ghtrends ↗

v0.27: one domain report, no light/deep tiers. Assess search demand, commercial supply, open-source supply and user needs first; derive 0–3 directions from evidence, never invent gaps first. Target: one minute after scope confirmation, with a 55-second server deadline (18s collection, 9s original pages, up to 25s writing). One report-writing call, at most 3,600 output tokens, thinking disabled, no model repair loop. Scope preparation happens before confirmation. Failed delivery returns the attempt's credit; the deadline is not a guaranteed success SLA. Historical reports and purchased-credit records remain readable; new deep tasks, retries and research-pack purchase links are closed. Legacy focused-research sections below describe historical behavior.

English · 简体中文

Know where to build.

GitHub supply × Google search demand. Find growing categories, inspect the competition, and share the evidence.

Open the radar · Methodology · Report a bug

What does it tell you?

Ocean names summarize the observed search and open-source signals; they do not establish commercial competition. A quiet ocean can still be a valuable niche.

Every result includes its source queries, dates, methodology version and limitations. Category and report pages include readable HTML evidence before JavaScript loads, and shared links show the specific report in their previews. Search interest measures attention, not paying customers. A quadrant is a research starting point, not a prediction of commercial success.

Broad topics retain the original query and a scoped ocean verdict. The report distinguishes a source-backed market assessment from search × open-source competition. When AI synthesis needs another pass, the original query, measured facts and scoped rules-based conclusion remain available. Product discovery pages remain visible when the competitor synthesis is pending; an advertising section appears only for captured ads.

Related MCP server: GitHub MCP Server

A useful answer when the data is thin

A scan gives an evidence-based recommendation even when a reliable quadrant cannot be established. It separates known facts, an initial recommendation, and the next validation steps. Sparse search data is never turned into a confident “blue ocean” or “dead market” verdict.

For example, ai4s resolves to AI for Science. GitHub searches ai4science, ai-for-science, and ai4s separately and deduplicates returned repositories. Incomplete unions report a lower bound. Google Trends measures the full field name. Existing abbreviation-only reports keep their original evidence and offer a one-click rescan.

Scans show source progress and a preliminary result before optional project details finish. Interactive scans have priority over scheduled refreshes, and repository enrichment uses three bounded workers. Source rate limits can still increase latency. Without a model key, self-hosted scans use the built-in topic mappings. With a DeepSeek key, input normalization and a source-grounded product strategy are enabled.

The website supports English and Simplified Chinese, including report text, Markdown and PNG exports. It follows the browser language on first visit; the header switch saves your preference and preserves the current page. Add ?lang=zh or ?lang=en to open a specific language.

Try it in 60 seconds

No installation: ghtrends.dev/radar/. Browse public reports without signing in.

CLI, local UI and MCP: Node.js 22.13 or newer.

Launch without a global installation:

npx --yes --package=https://ghtrends.dev/radar/ghtrends.tgz ghtrends ui

Or install the CLI:

npm install -g https://github.com/noahbenjamin1994/ghtrends-radar/releases/download/v0.21.1/ghtrends-radar-0.21.1.tgz

ghtrends ui
ghtrends scan --topic mcp-server --json
ghtrends repo facebook/react
ghtrends compare facebook/react vuejs/core --format md

The release package is ghtrends-radar; the executable is ghtrends. The existing npm package named ghtrends belongs to a different project. This release is distributed through GitHub Releases.

One engine, three ways to use it

• Web: search first, read public examples, and save reports and projects in My research. Select two to six repositories inside a report to compare them. The category map and detailed method are optional views. Saved projects refresh when opened; they do not send alerts. Share a report or export PNG, Markdown or JSON.

• CLI: scan a category, inspect a repository, maintain a persistent watchlist, compare repositories and generate a report.

• MCP: give your agent structured GitHub and search-demand evidence without leaving its workflow.

ghtrends scan --topic agent-memory --keyword "ai agent memory" --geo US
ghtrends scan --topic mcp-servers --report findings.md
ghtrends watch add facebook/react
ghtrends watch list
ghtrends watch run
ghtrends watch remove facebook/react
ghtrends report --topic coding-agents --format md
ghtrends compare facebook/react vuejs/core sveltejs/svelte --format json
ghtrends collect
ghtrends ui --port 3721 --no-open
ghtrends mcp

scan is a category scan. report formats the full available evidence, leading repositories, issue signals and limitations into a document. watch run performs one refresh; an external scheduler can run it periodically.

MCP configuration

Install the release first, then add to your client's MCP configuration:

{
  "mcpServers": {
    "ghtrends": {
      "command": "ghtrends",
      "args": ["mcp"]
    }
  }
}

Tools: ghtrends_scan, ghtrends_repo, ghtrends_compare, ghtrends_watch_list. Transport: stdio. Ask your agent: “Compare active GitHub supply and search demand for agent memory. Show the evidence and its limits.”

