Gumshoe.

Private investigations for your pipeline · est. 2026

Send a question.
Get a closed case.

Gumshoe is an agentic research API. Point it at any URL with a question in plain English. It reads the page, weighs the evidence, and files a typed JSON verdict with the exact quotes it relied on. Sites that need clicking get a real browser, on request.

Free plan: 1,000 credits a month. No card. No small talk.

GUMSHOE INVESTIGATIONSCONFIDENTIAL
Case closed
Screen-printed poster of a private eye holding out a magnifying glass

Doc 2 of 8 · internal memo

You own a shovel.
That doesn't make you a detective.

Search APIs hand you links. Scraping APIs hand you HTML. Then the real work starts: reading forty tabs at 2 a.m., deciding what counts, formatting it for your CRM. You didn't buy research. You bought homework.

Gumshoe is the layer above the plumbing. You file a brief in plain English. The agent works the case across search and primary sources, and when a JS-heavy site slams the door it drives a real cloud browser that clicks, filters, and reads like a human. What comes back isn't raw material. It's a verdict.

Keep your search API. It's a fine shovel. Hire the gumshoe for the cases that matter.

Attachment 2-A · the usual suspects

SUSPECTM.O.VERDICT
Search APIsSerper & associatesHands you ten blue links and wishes you luck with the reading.Fine shovel
Scraper APIsFirecrawl & co.Delivers the raw HTML crime scene. Cleanup is your problem.Good tweezers
Spreadsheet empiresClaySells you a beautiful table and credits priced like jewelry. You still do the judging.Nice furniture
Deep research chatsClaude, Perplexity & co.Brilliant on one case, if you paste, wait, copy, and reformat it yourself. Nobody does that ten thousand times a night.Great sidekick
Gumshoethat's usWorks the case end to end. Returns the verdict with citations and photographic evidence.Hired

Intake desk · your deposition

You bring the hunch.
We write the brief.

Prompt engineering is not your job. Mumble what you want in one sloppy line to POST /v1/brief, or to draft_brief over MCP. Gumshoe reads it back as a structured brief: who to look at, which signals to check, what the report must contain, plus the schema. You approve it, or tweak a line, then open the cases. Drafting is free, no credits. Bad prompts can't sink good research when the agency writes the prompt.

WHAT YOU SCRIBBLE

“find me companies kinda like the ones that already buy from us and might be ready to switch”

CASE BRIEF · DRAFT FOR APPROVAL

TARGET   B2B SaaS, 50-500 employees
LOOK FOR · runs a competing tool
         · hiring RevOps or GTM eng
         · raised in the last 12 mo
SOURCES  careers, news, tech traces
DELIVER  fit verdict + confidence
         + top 3 signals, cited
Approve & open casesTweak a line

Doc 3 of 8 · the caseload

What people hire us for

Everything you'd wire a Clay table to do, minus the table. One API call each.

ICP qualification

Does Acme fit? Headcount, motion, market, stack. A verdict with sources, not a hunch.

REPLACES A 6-COLUMN CLAY WATERFALL

Hiring signals

Who is hiring SDRs, engineers, RevOps. Read straight from careers pages, not stale job boards.

REPLACES A CLAYGENT HIRING PROMPT

Funding intel

Raises, rounds, investors, dates. Cited from primary coverage, not rumor mills.

REPLACES A NEWS ENRICHMENT COLUMN

Tech stack recon

Which CRM, which cloud, which tooling. Assembled from public traces and product pages.

REPLACES 2 TO 3 STACK DETECTION TOOLS

Competitor watch

Pricing moved, feature shipped, logo landed. Re-file the same brief on your cadence and diff two verdicts.

CLAY DOES NOT EVEN DO THIS

List sweeps

Re-verify 10,000 accounts every Monday morning. One loop, one webhook, zero tabs.

