Insider Buy Signals

Verify any U.S. SEC filing claim, field-by-field, against the filing it came from. Every figure here — shares, price, dollar value, percent, stake size — traces to a parsed field in a real EDGAR filing and ships with a link to that filing on sec.gov. Records that can't be linked to a filing are dropped, never guessed: no scores, no unsourced numbers. We call this filing-verified. We also ran the obvious question — does insider buying predict returns? — as four sealed, pre-registered tests. I tried four times to prove otherwise, and four times the data said no; the full record is at https://www.tradestarinsider.com/what-we-tested. This repo is the MIT-licensed MCP server, client library, and worked examples for the TradeStar Insider API.

• Website: https://www.tradestarinsider.com

• MCP server (Streamable HTTP, no auth): https://mcp.tradestarinsider.com/mcp

• REST (OpenAPI 3.0.2): https://api.tradestarinsider.com/v1/openapi.json

• Live status: https://status.tradestarinsider.com

• What we tested (negative results): https://www.tradestarinsider.com/what-we-tested

• GitHub: https://github.com/TradestarV5/insider-signals

• PyPI: https://pypi.org/project/insider-signals-mcp/

• npm: https://www.npmjs.com/package/insider-signals-mcp

• Hugging Face dataset: https://huggingface.co/datasets/tradestarinsider/insider-signals-form4

• Kaggle dataset: https://www.kaggle.com/datasets/tradestarinsider/insider-signals-form4

• RapidAPI: https://rapidapi.com/tradestarinsider-tradestarinsider-default/api/insider-buy-signals1

Free vs paid

The data is the same on every plan. Every row is filing-verified and links to its source on sec.gov. Plans differ in history, volume and export:

The free tier serves the last 30 days of filings. Older filings are not hidden: a request for an older date tells you the filing exists and is on the paid plan — it never claims that no filing exists. The verify_* tools check any filing across the full history for free. The paid plan is $29 a month on RapidAPI: 100,000 calls a month and full history since 2024, plus the key for the paid MCP endpoint from /v1/mcp-key.

(On RapidAPI the free Basic plan is 100 calls a day, or 3,000 calls a month where the console allows only a monthly quota.)

There are two free ways in, each with its own limit: the open, no-key endpoint (mcp.tradestarinsider.com/mcp and /public/v1) enforces a hard 100 data calls per day per IP, resetting at 00:00 UTC — over the cap it returns a plain message (a tool result on MCP, HTTP 429 on REST), never a transport error. MCP handshakes and verify_* checks both behave as expected (handshakes are free; a verify counts as one call). RapidAPI Basic meters per key at the same 100 calls/day. Both serve the same 30-day free window.

Full pricing and a key: https://www.tradestarinsider.com/pricing

Related MCP server: SEC EDGAR MCP Server

Run the MCP server locally

The server is published as insider-signals-mcp — a tiny stdio proxy to the remote endpoint. Run it with no install using uvx:

uvx insider-signals-mcp

To wire it into Claude Desktop, add this to claude_desktop_config.json:

{
  "mcpServers": {
    "insider-signals": {
      "command": "uvx",
      "args": ["insider-signals-mcp"]
    }
  }
}

Or point any MCP client straight at the remote endpoint (no proxy needed): https://mcp.tradestarinsider.com/mcp (Streamable HTTP, no auth). The verify tools are the core: check an insider purchase, an activist 13D stake, an 8-K event, or a 13F holding against its source filing.

What the data is

Five filing-verified families from SEC EDGAR (plus federal contract awards from USAspending):

• Insider Form 4 purchases (SEC transaction code P) — cluster buys (two or more insiders buying the same company inside a window of seven calendar days or fewer) and buys large relative to the insider's existing holdings.

• Activist Schedule 13D stakes, 13F fund position changes, and recent federal contract awards.

Code P does not by itself prove an open-market buy. A purchase filed under P can be an IPO / offering / private-placement allocation, so every purchase record carries an offering_context flag (true = allocation, not open-market) and clusters carry a uniform_price_cluster flag marking a single administered price. Option exercises, 10b5-1 planned sales, and sub-threshold noise (below $50,000) are dropped. Every field we return is parsed from a real filing and ships with its sec.gov link. Records that can't be linked to a filing are dropped, never guessed. No scores, no unsourced numbers. We call this filing-verified.