Credentials and storage

Public GitHub requests work without credentials within GitHub's unauthenticated limits. For larger scans, configure either a token or your own GitHub App:

export GITHUB_TOKEN=your_token
# Or:
export GITHUB_APP_ID=your_app_id
export GITHUB_INSTALLATION_ID=your_installation_id
export GITHUB_PRIVATE_KEY_PATH=/secure/path/github-app.pem

The App installation token refreshes automatically. Only public repositories are returned. Never put credentials in source control or expose them in a browser.

• The release includes dated public starter snapshots so the local radar is useful on first launch. They retain their source dates; snapshots older than 14 days are reclassified as insufficient evidence until refreshed.

• SQLite defaults to ~/.ghtrends; override with GHTRENDS_DATA_DIR.

• Hosted history and watchlists are saved by account in SQLite; self-hosted CLI, MCP and Web share the local workspace. Completed scans appear in My research.

• Anonymous Trends sessions first visit the Trends home page to obtain a session cookie, then use the JSON endpoints on the same proxy route. Session renewal happens every ten minutes; a sticky residential session of at least thirty minutes is suitable.

• GOOGLE_TRENDS_PROXY configures the primary HTTP(S) proxy. Optional GOOGLE_TRENDS_PROXY_FALLBACK adds one backup in the same region. Decodo sticky-session IDs are renewed for each new research; each complete Google token sequence keeps one exit. A limited primary attempt triggers one complete attempt through a separate backup session. A blocked session cools down without blocking fresh research on a new exit; successful cached evidence remains available. Admin shows the configured route count.

• Google Trends collection shares a paced queue (1.5 seconds between requests; GHTRENDS_TRENDS_INTERVAL_MS configures 1–10 seconds), and concurrent identical queries share a request. HTTP 429/403 pauses collection until Retry-After or a default 15-minute recovery window. Cooldown survives restarts, successful cached queries stay available, and a prior successful snapshot keeps its original date. The report displays the next refresh time.

• Google Trends web endpoints may rate-limit requests or change. Temporary connection/5xx errors get one bounded retry. Classification requires fresh, usable weekly evidence. Collection state and measured zero values have separate meanings; baseline claims require observed weekly data.

• Product copy and AI briefs use affirmative facts, current status and specific next actions. Saved narrative text is checked at display time; source evidence retains its original record.

• New signed-in hosted scans are private by default. Explicitly publish a report to share it; stop sharing to revoke future access. Someone who already downloaded a public report may keep their copy. Older anonymous reports remain public and cannot be assigned to an account automatically; save any accessible report to My research.

For reproducible imports:

ghtrends scan --topic mcp-servers --trends-file demand.json --json

demand.json follows DemandEvidence in src/core/types.ts: matching keyword and region, original Google Trends source URL, collection timestamp, weekly 0–100 observations, and optional anchor values. Partial weeks must be marked. Inputs are validated; daily data and missing weeks are not treated as weekly evidence.

Hosted and self-hosted

One codebase. Self-hosting needs no login provider and no external database. Hosted mode requires sign-in for resource-consuming research and stores user/report ownership, history, watchlists, sessions and daily usage in SQLite. Single instance with a persistent data volume; back up with SQLite's backup API or while stopped, rather than copying a live WAL database file alone.

Upgrade/rollback: back up SQLite before upgrading. Versions before 0.3 do not enforce private-report ownership; never run them against a database containing private reports. Restore the pre-upgrade backup before such a rollback, and preserve the newer database separately.

Optional self-hosted AI configuration (server environment only):

export DEEPSEEK_API_KEY=your_deepseek_key
export DEEPSEEK_MODEL=deepseek-flash
ghtrends ui

The model proposes one primary Trends phrase, up to two genuine synonyms, and bounded GitHub topic/phrase queries. It also reviews project roles using quoted repository metadata; deterministic code calculates pressure and search direction. Ambiguous acronyms request clarification. It then maps 3–5 distinct opportunity directions in English and Chinese, with one prioritized strategy to test. Market metrics remain a separate, deterministic layer. Use public research inputs: queries and bounded public-source excerpts go to DeepSeek, Google, DuckDuckGo and GitHub as needed. Collected evidence remains readable during model recovery.

Find the directions that fit you. On a report with researched directions, choose your experience, available time and goal. The model ranks the existing directions and explains the fit plus a small first step for each. This uses zero research credits and shares the scope-preparation request budget. The selected profile and recommendations stay private, are restored on return, and are retained for 30 days; Markdown export is available. Public reports and their demand/competition evidence keep their original content. The configured research model receives the profile and relevant report proposals; use research-appropriate context. Generation uses disabled thinking and bounded, targeted repairs; admin records it as personal direction ranking.

The strategy process

1. Select up to four relevant project READMEs across distinct user jobs, their latest published release notes, and three individual issue excerpts. Historical reports also filter adjacent-object Issues at display time. Request cards show who needs what, a possible contribution, source dates and the latest observed status. Expand a card to read the current workaround, desired outcome, verification step and original quote. Completed requests point to existing progress. Recent requests appear alongside highly discussed ones; repeated links and identical requests from the same identified author are merged. Broad fields retain their full original scope alongside open-source directions.