REPLACES THE WHOLE TABLE

Doc 4 of 8 · standard procedure

How a case gets worked

Illustration of an open case folder with a paperclipped photograph

Step 1 · File the case

Brief the agency

POST /v1/case with a mission in plain English. “Does Acme run Salesforce? Are they hiring SDRs? Did they raise?” Attach your JSON schema if you want the report typed your way.

Illustration of a detective walking under a street lamp at night

Step 2 · Fieldwork

We walk the streets

The agent searches, cross-references, reads primary sources. When a site resists (logins, filters, infinite JS), it drives a real browser and takes timestamped screenshots as it goes.

Illustration of a hand stamping a document in red ink

Step 3 · Case closed

The report lands

Structured JSON hits your webhook: verdict, confidence, citations, exhibits. Pipe it into your CRM, your outbound tool, your warehouse. Every claim traceable to its source.

Doc 5 of 8 · dispatch board

File one case.
Or ten thousand.

This is not a deep research chat you babysit one tab at a time. One detective works one lead at a time. Gumshoe is a bureau: file a thousand briefs in a loop, a dozen investigators work every call concurrently, and each report fans back to your webhook the second its case closes.

Sweep your whole TAM before lunch. Qualify every signup from last night. Re-verify 10,000 accounts every Monday. One case is a question. Ten thousand cases is a pipeline.

TONIGHT'S SHIFT · LIVE

  • CASE 0114Does Acme still run Salesforce?CLOSED
  • CASE 0115Is Initech hiring SDRs this quarter?CLOSED
  • CASE 0116Did Globex raise a Series B?CLOSED
  • CASE 0117Who owns data engineering at Hooli?WORKING
  • CASE 0118Pricing page changes at Umbrella Corp?WORKING
  • CASE 0119Which CRM does Vandelay Industries use?QUEUED

9,994 MORE IN THE QUEUE · REPORTS FAN BACK AS THEY CLOSE

Screen-printed illustration of dozens of identical detectives typing reports in parallel rows of desks
THE BUREAU FLOOR, 2:47 AM · EVERY DESK IS ONE OF YOUR CASES

Doc 6 of 8 · evidence locker

Send the browser in,
get the photos back.

Standard cases read the page and quote it back to you. When a site needs real clicking, you flag fieldwork and a cloud browser drives it for you. These are actual frames from one of those runs: 13 steps through a JavaScript-heavy directory, shot by shot, timestamped.

Timestamped browser screenshot used as investigation evidence, EXHIBIT A · ycombinator.com · step 1 of 13
EXHIBIT A · ycombinator.com · step 1 of 13
Timestamped browser screenshot used as investigation evidence, EXHIBIT B · company profile · step 5 of 13
EXHIBIT B · company profile · step 5 of 13
Timestamped browser screenshot used as investigation evidence, EXHIBIT C · final source · step 13 of 13
EXHIBIT C · final source · step 13 of 13

WHAT THIS EXACT CASE RETURNED · 53 SECONDS, 13 STEPS

{
  "case_id": "case_yc_ai_sales",
  "brief": "Top AI sales companies in recent
            YC batches, with sites",
  "verdict": "5 companies identified",
  "findings": [
    { "name": "Palisade",
      "note": "AI sales agents for marketplaces",
      "url": "palisade.run" },
    { "name": "Quotain",
      "note": "AI sales simulations on real deals",
      "url": "quotain.com" },
    { "name": "AiSDR",
      "note": "AI sales prospecting",
      "url": "aisdr.com" }
  ],
  "exhibits": ["ex/a.png", "ex/b.png", "ex/c.png"],
  "steps": 13,
  "duration_seconds": 53
}

Doc 7 of 8 · the paperwork

One endpoint. Typed reports.

No SDK ceremony, no eight-step pipeline builder. A case is one POST. The report is JSON your warehouse already understands.