What the data is not

It is not a prediction. We ran four pre-registered, sealed tests for whether insider buying precedes price moves over our own history, and all four returned NULL — no return edge distinguishable from base rate — so we do not claim one. The full record (question, method, sample, number, verdict per test) is public at What we tested. This is a clean, filing-verified lookup and verification layer over public purchases — you decide what they mean. It is also not real-time: ingestion is a once-daily weekday cron, so new filings surface next business day, not live.

Coverage

• Insider Form 4 buys: full history 2024-01-02 → 2026-10-09 (686 U.S. trading days), no non-holiday gaps (market holidays such as Presidents' Day carry no filings and are not gaps). Read the live number any time at /v1/health — it is the source of truth and advances every business day. Every served purchase is a filing-verified code-P purchase, never an open-market guess: records that can't be traced to a parsed SEC filing field are dropped, never inferred.

• Completeness: we capture 93.70% of in-window code P purchases (lower bound 87.28%); of the filings we skip, only 0.75% (95% CI 0.34–1.63%) turn out to be purchases we genuinely missed — the rest are sales, option exercises, amendments, and sub-$50k buys we exclude by design. (This recall figure is from a sealed, pre-registered 20-trading-day study, 2026-08-17 → 2026-09-14; see What we tested.)

Install

pip install insider-signals        # or, from a clone:  pip install -e .

Zero runtime dependencies — the client uses only the Python standard library.

Quickstart (free, no key)

The free tier answers on the open discovery surface with no key and per-IP rate limiting:

from insider_signals import InsiderSignals

api = InsiderSignals()                         # public base URL, no key

print(api.health()["datasets"]["insider"])     # coverage days + date range

for s in api.today():                          # today's filing-verified buys
    print(s["ticker"], s["value_usd"], s["filing_url"])

for s in api.clusters():                       # cluster buys (2+ insiders, ≤7 days)
    print(s["ticker"], s["cluster_insiders"], s["cluster_notional"])

# Verify a specific claim against the filings we hold:
v = api.verify("ETRA", "ORBIMED ADVISORS LLC")
print(v["status"], v["count"])                 # -> confirmed 2

See examples/rest_api.py for a runnable version.

Free tier vs. keyed tier

For the keyed tier, subscribe on RapidAPI and construct the client with your key:

api = InsiderSignals(base_url="https://<your-rapidapi-host>", rapidapi_key="...")
for s in api.by_ticker("GME"):
    print(s["insider"], s["shares"], s["price_per_share"], s["filing_url"])

Calling a keyed method without a key raises InsiderSignalsError with status == 403.

Worked examples — one real request and response per tool

Every response below is a real capture from the live public API, on the free tier (no key). The two free list tools (today, clusters) were re-captured on 2026-10-21 and show the free 30-day window in their coverage_* fields — a free caller sees exactly this. The verify_* examples are from an earlier capture (2026-09-23); verification is full-history on every plan, so their coverage_* fields reflect that capture date and are not clamped to the 30-day window. The live coverage window moves daily — read /v1/health for the current bounds.

today — today's insider buys

GET https://api.tradestarinsider.com/public/v1/signals/today

{
  "api_version": "v1",
  "endpoint": "signals/today",
  "rule2_guarantee": "Every record traces to a parsed SEC filing field and carries its source filing_url. Records without a filing link are dropped upstream and never returned.",
  "count": 0,
  "coverage_start": "2026-09-22", "coverage_end": "2026-10-20",
  "coverage_note": "Insider (Form-4 code P) signals collected 2026-09-22..2026-10-20. An empty result means no matching filing in this window — not absence at the SEC.",
  "total_matched": 0,
  "params": { "date": "2026-10-21", "limit": 25 },
  "signals": []
}

(An empty set is normal pre-market / on a non-trading day; it repopulates the moment new filings materialize.)

clusters — cluster buys (2+ insiders, one issuer, ≤7 days)

