next.js + Node project (runs with `npm run dev`, deploys to Vercel) containing: (1) a server-side /api/depwatch route that reads SYNORB_API_KEY + SYNORB_SECRET from .env and pulls cited Synorb manifests for the dependencies/AI-vendors the repo actually uses; (2) an in-IDE Cursor MCP App card and a clean web view that render each change as a cited claim card (what changed, claim_type/confidence chips, a 'View source' link to the real changelog/arXiv URL); (3) a saved Synorb Beacon scoped to those vendors + the 'resource' tag_type (API endpoints, model SKUs, SDK packages) wired into a Cursor Automation that runs daily/weekly, posts a cited digest to a PR comment or opens a Linear/GitHub issue, and self-verifies its own source links before declaring done.

It solves the single most visceral pain for Cursor's pro-engineer audience — breaking changes in dependencies, APIs, and model SKUs — and it is impossible to do credibly without citations, which is exactly Synorb's wow-factor. It auto-detects which vendors the repo uses from package.json/imports, so the interview is nearly zero-effort. It is native to BOTH new Cursor 2.6 surfaces at once: MCP Apps (in-chat cited cards) AND Automations (scheduled cited agent that comments on PRs). Synorb's 'resource' tag_type was purpose-built for exactly this (API endpoints, model SKUs, SDK packages), and orient-before-web-search means it says 'no tracked source covers this' instead of hallucinating a deprecation. This is the demo that makes Anysphere's partnership team see Synorb as the verifiable data spine under Automations.

Click, answer a few plain-English questions, get a working product.

No setup to learn. Cursor asks a handful of normal-person questions — each with a smart default, so “you pick” always works — then builds it with live Synorb data flowing in.

1

I can read your package.json and imports to figure out which tools and AI vendors to watch — want me to do that automatically, or should I just watch a specific list you name?

Smart default: Auto-detect from the codebase (recommended) — then show you the list so you can trim it.

2

How should this look — a clean in-editor card panel, or a small deployable web page too?

Smart default: Both: an in-Cursor cited-card panel for quick checks plus a deployable web view (recommended).

3

When something you depend on changes or gets deprecated, how do you want to hear about it — a daily PR comment, an auto-opened issue, or just the dashboard?

Smart default: A daily cited digest as a PR comment, plus the dashboard (recommended).

4

How far back should the first scan look?

Smart default: Last 30 days (recommended) — enough to catch recent deprecations without noise.

Pick a look. It ships with taste by default.

Terminal

Monospaced, dark, IDE-native: dependency rows as a compact log, claim_type/confidence as inline tags, source links as cyan underlines. Feels like it belongs in Cursor.

Changelog

Light, document-style cards grouped by vendor and date, with a colored severity rail (deprecation = red, new feature = green) and a prominent 'read the original' source pill.

Radar

Dark dashboard with a per-vendor volume sparkline grid and a 'what changed this week' lead column; provenance chips sit under each headline.

Synorb is the cited core. Cursor’s own tools make it pop.

INFRA-BACKED in a real Cursor codebase. BUILD TIME: the user clicks the Synorb 'Add to Cursor' deeplink so the MCP server (mcp.synorb.com/mcp, Streamable HTTP + token) is installed; Cursor calls synorb-stream-search to map the repo's dependencies to real streams, then synorb-manifests (compact=true, target_count 5-10, the 'resource' tag_type) to pull genuine cited examples so the UI is built against the true data shape — no invented fields. RUNTIME: a server-side Next.js /api/depwatch route reads SYNORB_API_KEY + SYNORB_SECRET from .env (server-only, NEVER in client/VITE code) and calls Synorb REST (api.synorb.com /streams/query, /streams/{id}/manifests). Credentials are requested only when .env is created, in plain terms, with a link to https://synorb.com/connect (free, no card). The MCP build-time token and the REST runtime key+secret are kept distinct and both server-side. STAY LIVE: a saved Synorb Beacon (vendors + resource tag_type) is executed by beacon_id inside a Cursor Automation on a daily/weekly schedule.

Cited data core
Synorb
StreamsarXiv AI Infrastructure, Inference & Ops (source_channel 17779468851298592, ~356 manifests/30d — real arxiv.org/pdf source_urls, claim_type=analysis, evidence_type=research_result), engineering-technology streams for dev-tool vendors: Anthropic, OpenAI, Google Gemini, Vercel / Vercel AI SDK, Supabase, LangChain, HuggingFace (HuggingFace NLP channel confirmed live), Stripe, Cloudflare, AWS, AI-infra topic tags confirmed live: Inference Optimization (75/sample), AI Agents (33), GPU Clusters (24), LLM Evals (19), Model Security (14), 'resource' tag_type for model SKUs / API endpoints / SDK packages (migration 343), distinct from topic/data
Primitivessynorb-stream-search (map repo dependencies/vendors to real engineering-technology streams, keyless, zero quota), synorb-manifests with tag_names + the 'resource' tag_type (API endpoints, model SKUs, SDK packages) and a date window for build-time grounding, structured claims with claim_type (statement/data/event/forecast/analysis) and confidence/evidence chips, citations[] / citation_contract (mcp-citations.v1, never_invent_urls=true) — every card links to the real changelog or arXiv source_url, synorb-configure-beacon -> synorb-save-beacon -> run by beacon_id (the saved, re-runnable monitor the Cursor Automation executes), orient-before-web-search + coverage.status honesty (say 'not tracked' rather than guess a deprecation), mode='count' zero-quota preflight to size a pull before billing
Experience layer
Cursor native tools
ComposesNext.js API route + Node fetch (server-side Synorb REST call; credentials never reach the browser), Cursor MCP App (v2.6) — render cited change cards as an interactive in-chat panel with claim chips and View-source links, Cursor Automations — cron/event-triggered cloud agent that runs the saved Beacon and posts a cited digest to a PR comment or opens a GitHub/Linear issue, package.json / import scan (local file read) to auto-detect which vendors and SDKs the repo actually depends on, Recharts sparkline for per-vendor change-volume over the last 30/90 days, a web-search MCP used ONLY to fill gaps for a dependency Synorb does not track, with web-sourced items visibly badged 'web (unverified)' and kept separate from Synorb-cited cards
Web searchgaps only — clearly marked, never the primary feed