2. Use direct generation to answer the original topic first: overall opportunity, demand drivers, competition, entry conditions and evidence coverage. Then map distinct customer jobs: typically five for a broad field and three for a narrow product. Every direction opens with a plain-language name, customer, need and offered service, followed by demand, competition, resources, a first-release estimate, upkeep and a concrete experiment. Broad consumer fields span multiple lifecycle stages; source availability guides evidence confidence while preserving the original scope.

3. The first pass produces a compact English research blueprint. Check each direction with one targeted GitHub repository search, up to one README and one issue search (up to three excerpts), using two source workers. An evidence editor then turns the blueprint and collected sources into the full bilingual report; a bounded editing pass repairs wording or citations. Directions, the priority strategy and the market overview have separate, bounded output budgets.

4. Show source signals separately from research inference. Parent-query search growth stays separate from direction-level demand. Project features establish supply; observed demand needs relevant user-request evidence. Strong demand requires multiple request sources and model review of their relevance. Limited search coverage keeps competition estimates provisional.

5. Develop the selected direction in depth with a causal mechanism, tradeoff, critical assumption and proposed continue/redirect criteria. Exact cited excerpts and direction IDs are checked. All directions appear in the interactive report, Markdown, JSON and server-rendered HTML. Old single-strategy reports remain readable.

Sparse source coverage produces a domain hypothesis; document-grounded recommendations are labeled source-led hypotheses. Both represent research proposals. The interface shows source progress and preliminary measurements while the strategy develops. AI-enabled scans can proceed with available evidence during Trends cooldown, within the same account and attempt limits. A compact reasoning blueprint is followed by separately validated bilingual directions and an overall judgment; direction writing runs with two requests at a time. Completed sections are cached. Strategy generation, direction/overall writing and corrective editing are recorded separately in admin usage. Provider reasoning text stays outside stored reports. Results use a six-hour strategy cache; deeper analysis adds latency and model usage.

PUBLIC_URL may include a directory, for example https://example.com/radar. The same build supports both directory hosting and a local root URL. Forward that prefix unchanged to the server and configure the matching Logto callback.

Account and credits: /account separates daily research, the introductory focused research and purchased packs. Hosted operators can connect Nexus with the server-only GHTRENDS_NEXUS_URL, GHTRENDS_NEXUS_PROJECT_ID and GHTRENDS_NEXUS_PROJECT_KEY. The account reads purchase confirmations, per-purchase expiry and credit activity through the authenticated backend. Entries distinguish reservation, completion, release, expiry and purchase adjustments. A temporary connection issue shows a sync message and the previous record timestamp. Purchased-task admission uses its own deployment flag, described below; set GHTRENDS_CASHIER_URL and GHTRENDS_RESEARCH_PLAN_ID to show the server-configured purchase link.

Focused research preview

Set GHTRENDS_DEEP_RESEARCH=1 with a research model key to try the private, selected-direction workflow. Open a report direction and choose one question: compare other products, scope a first release, build on open source, or find the first users. Optional personal-fit conditions travel with the task. Targeted searches, current project requests and original documents support a compact bilingual decision brief. Facts, proposals, resource estimates and source quotes stay distinct.

Hosted accounts receive one introductory focused-research credit for the account's lifetime. Starting a task reserves it; complete delivery uses it; partial delivery or a restart returns it. This balance is separate from daily standard research. Self-hosted tasks use the operator's provider keys. Requests are idempotent and tasks survive navigation and restarts; a partial task can resume up to three total attempts. Account attempts are capped at three per UTC day; GHTRENDS_DEEP_DAILY_REQUESTS sets daily new-task capacity (default 20). Admitted tasks can resume within their daily attempt budget when new-task capacity is full; attempts are counted on their actual UTC date. Reading and exporting consume zero research credits. Private history supports Markdown/JSON export and content deletion; minimal usage records retain attempt counts and used-credit status. The preview remains separately configurable while quality and cost evaluation proceeds; the hosted service offers an optional one-time research pack.

All requests hard-disable thinking, regardless of legacy flags. Deep research combines factual, decision and correction reviews into one call. Invalid review output stops without an extra review; original quote checks remain local. Writing uses at most 12 sources, with model excerpts capped at 2,000 characters (licenses: 6,000), labeled when truncated. Original evidence and interrupted drafts are retained. No automatic stronger-model fallback runs.

Free light briefs and deep research

New light briefs deliver exactly three directions, with a concise judgment, comparison and next step that follows the selected direction. Details and evidence open on demand; historical reports retain their original direction count. Standard research keeps its free daily allowance.