THE BRIEF · REQUEST

curl -X POST https://gumshoe.run/v1/case \
  -H "X-API-Key: gum_your_key" \
  -d '{
    "url": "https://sierra.ai",
    "brief": "Does this company sell
              autonomous AI agents?",
    "webhook": "https://you.com/hooks"
  }'

THE REPORT · WEBHOOK PAYLOAD

{
  "case_id": "case_0001",
  "url": "https://sierra.ai",
  "verdict": "yes",
  "confidence": 90,
  "summary": "AI agents platform: build,
              deploy, manage agents.",
  "evidence": [
    "Build powerful AI agents quickly,
     with or without engineering support",
    "Deploy a single agent across chat,
     SMS, WhatsApp, email, voice",
    "Ghostwriter builds a production-ready,
     multilingual, multichannel agent"
  ],
  "read_at": "2026-08-29T11:04:22Z"
}

Also in the file · bulk work

One call. Not fifty.

Batch: you already have the list.
POST up to 250 companies to cases/batch and they get worked a dozen at a time, one investigator per row, filling in the facts. Loop the call for the next 250.

Sweep: you don't have a list yet.
Give a brief and a target count. We fan out the searches, verify every candidate, dedupe server-side, and deliver unique companies. You never file the same request twice, and you never pay for a duplicate.

THE SWEEP · ONE CALL, 100 UNIQUE ROWS

curl -X POST https://gumshoe.run/v1/sweep \
  -H "X-API-Key: gum_your_key" \
  -d '{
    "brief": "Companies selling AI agents
              built on LangChain",
    "count": 100,
    "webhook": "https://you.com/hooks"
  }'

// We fan out, verify, dedupe.
// Up to 100 unique rows per call.
// Duplicates are never billed.
// Credits reserved up front,
// undelivered rows refunded.

Also in the file · agent-first

Your agents can hire ours.

Gumshoe ships as an MCP server. Claude, Cursor, or your own agent mounts the agency and gets file_case, get_report, list_cases, file_sweep and draft_brief as native tools. Your agent decides what needs investigating. Ours does the legwork and returns the closed case.

Read the MCP field manual

MOUNT THE AGENCY · MCP

{
  "mcpServers": {
    "gumshoe": {
      "url": "https://gumshoe.run/mcp",
      "headers": {
        "X-API-Key": "gum_your_key"
      }
    }
  }
}

Doc 8 of 8 · rate card

Pick your caseload

One credit = one URL investigated. LLM included, math included, surprises not included.

The meter, in plain English

  • One credit is one URL.
    Company, product page, directory listing: one credit each.
  • Every claim quoted from the page.
    Three quotes or ten: same credit. Depth is included.
  • Many URLs, many credits.
    Batches and sweeps just multiply the rows. You pay per URL investigated, never per fact returned.
  • Browser fieldwork: 42 credits.
    Deep research dossier: 15. Only when the brief calls for it.

SPECIMEN · ONE MONTH ON THE METER

"Does acme.com sell AI agents?"
  1 URL read ................ 1 credit

"Qualify last night's 250 signups"
  cases/batch, 250 URLs ... 250 credits

"Full dossier on sierra.ai"
  deep research: 4 pages read,
  2 outside searches ....... 15 credits

"Read the YC directory,
  it needs clicking"
  fieldwork, real browser .. 42 credits

TOTAL DEBITED ............. 308 credits

7,500

CREDITS A MONTH · ONE CREDIT, ONE URL

$29/mo

PER 1,000 URLS
$3.87
VS OUR ENTRY RATE
entry
BROWSER + DEEP RESEARCH
included
Get 7.5K a month

Move up or down a rung any month. Change takes effect the same day, prorated.