Credential safetySynorb credentials stay server-side or local — never in client/frontend code. You only enter them when the agent asks, with credentials from synorb.com/connect.

One click from the original source.

Provenance is the product. Every change card and every claim renders the real source_url from Synorb's citations[] as a 'View source' pill plus a source-domain chip (e.g. docs.anthropic.com, arxiv.org) and the source_channel_display label. The PR-comment digest links each line item to its original changelog/arXiv URL. URLs are ONLY ever the real source_url returned by Synorb (never_invent_urls=true); web-search gap-fillers are visibly badged 'web (unverified)' so a reviewer can always tell Synorb-cited facts from web-sourced ones, and is always one click from the original.

Questions builders ask.

Will this show real live data on the first run?

Yes. The build prompt connects Synorb at build time to learn the real data shape, then wires live Synorb data into the running product. A working watchdog that, on first run, shows you the real, source-linked changes and AI-infra research affecting the exact libraries and AI vendors in your codebase — and keeps watching on a schedule via a saved Synorb Beacon.

Where do my Synorb credentials go?

Only where they are safe. INFRA-BACKED in a real Cursor codebase. BUILD TIME: the user clicks the Synorb 'Add to Cursor' deeplink so the MCP server (mcp.synorb.com/mcp, Streamable HTTP + token) is installed; Cursor calls synorb-stream-search to map the repo's dependencies to real streams, then synorb-manifests (compact=true, target_count 5-10, the 'resource' tag_type) to pull genuine cited examples so the UI is built against the true data shape — no invented fields. RUNTIME: a server-side Next.js /api/depwatch route reads SYNORB_API_KEY + SYNORB_SECRET from .env (server-only, NEVER in client/VITE code) and calls Synorb REST (api.synorb.com /streams/query, /streams/{id}/manifests). Credentials are requested only when .env is created, in plain terms, with a link to https://synorb.com/connect (free, no card). The MCP build-time token and the REST runtime key+secret are kept distinct and both server-side. STAY LIVE: a saved Synorb Beacon (vendors + resource tag_type) is executed by beacon_id inside a Cursor Automation on a daily/weekly schedule.

How are sources shown?

Provenance is the product. Every change card and every claim renders the real source_url from Synorb's citations[] as a 'View source' pill plus a source-domain chip (e.g. docs.anthropic.com, arxiv.org) and the source_channel_display label. The PR-comment digest links each line item to its original changelog/arXiv URL. URLs are ONLY ever the real source_url returned by Synorb (never_invent_urls=true); web-search gap-fillers are visibly badged 'web (unverified)' so a reviewer can always tell Synorb-cited facts from web-sourced ones, and is always one click from the original.

Is this just a Synorb feed, or a real app?

A real app. Synorb is the cited data core; it also composes the platform's own tools: Next.js API route + Node fetch (server-side Synorb REST call; credentials never reach the browser), Cursor MCP App (v2.6) — render cited change cards as an interactive in-chat panel with claim chips and View-source links, Cursor Automations — cron/event-triggered cloud agent that runs the saved Beacon and posts a cited digest to a PR comment or opens a GitHub/Linear issue, package.json / import scan (local file read) to auto-detect which vendors and SDKs the repo actually depends on, Recharts sparkline for per-vendor change-volume over the last 30/90 days, a web-search MCP used ONLY to fill gaps for a dependency Synorb does not track, with web-sourced items visibly badged 'web (unverified)' and kept separate from Synorb-cited cards.

What will it ask me before building?

A few plain-English questions, each with a smart default so you can just say 'you pick': I can read your package.json and imports to figure out which tools and AI vendors to watch — want me to do that automatically, or should I just watch a specific list you name?; How should this look — a clean in-editor card panel, or a small deployable web page too?; When something you depend on changes or gets deprecated, how do you want to hear about it — a daily PR comment, an auto-opened issue, or just the dashboard?; How far back should the first scan look?.

Using a different agent?

The same Synorb MCP connects to every major coding agent. Pick yours.

Free credentials. No card.

Generate credentials, connect Cursor, and ship something with live, cited intelligence.