GET https://api.tradestarinsider.com/public/v1/signals/clusters

{
  "api_version": "v1",
  "endpoint": "signals/clusters",
  "count": 1,
  "coverage_start": "2026-09-22", "coverage_end": "2026-10-20",
  "coverage_note": "Insider (Form-4 code P) signals collected 2026-09-22..2026-10-20. An empty result means no matching filing in this window — not absence at the SEC.",
  "total_matched": 177,
  "params": { "limit": 1 },
  "signals": [
    {
      "filed_date": "2026-10-19",
      "ticker": "APLD",
      "issuer_name": "Applied Digital Corp.",
      "issuer_cik": "0001144879",
      "insider": "Jason Zhang",
      "relationship": { "director": false, "officer": true, "ten_pct_owner": false, "officer_title": "President" },
      "code": "P",
      "shares": 300000.0, "price_per_share": 18.86, "value_usd": 5658000.0,
      "transaction_date": "2026-10-16", "acceptance_datetime": "2026-10-19T07:06:33-04:00",
      "cluster": true, "cluster_insiders": 2, "cluster_notional": 6664686.12,
      "offering_context": false, "uniform_price_cluster": false,
      "accession": "0001900445-26-000006",
      "filing_url": "https://www.sec.gov/Archives/edgar/data/1900445/000190044526000006/0001900445-26-000006-index.htm"
    }
  ]
}

verify — confirm an insider code P purchase claim

GET https://api.tradestarinsider.com/public/v1/verify?ticker=ETRA&person=ORBIMED ADVISORS LLC

{
  "api_version": "v1",
  "endpoint": "verify",
  "coverage_start": "2026-01-20", "coverage_end": "2026-09-23",
  "coverage_note": "Checked SEC Form-4 code P purchases collected 2026-01-20..2026-09-23. Absence here means not in this window — not absence at the SEC.",
  "status": "confirmed",
  "matched": ["ticker/issuer", "person"],
  "filings": [
    {
      "ticker": "ETRA", "issuer_name": "Electra Therapeutics, Inc.",
      "insider": "ORBIMED ADVISORS LLC",
      "code": "P", "shares": 1000000.0, "price_per_share": 15.0, "value_usd": 15000000.0,
      "filed_date": "2026-09-23", "accession": "0000947871-26-000880",
      "transaction_date": "2026-09-19",
      "acceptance_datetime": "2026-09-23T18:04:11-04:00",
      "aff10b5One": 0,
      "is_amendment": false,
      "original_accession": null,
      "filing_url": "https://www.sec.gov/Archives/edgar/data/2088082/000094787126000880/0000947871-26-000880-index.htm"
    }
  ],
  "count": 2,
  "message": "Confirmed: 2 matching Form-4 code P purchase filing(s)."
}

A claim with no matching filing returns "status": "not_found" (in-window) or "out_of_coverage" (outside the collected days) — never a guess.

Purchase record fields. Every confirmed purchase filing — on both verify (by claim) and the by-accession fetch — carries the same four extra fields, so a consumer reads one shape regardless of how the row was pulled:

Any field that could not be resolved is null and explained under a partial_fields object (field → reason) on that record — a missing value is never silent. A record with everything resolved omits partial_fields entirely. Example:

"partial_fields": { "acceptance_datetime": "row collected before the acceptance stamp was captured" }

Form 4/A amendments. When an insider files a Form 4/A that corrects an earlier Form 4, verify returns the amended filing as the current record — the superseded original is not returned as a co-equal filing. If your claim cited the figure the amendment changed, the verdict carries a superseded object naming the original_accession, the amendment_accession, and the corrected figure, and the message says the figure was corrected by 4/A. The keyed amendments_recent (GET /v1/amendments/recent) lists these amendments newest-first, each with its original_accession and a changed_fields map of what the 4/A changed (before → after); an amendment whose original is outside our collected window is still listed, flagged original_in_corpus: false, with the before/after left null — never guessed.

verify_13d — confirm a Schedule 13D activist-stake claim

GET https://api.tradestarinsider.com/public/v1/verify/13d?subject_name=New Fortress Energy Inc.&filer_name=Strategic Value Partners, LLC

{
  "endpoint": "verify/13d",
  "coverage_start": "2026-06-22", "coverage_end": "2026-09-18",
  "status": "confirmed",
  "filings": [
    {
      "filed_date": "2026-09-18", "form_type": "SCHEDULE 13D",
      "subject_name": "New Fortress Energy Inc.", "subject_tickers": ["NFE", "NFEGP"],
      "filer_name": "Strategic Value Partners, LLC",
      "percent_of_class": 14.7, "aggregate_amount_owned": 19208710.0,
      "signal": "new_13d_bare", "stated_intent": "New 13D. No activist purpose stated.",
      "accession": "0001193125-26-395888",
      "filing_url": "https://www.sec.gov/Archives/edgar/data/1749723/0001193125-26-395888.txt"
    }
  ],
  "count": 1,
  "message": "Confirmed: 1 matching Schedule 13D/13D-A filing(s)."
}

verify_8k — confirm an 8-K material-event claim

GET https://api.tradestarinsider.com/public/v1/verify/8k?cik=1650648&company_name=4D Molecular Therapeutics, Inc.