Every rung, at a glance

PLANCREDITS / MOPRICE / MOPER 1KREQ / MINYOU SAVE
Walk-in1,000$0—60no card
7.5K7,500$29$3.87120entry rate
20K20,000$74$3.703004%
50KPOPULAR50,000$175$3.506009%
100K100,000$329$3.2996015%
250K250,000$779$3.1296019%
500K500,000$1,449$2.9096025%
Bureaunegotiatedcustomunder $2.90dedicatedtalk to the chief

Every paid rung includes browser fieldwork, deep research and list sweeps. Move up or down any month, prorated the day you switch. The free tier runs standard cases only. REQ / MIN is the enforced ceiling on cases, and every response tells you where you stand: batch calls carry up to 250 URLs each, so the 50K rung admits 12,500 URLs a minute.

Exhibit D · full disclosure

Side by side, no squinting

Two honest ways to do this today: rent a spreadsheet empire, or glue search and scrape APIs together yourself. Here is exactly what each one costs you, in dollars and in evenings.

CLAY (CLAYGENT)DIY: SEARCH + SCRAPEGUMSHOE
1,000 URLS COST$150 to $250 in creditsCheap per call, plus the pipeline you build and babysit$3.87 at the entry rung, $2.90 at volume, pipeline included
WHAT COMES BACKEnriched cells in a spreadsheetRaw links and HTML you still have to read and judgeTyped JSON: verdict, confidence, quotes from the page
PROOF ATTACHEDSource linksWhatever you remember to logExact quotes, plus screenshots on fieldwork runs
REAL BROWSER ON JS SITESNoOnly if you build and host itYes, opt-in, 42 credits, screenshots attached
LIST BUILDING DEDUPYour problemYour problemServer-side, duplicates never billed
BUILT FOROps teams who live in tablesEngineers who enjoy the glue codeSystems that act on verdicts

Ran dry mid-month? · top-ups at the rate your rung already earned

Credits

+2,500

$9.67

at the 7.5K rate · less on every rung above

Credits

+10,000

$38.67

at the 7.5K rate · less on every rung above

Credits

+50,000

$193.33

at the 7.5K rate · less on every rung above

YOU PAY YOUR OWN RATE, THE ONE YOUR RUNG ALREADY EARNED YOU. NO SURCHARGE FOR RUNNING HOT. ONE CLICK FROM YOUR DASHBOARD, CREDITED IN SECONDS. YOUR MONTHLY CREDITS RESET EACH CYCLE, TOP-UP CREDITS NEVER DO: THEY SIT THERE UNTIL YOU SPEND THEM. RUNNING OUT EVERY MONTH? THE NEXT RUNG UP IS CHEAPER PER URL, THAT IS THE WHOLE REASON TO CLIMB.

When does the meter move past 1?

  • Fieldwork is a real browser.
    A cloud Chromium clicks buttons, fills forms, passes JS walls, and screenshots every step. The screenshots become your exhibits.
  • It only runs when you flag it.
    Set fieldwork: true on a case, or let your agent request it over MCP. Never triggered on its own, never billed silently.
  • Fieldwork: 42 credits. Deep research: 15.
    A browser run is the one thing here that costs real money to operate: a cloud Chromium, up to fifteen steps, a screenshot at each one. 42 credits, flat, however long it takes. A deep research dossier reads several pages, searches outside the site, and costs 15. No per-minute meter in the trunk.
  • Blocked standard case? Still 1.
    If a site stonewalls a standard case, the report says exactly what blocked it. Refiling with fieldwork is your call, never ours.

THE THREE RATES · MEASURED

// 1 credit - read the page
POST /v1/case
{ "url": "https://acme.com",
  "brief": "Do they sell agents?" }
DEBIT ...................... 1 credit


// 42 credits - the page needs clicking
POST /v1/case
{ "url": "https://portal.acme.com",
  "brief": "List the plans behind
            the filter",
  "fieldwork": true }
// Real browser, screenshots kept
// as exhibits. Up to 15 steps.
DEBIT .................... 42 credits


// 15 credits - the full dossier
POST /v1/case
{ "url": "https://acme.com",
  "brief": "Everything: pricing,
            customers, funding,
            hiring, positioning",
  "depth": "deep" }
// Several pages + outside searches.
DEBIT .................... 15 credits

