Market Signal MCP

A local MCP server + Claude Code skill for researching whether a startup idea, product, or topic shows real market signal — search interest, reading interest, community sentiment, builder activity, and company registration — for case competitions and startup-idea validation.

Started as a single-source Google Trends wrapper; evolved into a multi-source signal aggregator with token-conscious defaults and a real eval suite, so it's something you can trust the output of, not just a demo. No database, no config file beyond optional API tokens for the tools that need them.

/market-signal — the actual product

The 9 raw tools below return structured data with no interpretation. The .claude/skills/market-signal/SKILL.md skill is where the judgment happens — and it's a conversation, not a one-shot command: it scopes the request, researches with the grounded tools plus live web search, writes a fixed-format SIGNAL REPORT so results are comparable across ideas and across sessions, then talks through what's real, what's noise, and whether it changes your plan.

/market-signal AI resume builder for Indian college students

Real output, from a live run against Trends + Wikipedia:

SIGNAL REPORT
════════════════════════════════════════════════════
Idea:            AI resume builder for Indian college students
Verdict:         MODERATE_SIGNAL
Confidence:      MEDIUM (3 sources, 12mo default window)

── Google Trends (interest_over_time + related_queries, India, 12mo) ──
"AI resume builder" holds steady in the high-30s/40s over the last 90 days
(most recent point isPartial - this week isn't finished yet, don't read the
dip as real). Top related query is generic ("resume ai free") - this space
isn't short on competitors already using the obvious framing. One rising
query, "jobsuit ai", spikes to an extreme relative value - worth a manual
check on whether that's a real emerging competitor or a data artifact.

── Wikipedia (Applicant_tracking_system pageviews) ──
Steady ~4,000-4,900 monthly views for most of the window, with a step up in
Dec 2025 (6,763) - no clear correlated spike with the Trends rising-query
signal. Reading interest in "how ATS works" isn't obviously accelerating
alongside search interest in resume tools.

Caveats:
- Default tier only - no Reddit/HN/Product Hunt signal in this run.
- The "jobsuit ai" rising-query spike is unverified from this data alone.
- Wikipedia's Sept 2026 figure (735) is a partial month, not a real drop.

Sources used:    interest_over_time, related_queries, wikipedia_pageviews
Tokens spent:    ~620 (default tier)
════════════════════════════════════════════════════
Pitch line: "Demand for AI resume tools is steady but crowded with generic
entrants - the real question worth digging into next is who's actually
complaining about existing tools, which needs the --deep tier to answer."

That report is still the fixed core of every run — but it's not the whole flow anymore:

• Scopes before researching. Asks at most 2 clarifying questions, and only when the request is genuinely ambiguous — a request that already states the decision, topic, and stakes gets no questions at all.

• Blends grounded data with live web search. The tools above give quantitative signal; native web search adds qualitative context (competitor mentions, news, forum chatter), cited, never turned into a fabricated confidence score.

• Structures a competitive landscape when there's one to show. Three or more named competitors surfaced by web search get grouped into a Differentiation Landscape (business-model archetype, funding, website) instead of loose prose.

• Handles multi-sided ideas entity-by-entity. An idea naming 2+ distinct impacted stakeholders (e.g. a marketplace's supply and demand sides) gets researched and verdicted per entity — capped at 3, never a numeric score — then closes with an elevator pitch for whichever entity has the strongest signal.

• Can render a shareable HTML version. Ask for something "presentable" or name a non-technical recipient, and the report publishes as a page instead of staying chat-only text.

• Stays out of pricing/segment/business-model territory. If a request drifts there, it hands off to /idea-validator or /office-hours when either is installed, or asks one pointed question and stops if neither is.

After the report, it opens a real back-and-forth — what's real, what's noise, does this change your plan — capped at 2 exchanges. Every run is appended to a local history file; researching a similar idea again surfaces a Prior check: line citing your earlier verdict, and a substantive answer to "does this change your plan" gets logged too, as outcome.

Add --deep to also pull Reddit + Hacker News/Product Hunt (community and builder-activity signal, at higher token cost — see Tools below).

Tools

All 9 tools accept response_format="concise" (default — truncated, rounded, token-conscious) or "full" (complete, unrounded). /market-signal calls them itself; you can also call any tool directly.

Default tier (no credentials needed):

interest_over_time(keywords, timeframe="today 12-m", geo="IN", response_format="concise")

Relative Google search interest (0–100) over time for up to 5 keywords. concise returns the most recent 90 days.

related_queries(keyword, timeframe="today 12-m", geo="IN", response_format="concise")

Top and rising related search queries. A rising value of 5000% is Google's "Breakout" marker (explosive growth from near-zero), not a literal percentage. concise returns the top 10 of each.

related_topics(keyword, timeframe="today 12-m", geo="IN", response_format="concise")

Same as related_queries, but topic clusters instead of raw query strings.

interest_by_region(keyword, timeframe="today 12-m", geo="IN", response_format="concise")

Search interest by state/region. concise returns the top 10 regions by interest.

trending_now(geo="india", response_format="concise")

Today's top trending searches for a country. Known limitation: currently fails with HTTP 404 — Google retired the legacy endpoint this depends on. Fails cleanly with a readable error string; the other 8 tools are unaffected.

wikipedia_pageviews(article, timeframe="P1Y", response_format="concise")

Monthly Wikipedia pageview counts — a free, no-auth reading/reference-interest signal that complements Trends' search-interest signal. concise returns the last 12 months.