Deep research covers four search jobs: official products/pricing, user problems, independent community discussion and open-source implementations. Scope and competitor questions can follow up on at most two discovered direct-product domains to find their offers. It combines the existing residential search route with public sources, reading at most eight original pages (up to two per domain, prioritizing pricing), four concurrently, with a 40-second overall reading budget. Light research retains four domains and 20 seconds. Repository-search aggregates stay outside synthesis. Reports lead with the decision, next step and observable success rule; details expand on demand and collected-but-uncited leads are labeled. Coverage counts do not establish buyers or willingness to pay. Factual negation is preserved; commercial pilots distinguish expressed interest from paid commitment.

A separate, bounded high-effort decision review checks buyer/job alignment, claimed differentiation and whether the experiment supports its conclusion. Failed review cannot silently approve delivery; repairs share the same attempt limit. Missing buyer evidence may lead to an explicit pause and concrete workflow-discovery test, not an invented niche or a claim that an interview or one purchase validates subscriptions.

These limits supersede the older three-to-five direction and four-domain focused-research descriptions below. Introductory and purchased credits retain complete-delivery settlement.

Purchased research

Hosted operators can set GHTRENDS_PAID_RESEARCH=1 after configuring Nexus and focused research. The UI uses the introductory credit first, then shows the purchased balance and an explicit one-credit confirmation. Each account can have one active research task. Purchased research has a separate daily new-task capacity (GHTRENDS_PAID_DEEP_DAILY_REQUESTS, default 20) and an account attempt budget (GHTRENDS_PAID_DEEP_DAILY_ATTEMPTS, default 10). Each research still allows three attempts.

Tasks are saved before admission; the credit is reserved when execution begins. Queued tasks can be canceled. Complete delivery and its settlement intent commit together in SQLite; interruptions release the original purchase reservation, retaining its expiry date. Recovery retries use the same Nexus receipt, including after a lost response or restart. Deleting research content preserves pending financial recovery. The administrator panel shows active settlements and records requiring ledger review. Execution checkpoints are bounded to 15 minutes within a one-hour reservation. This implementation targets one persistent application instance; the hosted service offers 10 focused studies for US$19.90, valid for 90 days from receipt, with tax shown at checkout. Paddle handles one-time payment. The live account flow and canceled-order callbacks have been verified; completed purchases and refunds were verified in Paddle sandbox. Research quality and same-version cost benchmarking continue separately.

Original text and licenses

GitHub direction searches match issue titles and bodies, ordered by relevance. README and release excerpts contain publisher text; publication dates, request states and review guidance remain separate. A repository name used as a citation ID is recoverable only when its exact quote uniquely matches a collected source from that same repository.

Focused research reads relevant Markdown sections from the selected project's README and release notes within the existing excerpt budget. Within its four-domain original-page budget, documentation links matching the selected direction take priority over general comparison articles; public discussion priority and each page's provenance remain intact. URL and term matching guide reading order; source authority and factual support still require review. Pages use gzip, Brotli or deflate when offered. Both wire traffic and decoded content are capped at 2 MB; retained excerpts stay within 6,000 characters, with truncation marked. The introduction and selected original spans retain their source URL and collection date; omission markers and truncation metadata describe partial coverage. Markdown code blocks, inline code and autolinks survive presentation cleanup, including shell redirection syntax.

Focused research also reads up to two selected GitHub issue threads. It keeps one early and two later public comment excerpts per thread, with dates, original links, account deduplication keys and project-association metadata. Thread message totals, included comments and distinct sampled accounts stay separate in saved sources. Focused writing receives the original messages, roles and dates; collection totals stay outside its evidence quotations. Large threads use the first and last reported pages (up to 30 comments each); read failures and partial coverage remain visible. Invitations are proposed from specific user experiences, while project replies provide context and capability checks.

New research selects original pages from up to four domains already observed in organic search. Publisher pages, individual Hacker News discussions and GitHub Discussions retain their source identity. Hacker News uses its official public API and reads at most three direct comments per topic. GitHub Discussions are read from observed links, with accepted answers retained; the backend GitHub credential needs the relevant read permission. Up to two selected repositories contribute their license files for checking attribution, distribution, source-disclosure and third-party conditions.

The report's “Original text and licenses” section keeps collection dates, excerpts and reading status, also included in Markdown, JSON and static HTML. Publisher pages describe publisher claims; discussions describe individual experiences. Page reads respect robots.txt, bound redirects, time and size, and pin connections to checked public addresses. Transient connection failures receive one retry within the shared 20-second collection deadline, respecting crawl delays. Every attempt appears in provider usage; access, size, certificate and format failures retain their reading status. Original text is cached for one hour; restricted pages retain the search excerpt and reading status. Public-page requests use a separate transport; backend identity credentials go only to the fixed GitHub API. Set GHTRENDS_SOURCE_DOCUMENTS=0 to disable this layer. Reddit and X body integrations follow each platform's authorization requirements.