{
  "endpoint": "verify/8k",
  "coverage_start": "2025-03-06", "coverage_end": "2026-09-22",
  "status": "confirmed",
  "filings": [
    {
      "cik": "1650648", "company_name": "4D Molecular Therapeutics, Inc.",
      "accession": "0001193125-26-277563",
      "filing_date": "2026-06-22", "event_date": "2026-06-17", "form": "8-K",
      "items": [{ "code": "5.07", "label": "Submission of Matters to a Vote of Security Holders" }],
      "filing_url": "https://www.sec.gov/Archives/edgar/data/1650648/000119312526277563/"
    }
  ],
  "count": 4,
  "message": "Confirmed: 4 matching 8-K filing(s) carrying the claimed item(s)."
}

verify_13f — confirm a 13F institutional-holding claim

GET https://api.tradestarinsider.com/public/v1/verify/13f?filer=BERKSHIRE HATHAWAY INC&issuer_name=ALPHABET INC&quarter=2026-06-30

{
  "endpoint": "verify/13f",
  "coverage_start": "2024-06-30", "coverage_end": "2026-06-30",
  "coverage_note": "Values are AS-OF quarter-end and up to ~45 days stale.",
  "status": "held",
  "filings": [
    {
      "filer_name": "BERKSHIRE HATHAWAY INC",
      "as_of_quarter_end": "2026-06-30", "filed_date": "2026-08-14", "lag_days": 45,
      "change": "added", "issuer_name": "ALPHABET INC", "title_of_class": "CAP STK CL A",
      "value_usd": 28157599351, "prior_value_usd": 15600071913, "ssh_prnamt": 78791167,
      "accession": "0001193125-26-352200",
      "filing_url": "https://www.sec.gov/Archives/edgar/data/1067983/000119312526352200/"
    }
  ],
  "count": 2,
  "message": "Held: 2 matching 13F position(s) reported as of the covered quarter."
}

Response shape

Every data route returns the same envelope; the rows live under signals, and the client's list-returning methods hand you that list directly. The verify* methods return the full envelope (you want the status and matched/filings fields).

List controls

Every list route (today, clusters, ticker, range, amendments/recent) accepts the same optional query parameters. The response echoes the ones you set under params and always reports total_matched, so you page rather than hit a silent cap.

Every listing row carries first_served_at — the UTC ISO-8601 instant that row first crossed our publish boundary (the atomic write that made it public). It is stamped once, per ingest batch, and never rewritten, so it is a stable anchor for as_of replay and for measuring publish latency. On rows materialized before this field existed it is null, explained under partial_fields (and such rows are kept by an as_of query, since they were already public before the field). as_of is the companion to days: days bounds the filing date, as_of bounds when we served it.

MCP server (for AI assistants)

Ask an AI client (Claude and other MCP clients) for these signals directly. Open, free, per-IP rate limited — no API key, no OAuth:

claude mcp add --transport http insider-signals https://mcp.tradestarinsider.com/mcp

Then ask: "Show today's insider buy signals" or "Any insider cluster buys this week? Link the filings." See examples/mcp_client.md for the tool list and a programmatic SDK example.

Pricing

RapidAPI is the checkout. Two plans: Basic (free) and one $29-a-month plan with 100,000 calls a month and full history since 2024. After subscribing, call /v1/mcp-key to get your key and both keyed URLs — the key also unlocks the paid MCP endpoint.

Hard limits, no overage — requests over the cap return 429 until the window resets. The free discovery surface and the MCP server are open and per-IP rate limited; metered API billing runs through RapidAPI.

Disclaimer

Information, not investment advice. Every figure is parsed from a public SEC filing and links to its source on sec.gov. Insider buying guarantees nothing — insiders are wrong plenty — but code P purchases are real money on the line.

License

MIT. Changelog: CHANGELOG.md.