--deep tier (requires free credentials):

reddit_signal(query, subreddits=None, limit=25)

Qualitative community signal via Reddit search. Requires a free Reddit app — create one at https://www.reddit.com/prefs/apps (type "script") and set REDDIT_CLIENT_ID / REDDIT_CLIENT_SECRET. As of 2026-09-07, Reddit requires manual approval for new apps (self-service registration closed in late 2025) — expect a wait, not instant signup. Fully optional: --deep works fine without it, and this tool degrades to a clear setup message rather than erroring if unconfigured.

builder_activity(query)

Builder/launch-activity signal: Hacker News (Algolia search, no auth) + Product Hunt (requires a free developer token — create one at https://api.producthunt.com/v2/oauth/applications and set PRODUCTHUNT_TOKEN). Both hn and product_hunt are always present as lists; HN works regardless of whether the PH token is set.

Opt-in (not part of either tier — called only when a query names a specific company):

company_registration(name, jurisdiction=None)

Company registration lookup (incorporation date, status, company number) via OpenCorporates. Scope note: registration facts only, not funding/valuation/traction data — no free API exists for that (Tracxn, Crunchbase, and similar are sales-gated enterprise products; we looked). Requires a free OpenCorporates API token — register at https://opencorporates.com/api_accounts/new and set OPENCORPORATES_API_TOKEN. Free tier is roughly 50 requests/day, 200/month.

All tools catch failures (rate limits, missing credentials, network errors) and return a plain error string instead of crashing.

Setup

Recommended — install as a Claude Code Plugin. Two commands, no cloning, no manual path-finding. Run these inside any Claude Code session:

/plugin marketplace add jain-eshan/market-signal-mcp
/plugin install market-signal

Start a new Claude Code conversation after installing — sessions already running won't pick up a newly added plugin. Check it connected:

claude mcp list

You should see market-signal listed as ✔ Connected with 9 tools. Then run /market-signal <your idea> (add --deep for the community/builder tier).

Optional — set any of these before installing if you want the corresponding tool to work (all degrade gracefully if unset):

export REDDIT_CLIENT_ID=...
export REDDIT_CLIENT_SECRET=...
export PRODUCTHUNT_TOKEN=...
export OPENCORPORATES_API_TOKEN=...

Alternative — manual clone, if you want to read or modify the code:

git clone https://github.com/jain-eshan/market-signal-mcp.git
cd market-signal-mcp
uv sync
claude mcp add market-signal -- uv run --directory "$(pwd)" server.py

Same verification and env vars as above apply to this path too.

Evals

This isn't just an API wrapper — it has a real eval suite, because "does the tool exist" and "does it produce a trustworthy answer" are different questions.

Both eval scripts run entirely through claude -p (Claude Code's own headless CLI, --mcp-config pointed at this repo's server.py) — no separate LLM provider API key, unlike an earlier version of this eval built on the mcp-eval framework. A tool meant to live inside Claude Code should use Claude Code's own capability, the same way this project's self-update check and other tooling lean on already-authenticated CLIs rather than holding credentials of their own.

CI runs the always-green contract + structural suites on every push: . The two claude -p-driven evals aren't in CI (a runner has no authenticated claude CLI) — run them locally from a logged-in machine.

Design notes

• Data only in the server, judgment only in the skill. Every tool returns raw structured data — the verdict, caveats, and pitch line all get decided in .claude/skills/market-signal/SKILL.md, not baked into server.py. This keeps the server simple and lets the skill's judgment evolve independently.

• Token-conscious by default. response_format="concise" (the default) truncates and rounds; the default research tier costs ~600-1000 tokens, --deep ~2500-3500 — both measured, not estimated, against real API responses.

• Self-updating awareness, not self-updating. A cached, throttled (24h) check compares your local VERSION against GitHub and tells you if a newer one exists — it doesn't modify your install.

• Remembers what you've researched. Every run appends to ~/.config/market-signal-mcp/history.jsonl; researching something similar again surfaces what you found last time.

• The guardrail hands off, it doesn't absorb. When a request drifts into segment/pricing/business-model territory, the skill either calls a real sibling skill (/idea-validator, /office-hours) or asks one question and stops — it never grows a second synthesis engine of its own.

Known limitations

• trending_now is broken upstream (see Tools) — not fixable here.

• company_registration covers registration facts only, never funding/traction/valuation data — no free API exists for that.

• The self-update check notifies only; it does not modify your local install.

• reddit_signal, the Product Hunt half of builder_activity, and company_registration's success path were built and their error/setup paths verified live, but their successful credentialed calls haven't been verified end-to-end — no API tokens were available while building this. If you set the corresponding env var and hit an issue, please file one.

• Reddit's self-service app registration is closed (see Tools) — getting reddit_signal working involves a manual approval queue, not instant signup. Don't count on it for a same-day setup.

• The tool-selection and output-quality evals (see Evals) are built and their parsing/harness logic verified, but not yet run end-to-end — the claude CLI needs a logged-in session, which wasn't available while building this.

• The guardrail's handoff to /idea-validator was verified live (it receives the drifted question correctly); the /office-hours handoff and the shareable HTML rendering (which needs the Artifact tool) haven't been verified end-to-end yet.

License

MIT — see LICENSE.