Web search (optional)

With GHTRENDS_PUBLIC_SOURCES=1, the three-query budget is retained: community discovery targets public Reddit, Hacker News and Trustpilot indexes; physical-goods topics replace the open-source query with Amazon, eBay and Etsy discovery. Software topics also run one anonymous HN / Algolia search in parallel (four hits, six-second timeout, six-hour success cache). Deduplicated candidates are capped at 30, reuse the existing relevance review, and contribute at most 16 search sources to writing; original-page reads remain capped at four. No platform account, session cookies or additional paid API is required; existing search proxy usage still applies. Indexed snippets are not complete discussions, reviews or live prices. Listings do not establish sales, authorization or unmet demand. Access challenges remain explicit collection failures; login and CAPTCHA are not bypassed.

The default GHTRENDS_SEARCH_MODE=direct reuses GOOGLE_TRENDS_PROXY and its fallback. A lightweight Google mobile-page request uses curl_cffi with a matching browser connection profile. Docker bundles the Python transport; source/npm installs can use python3 -m venv .venv-search, .venv-search/bin/pip install curl_cffi==0.16.3, then set GHTRENDS_SEARCH_PYTHON to that environment’s Python path. GOOGLE_SEARCH_PROXY and GOOGLE_SEARCH_PROXY_FALLBACK optionally select dedicated routes. Each search fetches a single page. Decodo’s port 7000 uses rotating exits for these requests while preserving country targeting; Trends retains its sticky sessions. Search sends only a fixed consent preference and uses the existing residential traffic balance. The optional api mode uses DECODO_SCRAPER_TOKEN and a separate managed-search entitlement; off disables search.

A scan runs up to three first-page queries: commercial alternatives, user problems and open-source assets. Direct mode gives Google one fresh-exit retry, then uses DuckDuckGo Lite with one bounded retry. Fixed routes are attempted once per engine; rotating exits retain configured country targeting. Up to four requests share a 45-second query budget, with an 18-second maximum per request. A challenged engine pauses for five minutes independently of the other engine. Google results are cached for six hours; fallback results for 30 minutes. Each query keeps its actual engine and original collection time, including cached results. DuckDuckGo contributes organic results; Google ad coverage stays separate. Query, region, collection date, links, snippets and organic/ad labels are retained. Worldwide reports request US search settings; other regions use the selected engine’s supported market settings or worldwide results. Only bounded excerpts enter AI research, while credentials remain on the server. Admin records search calls, cache reuse, latency and failures separately; measured direct-request HTTP bytes join residential usage. Lightweight Google pages have limited ad-slot coverage. Reports distinguish this from a captured full ad sample with zero observed ads, including older saved reports. A collapsible, dated CRM browser example shows real HubSpot and Zendesk ad cards separately from each topic’s evidence and scoring. Google Search and Ads Transparency Center links support manual checks; advertised brands, legal payer identities and keyword appearances remain separate facts. Automated full-browser ad collection remains a separate integration.

Reports separate open-source projects, commercial competitors and observed ads. Projects show activity, age and community data. Competitors show who they serve, source-backed pricing or buyback quotes, existing advantages and possible openings. Audience and pricing facts carry individual quotes; trade-in payouts keep their eligibility and device scope. Ads show the destination website, copy, landing page, query and region. Advertiser identity, spend, clicks and conversions require additional evidence. Model inputs prioritize independent websites, then fill remaining slots with same-site pages, with up to four organic and two ad sources per query, plus up to two pricing pages for the competition query.

The report combines Trends, commercial alternatives and the open-source ecosystem. The research landscape adds blue-ocean candidates, growing red oceans, red oceans, quiet oceans and an opportunity watch. This qualitative assessment keeps the measured GitHub quadrant intact. Ads indicate marketing intent; direct user behavior supports demand judgments. Incumbent analysis explains data, distribution, ecosystem or switching barriers. Search ranks and project stars describe their own sample. Open-source directions cite a concrete project and explain the complementary contribution and license check.

Direct generation produces a compact blueprint; direct generation writes concise bilingual direction, priority and overall sections. Each section is validated and cached. A final review checks source scope, actual capabilities, proposed services and bilingual meaning; oversized passages receive targeted edits. Issue interpretations use only supplied request sources. Admin records the usage of each step.

Community readings distinguish specific requests and user experiences from existing-workaround advice and author promotion. Advice and promotion remain source context and contribute zero demand votes or demand cards, including duplicate retrievals of the same comment. Research proposals state skills, permissions, recruiting access and participants as requirements, preserving source dates, versions and testing restrictions.

A mistaken citation ID can be repaired by selecting from supplied sources containing its exact quotation; valid citations and source text stay fixed. Unsupported references continue through quality checks.

Section format or output-budget errors get one bounded direct-generation recovery for that section; completed sections retain their cached results. Evidence review and community reading may recover once with thinking disabled after an output-budget or format failure. Provider throttling and network failures keep their existing recovery path. Citation length repair preserves source identities and verbatim quotations with the actual field limit.

Authored prose is checked for clear English/Chinese language swaps and repaired with its paired meaning; source quotes retain their original language. Request interpretations preserve the author’s stated workaround and treat vendor code, permissions and device access as prerequisites for proposed contributions.

