edgrapi

US government data as clean JSON. Zero dependencies.

pip install edgrapi

Nothing else gets installed. No requests, no pandas — the client is standard library only, so it drops into a Lambda, a container, or someone else's dependency tree without an argument.

from edgrapi import Client

c = Client("edgr_...")                 # or set EDGRAPI_KEY
c.holdings("berkshire")                # 13F, CUSIP-aggregated, diffed QoQ
c.insider("AAPL")                      # Form 4, code P separated from 10b5-1
c.clusters(days=15, min_insiders=3)    # several insiders buying at once
c.search("climate risk", forms="10-K") # full text, every filing since 2001

Or without writing any Python at all:

export EDGRAPI_KEY=edgr_...
edgrapi holdings berkshire --changes
edgrapi clusters --days 15 --min-insiders 3
edgrapi insider AAPL --limit 5 | jq '.transactions[0]'

Get a free key — 100 credits every month, no card — at edgrapi.com/app.

Why this exists

The SEC's data is free. Parsing it correctly is the work, and most of the bugs are silent:

• A 13F lists one holding once per sub-manager. Berkshire's latest filing has 89 rows and 29 actual positions — Apple alone appears 12 times. Read the table as it arrives and you triple-count the portfolio.
• A putCall row is a bet against the stock, not a holding. Every quarter someone reports a fund is "long" a name it is actually short.
• Filings before 2023 report value in thousands; newer ones in whole dollars. Diff across that boundary without normalising and every position looks like it grew 1000×.
• Share classes are separate CUSIPs. Alphabet is 02079K305 (Class A) and 02079K107 (Class C). Merging them is as wrong as splitting Apple.
• XBRL tags drift between filers and across years, so the same line item arrives under a dozen different names.

holdings() handles all of that and labels every position new / added / reduced / exited against the prior quarter, so you get what changed rather than one quarter's snapshot.

What you can call

US government data (same key):

Anything added after this release is reachable with c.get("/v1/whatever", param=...).

Credits, errors and retries

Every response is the API's JSON dict with billing attached:

r = c.insider("AAPL")
r["count"]            # the data
r.credits_cost        # what this call was billed
r.credits_remaining   # balance afterwards

Calls cost 1, 2, 3 or 5 credits depending on the endpoint — a light lookup is 1, an XBRL parse is 3, pulling a full document is 5. A hard error refunds in full. A valid response that carried no rows keeps a 1-credit floor, on the endpoints that report a count.

Errors are typed, so you can branch on them:

from edgrapi import Client, OutOfCredits, RateLimited, EdgrapiError

try:
    c.holdings("berkshire")
except OutOfCredits as e:
    print("needed", e.endpoint_cost, "credits")
except RateLimited as e:
    time.sleep(e.retry_after or 5)
except EdgrapiError as e:
    print(e.status, e.error, e.request_id)

429s and 5xx are retried automatically with exponential backoff, honouring Retry-After. A 404 is never retried — it will not start working.

Worth knowing

• holdings() takes a fund, not a ticker — a name like berkshire, a CIK, or a filer ticker.
• activist() and formd() accept "latest" for the market-wide feed.
• Only open-market purchases (code P) mean an insider spent their own money. Option exercises and vesting also appear on Form 4 as acquisitions but say nothing; most sales are pre-scheduled 10b5-1.
• Data is public-domain US government data (SEC EDGAR, SAM.gov, USAspending, Grants.gov, Congress). Nothing here is investment advice.

Also available

The same data is exposed as a hosted MCP server at https://api.edgrapi.com/mcp, so an agent in Claude, Cursor or Cline can call it directly:

{ "mcpServers": { "edgrapi": {
    "url": "https://api.edgrapi.com/mcp",
    "headers": { "Authorization": "Bearer edgr_YOUR_KEY" } } } }

Full REST docs at edgrapi.com/docs · OpenAPI at api.edgrapi.com/openapi.json

License

MIT