Appendix · interrogation transcript

Questions, answered under oath

Q: Why not just use a search or scraping API?
A: Use both, for plumbing. They return raw material you still have to read, judge, and format. Gumshoe returns the finished conclusion with the receipts attached. Different job, different tool.
Q: How is this different from Clay?
A: Clay is a spreadsheet empire with credits priced like jewelry. Gumshoe is one API call, priced like infrastructure. If you live in tables, stay there. If you build systems, this is your layer.
Q: What's a browser case?
A: When a target site needs real interaction (search boxes, filters, JS walls), the agent drives an actual cloud browser and keeps the screenshots it takes along the way: they come back in the report as exhibit URLs. 42 credits instead of 1, and only when you set fieldwork: true. That premium is browser time, not margin: a real run is a dozen steps of cloud Chromium and it costs us close to a hundred times what reading a page costs. A run is capped at fifteen steps, so the price never moves. A standard case that gets blocked still costs 1 credit, and its report tells you exactly what blocked it so refiling with fieldwork is your call.
Q: What if I'm bad at writing prompts?
A: Then you're normal. Describe the job in one plain sentence and Gumshoe drafts a structured brief back: target, signals, sources, output shape. Approve it or tweak a line before anything runs. The agency writes the prompt so your sloppy one-liner still closes a clean case.
Q: Couldn't I just ask Claude or ChatGPT to research this?
A: For one company, sure, and it would do fine. Gumshoe exists for the other 9,999: same JSON schema every run, webhooks, retries, screenshot evidence, and a cost per case in cents. A chat is a conversation you supervise. Gumshoe is infrastructure your code calls.
Q: Do I bring my own LLM key?
A: No. The platform model is included in every case. Your bill is one number: cases.
Q: How many rows can one case return?
A: One credit buys one URL investigated: a verdict, a confidence score, and the exact quotes the agent relied on. Enriching a list? Up to 250 URLs per cases/batch call, one credit each, loop the call for more. Building a list from scratch? One sweep call with a target count up to 100: we fan out, verify, and dedupe server-side, billed one credit per unique row delivered. Sweeps reserve credits up front and refund every undelivered row.
Q: What if I run out of cases mid-month?
A: Buy a top-up from your dashboard, credited in seconds, no plan change and no call. You pay your own rate: a top-up costs exactly what your rung costs per URL, with no surcharge for running hot. Your monthly credits reset at the start of each cycle, so they are use-it-or-lose-it, but top-up credits never expire and are only spent once the monthly ones are gone. Top-ups sit on top of a paid rung: on the free tier, pick a rung first. And if you top up every cycle, look one rung up: more credits for less per URL is the only reason we ask you to climb.
Q: What are the rate limits?
A: They climb with your rung. Cases per minute: 60 on the free tier, 120 at 7.5K, 600 at 50K, 960 at 100K and above. Batch calls follow the same ladder and each one carries up to 250 URLs, so the 50K rung is 12,500 URLs a minute of headroom. Every response carries X-RateLimit-Limit, X-RateLimit-Remaining and X-RateLimit-Reset, so you never guess where you stand, and a 429 adds Retry-After in seconds. The full table is in the field manual, generated from the code that enforces it. Past those ceilings we move you to a dedicated cluster: tell us the volume and the latency you need.
Q: How do I know a webhook really came from you?
A: Every delivery carries X-Gumshoe-Timestamp and X-Gumshoe-Signature. The signature is an HMAC SHA256 of "timestamp.body", keyed with your API key. Recompute it on your side and compare: if it matches, the report is ours and untouched. We retry three times with growing gaps on a timeout or a 5xx, and we stop immediately on a 4xx because your endpoint has already made its mind up.
Q: How fast is a case?
A: A standard case closes in one to two seconds, measured end to end on real company sites. A deep research dossier takes about seven, because it reads several pages and searches outside the site. Browser fieldwork takes a minute or two: surveillance can't be rushed.