The recommended direction has one shared pilot definition: recruitment, task, time window, measurement and proposed continue/redirect criteria. The report summary, direction card and exports derive their experiment text from that definition. Review edits apply to the shared fields; source quotations and other directions stay independent.

The recommended pilot uses shared numeric counts for people, tasks per person and qualifying outcomes. The application checks the denominators and generates equivalent continue/redirect criteria in both languages, with an intermediate range for further evidence. Pilot authoring uses a separate short call with bounded original evidence and thinking disabled. Source-backed project-use notices also appear beside resources and first-release scope, and in text/Markdown exports. These observations support checking the actual license and permissions for the proposed reuse.

Before writing directions, a compact capability audit checks original project/product documents. It retains exact feature and constraint quotes, separates proposed additions from existing functions, and records code/data permissions, compatibility and access checks. Search snippets remain discovery leads. The audit runs with thinking disabled, allows one bounded correction and is stored with the report for inspection. A pending evidence review keeps delivery in the source-recovery flow.

Standard research uses a compact direct-generation blueprint and overlaps independent source collection and report sections. A source-bound capability audit precedes up to five parallel direction writers, market coverage and request interpretation. The selected pilot and summary are independent. There is no mandatory full-report second review: exact citations, evidence scope, bilingual structure and shared experiment criteria remain delivery gates, with targeted semantic review when a concrete check fails. Style checks preserve technical terms such as untrusted inputs and immutable logs. Thinking is hard-disabled regardless of legacy flags. Both report types stream live activity with polling recovery. Prompts, draft JSON and reasoning text stay on the server. Provider timings and specific validation failures support operator diagnosis; upstream rate limits and format recovery can still extend latency.

Reports include an optional progress check-in: useful insight, a request for more detail, a chosen direction, a changed plan or a completed validation. Signed-in users can save a note (up to 500 characters), edit or delete it, and export their feedback from Account. These records stay separate from shared reports and model inputs; only the author and deployment administrators can read them. Removing a report from history or deleting focused research also removes that author’s related feedback. Admin counts unique people separately from responses and internal testing; feedback timestamps record submission, with user interviews and observed behavior providing additional validation.

Access and research credits

A fresh scan, project analysis, or comparison uses one credit. Cached results are free. The header, research form, and account menu show the balance and reset time in your timezone. A credit is reserved while research runs; collection issues and interrupted work return it exactly once, including after a server restart. Project and comparison collection starts through an explicit action protected by the session and CSRF token; opening a saved page reads cached evidence.

GHTRENDS_DAILY_REQUESTS bounds hosted collection attempts across accounts (default 200/day, including returned credits). Each account can attempt up to three times its daily credit allowance. One fresh research job per account runs at a time. Admin shows today's reserved, used and returned credits, the service budget, and the Trends connection/recovery status. GOOGLE_TRENDS_PROXY supports authenticated HTTP(S) proxies; GHTRENDS_TRENDS_PROXY_REGION supplies a safe region label. Proxy credentials stay server-side, and provider recovery windows apply to each configured route.

Before collection, the web app prepares a research scope for confirmation. Greetings and placeholder text receive immediate guidance; ambiguous terms offer specific meanings. Scope preparation is free and has its own request budget. Confirmed scopes are bound to the account, input, region and search phrase for ten minutes. A temporary model outage offers an original-phrase scope for explicit confirmation. Brief connection interruptions keep the research task open and reconnect automatically. Administrators can identify preparation requests separately from research runs.

For a public hosted instance configure GHTRENDS_HOSTED=1, HTTPS PUBLIC_URL, LOGTO_ENDPOINT, LOGTO_APP_ID, LOGTO_APP_SECRET and optionally GHTRENDS_DAILY_SCANS (default 10). Create a Traditional Logto application with ${PUBLIC_URL}/auth/callback as its redirect. Keep GitHub/DeepSeek credentials in server secrets, never VITE_* or browser storage. OIDC uses PKCE, nonce/state and signed-token validation; the browser gets an HttpOnly, Secure session cookie. Personal mutations also require CSRF validation.

Administration

The account menu contains the administrator entry. Anonymous daily counters show report reads, shares, exports and open-source entry clicks; completed user scans exclude scheduled collection. These are action counts, not unique visitors, installations or GitHub stars. Counters contain no query text or visitor identifiers, respect browser Do Not Track, share the operational retention period, and can be disabled with GHTRENDS_ANALYTICS=0. They are enabled only in hosted mode. Historical activity before this upgrade is unavailable.

/admin shows scan status, queue, source errors/latency, GitHub quota snapshots, accounts, and actual DeepSeek input/output/cache-token usage. Operational logs survive restarts; interrupted scans are marked interrupted. They retain 30 days by default (GHTRENDS_LOG_RETENTION_DAYS, 1–365). Saved reports have separate retention. No credential values, session tokens or private report bodies are returned by the admin API.

# Deployment-specific Logto user IDs, copied from the user's profile in Logto.
# These are immutable sub IDs, not usernames, emails or display names.
export GHTRENDS_ADMIN_USER_IDS=your_logto_user_id,another_logto_user_id

Restart after changing deployment configuration. There is no hosted administrator by default; the server checks the allowlist on every request. In local mode without Logto, the workspace owner has admin access—keep that unauthenticated server on loopback, or enable hosted authentication before publishing it.

AI usage recording begins at upgrade: earlier consumption is unknown. Missing usage on failed responses also stays unknown. Optional GHTRENDS_LLM_PRICING_JSON maps each requested model to USD per million token rates: {"your-model":{"input":0.3,"cachedInput":0.006,"output":1.2}} (illustration only, verify your provider's current prices). An optional offPeakMultiplier follows DeepSeek's Monday–Friday 01:00–04:00 / 06:00–10:00 UTC peak schedule. Omit it for flat pricing. Per-call estimates are saved at request time, exclude unpriced calls, and are not invoices. Provider pricing and usage fields.

With DECODO_API_KEY configured on the server, administration also shows the official residential plan balance, expiry and daily billed traffic. It refreshes every 15 minutes and retains the last successful reading during provider recovery. Separate HTTP measurements show compressed traffic per research run, request success and 429 counts by primary/backup route. These measurements begin at upgrade and cover HTTP payloads and headers; official provider accounting includes its own overhead and remains the billing reference. Proxy credentials and the management API key stay on the server.

How the method works

Competition pressure · method 2.0.0

GitHub searches use topics and specific name/description phrases. Initial coverage includes original, active repositories with at least 1 star and a push within 365 days. Each query returns up to 100 leaders; overlapping results are deduplicated. An exact search count and an enumerated project sample are separate properties. Known categories keep their published query scope; compound requirements use topic intersections.

Projects are classified as direct alternatives, adjacent integrations, resources, or awaiting review. An optional model reviews the top 60 project descriptions in three bounded batches, with a verified verbatim source quotation for each accepted result. Successful batches remain useful when another batch needs source review. Invalid or ambiguous review items retain local metadata rules. Reports show every project's role, quote and review method. Direct alternatives are grouped by GitHub owner as a proxy for independent teams; each owner's strongest project contributes to three components:

• Breadth, up to 50 points: 50 × (1 − exp(−effectiveTeams / 12)). Team weights combine log-scaled stars, forks and maintenance.

• Established alternatives, up to 30 points: 30 × (1 − exp(−2 × sum(maturity))). Maturity combines project age, stars, forks and maintenance.

• Leading project strength, up to 20 points: 20 × max(maturity). Top-three owner star share provides separate descriptive context.

The operational reference line is 45/100, with lower confidence within 5 points. A lower bound at or above 45 supports established competition even from a partial sample. Limited competition requires complete enumeration of the displayed scope, at least one direct alternative, and an upper bound below 45 after including projects awaiting review. Truncated samples show ≥ score; complete samples can show a role-uncertainty interval. Zero direct matches prompt further research. See the exact weights and equations.

These versioned heuristics describe observed open-source competition. Stars indicate developer attention, forks indicate reuse, and owners approximate teams. Commercial products, customer adoption and willingness to pay deserve separate evidence. Broad fields and physical-product markets retain a scoped ocean verdict when supported, alongside concrete competitors and user tasks.

For sparse candidate results, DeepSeek can refine the search once using equivalent product names within the original task. Generic delivery words such as “app” can be omitted from exact GitHub phrases; subject and feature requirements remain explicit. The report retains original and added queries. Search evidence, competition coverage and the recommendation appear separately, so a useful observed signal leads to a concrete next step even while the landscape remains provisional.

Search direction

Two years of Google Trends data provide 4-, 8- and 13-week comparisons. Each phrase is independently normalized for the same region and period. Primary-term fallback follows the planned synonym order and data quality. Values stay separate; opposite measured synonym directions display as mixed.

• Percentage direction requires 26 consecutive complete weeks, at least 60% positive values in that window, and a baseline median of at least 3. Completed weeks are checked against the source collection time.

• Rising/falling normally requires an 8-week median change of at least ±10%, an absolute change of at least 2 index points, a two-week-block resampling band on the same side of zero, and supporting 4-/13-week checks.

• Gradual change can qualify over 13 weeks when its percentage, absolute change and resampling band pass the same checks and shorter windows agree. The report names the direction basis.

• Recurring annual shapes require at least 40 date-paired weeks, correlation ≥0.75, and material rises and falls in both annual profiles using four-week block medians. Direction then uses the same eight weeks a year earlier. The recent-window percentage remains visible separately.

• A baseline below 3 can produce an early rise when at least six of eight weeks reach 10 and the final four-week median retains 80% of the preceding four. Its percentage stays empty and confidence stays low.

• Conflicting, rounded or sparse measurements retain a qualified direction. Collection gaps display recovery actions. Established competition can still support red-ocean guidance with a separate pending/mixed search status; growing-red and blue labels require usable rising search evidence.

The resampling bands diagnose stability. The overall opportunity score stays null; competition pressure has its own named scale. Confidence is capped at moderate; low evidence, boundary cases and early signals receive low confidence. Every report retains source dates, queries, roles and method version. Refreshing creates a new snapshot.

Repository evidence: official GitHub star-history calendar buckets, a bounded recent issue sample, human maintainer responses, and returned contributor commit counts. Calendar buckets are not rolling 24-hour net star changes. Contributor and issue sample limits appear beside the results. Open issues are leads for research, not proven market gaps.

Reproducibility: source evidence, analysis date and method version determine each report ID. Hosted report links are immutable snapshots. Local reports stay local unless you deliberately share them.

Development

git clone https://github.com/noahbenjamin1994/ghtrends-radar.git
cd ghtrends-radar
npm ci
npm run check
npm run build
npm test
npm start

React + TypeScript + Vite for the UI; TypeScript core, Express, SQLite, Commander and the official MCP SDK. One package, no required database service, no required LLM key.

docker build -t ghtrends .
docker run --rm -p 3721:3721 -v ghtrends-data:/app/data \
  -e HOST=0.0.0.0 ghtrends

For a shared public instance, set GHTRENDS_HOSTED=1 to keep the homepage restricted to curated categories; custom scans require sign-in and are saved as private personal reports. Search-term variants are stored separately and cannot replace canonical category evidence. Set PUBLIC_URL and optionally GHTRENDS_AUTO_COLLECT=1. Set TRUST_PROXY only to your actual trusted reverse-proxy network. The collector and public scan queue share a bounded, paced workflow. Public scans are rate limited; scheduled collection refreshes curated categories daily using source collection times. Recalculating a report does not postpone collection. Method upgrades invalidate old analysis caches.

Contributing

Web-only source APIs

Admin session required; POST additionally requires X-CSRF-Token. Routes live under the configured base path, e.g. /radar/api/sources. This collection surface never calls a platform API or LLM, and never writes reports.

Successful sources[].links includes up to 12 same-site pricing/documentation/features/license navigation hints, not verified evidence or automatic crawl targets. Search always returns a boolean cached; cached pages keep their original observation dates. Batch workers schedule origin groups so one site's URLs cannot occupy all slots.

Two pre-release residential-proxy samples on 2026-09-24: cold search 3.7–5.1 seconds, two-page cold batches 9.1–14.6 seconds; warm search 15–16 ms and batches 8 ms measured over server loopback (excluding public client round trips). One robots timeout recovered. These samples verify connectivity, not a long-term success rate or P95 guarantee. No LLM calls were made.

Search uses residential Google/DuckDuckGo HTML collection with a 20-second budget including queue time. HTML reads use GHTRENDS_DOCUMENT_PROXY, falling back to the search/Trends proxy, with a shared 15-second page/batch deadline. Limits are deadlines, not measured latency guarantees. Successful pages cache for one hour, failures for one minute; duplicates coalesce, same-origin reads serialize and batches have at most four workers. At most two distinct API requests run concurrently, with 12 new requests per admin/minute. Downloads and decoded bodies are each capped at 2 MB; excerpts at 6,000 characters with truncation labels.

Results use state: ready|partial|failed and per-page read.status. Invalid input returns 400, missing authentication 401, missing admin/CSRF 403, overload 429, search timeout 504, setup/upstream failure 503. Partial page failures return 200 with explicit per-page status, never fabricated empty evidence. Caller-supplied cookies, proxies and headers are rejected. Public HTML/text only: no login, browser rendering, CAPTCHA handling or complete comment pagination. HN reads HTML here, not Firebase/Algolia. Robots and existing restricted-platform protections remain enforced. Proxy tunnels pin checked public IPs while preserving Host and verified TLS identity. Existing report workflows are not switched automatically.

Research model budgets

Scheduled refreshes collect public metrics without AI. Document selection is deterministic. Remaining model calls cover planning, relevance, capabilities/issues, blueprint/sections/pilot, optional fit, deep writing and one combined review, conditional repairs, and CLI briefs. Successful caches are reused.

Output ceilings (tokens): planning 900; relevance/copy 2,200; evidence/issues 3,500; default 4,000; deep write/repair/review 4,500/3,000/1,800; legacy full bilingual recovery 9,000. Copy batches have at most six fields. Context-tracked runs stop at 24 calls (deep: 12) or 60,000 reserved output tokens. Unscoped CLI calls retain per-request caps. These are ceilings, not targets or currency guarantees. Cost-policy tests use mocks, not paid generations.

Useful first contributions: improve keyword/topic mappings with evidence, add adversarial algorithm tests, document actual collection failures, and improve accessibility. Open an issue with the original source URLs and timestamps when a classification looks wrong. Run the checks above before submitting changes.

If ghtrends helps you find something worth building, share the report and star the project so you can find it again.

MIT. Not affiliated with GitHub or Google. Built with lessons from the existing trendscout project's Google Trends collection workflow.