PSLRA Case Tracker
Reads Business Wire securities class action announcements through the Google News index, since Business Wire's own site refuses automated readers; used to collect headlines from that source.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@PSLRA Case Tracker refresh the tracker, then show deadlines due in the next 14 days"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
PSLRA Case Tracker
An open-source tracker for newly filed US securities class actions and their PSLRA lead plaintiff deadlines, with an MCP server so Claude can query it.
One securities class action is announced separately by eight to fifteen plaintiffs' firms. This tool reads those announcements from free public sources, recognises when many of them describe one lawsuit, and keeps one case per lawsuit with its ticker, class period, deadline, every firm that announced it, and the court docket when one can be found.
Runs with no API keys: extraction and classification are plain regular expressions. Optionally, set a TypeSafe key and the classification decisions are made by the Jev model instead (see Optional: Jev).
Not legal advice. Deadlines are read by regex from law firm press releases and can be wrong. Verify against the published notice or the court docket before relying on a date.
Quick start
uv sync # or: pip install -e .
uv run pslra-tracker run # one pass over the live sources (first run takes a few minutes)
uv run pslra-tracker cases --due-within 14run also writes output/report.html, cases.csv and cases.json. The record itself lives in SQLite at ~/.pslra-tracker/tracker.db (override with --db or PSLRA_DB).
Related MCP server: MyCase MCP Server
Connect it to Claude
Claude Code (this repo ships a .mcp.json, or add it anywhere):
claude mcp add pslra-tracker -- uv run --directory /path/to/pslra-tracker pslra-tracker serveClaude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"pslra-tracker": {
"command": "uv",
"args": ["run", "--directory", "/path/to/pslra-tracker", "pslra-tracker", "serve"]
}
}
}Remote connector (claude.ai and Claude Desktop "custom connector"): see Remote connector with sign-in.
Then ask things like "refresh the tracker, then which lead plaintiff deadlines fall in the next two weeks?" or "which securities dockets were filed this week that no firm has announced yet?"
Tools
Tool | What it does |
| One pass over the sources. The only tool that touches the network. |
| Cases by status, deadline window, ticker, company, or first-seen date. |
| One case in full: deadline votes, class period, firms, docket, every source announcement and how it was matched. |
| Every announcement ever read, including those set aside, with the decision made about each. |
| Federal securities dockets from CourtListener, matched or not yet matched to a case. |
| Each source's last success, last error and newest item reached, plus record counts. |
Remote connector with sign-in
Custom connectors are reached from Anthropic's servers, not from your computer, so the tracker needs a public https address and a login.
export PSLRA_AUTH_PASSWORD='a long passphrase'
pslra-tracker serve --http --port 8765 --public-url https://tracker.example.com--public-url is the address Claude will use. It switches on OAuth 2.1: Claude registers itself, opens a sign-in page in your browser, and you approve by typing the password. In Claude, add a custom connector with the URL https://tracker.example.com/mcp.
Put TLS in front (a reverse proxy, or a tunnel such as
cloudflared tunnel --url http://127.0.0.1:8765for testing; pass the tunnel's https address as--public-url).Access tokens last an hour and refresh automatically for 30 days. Only token hashes are stored, in the tracker's SQLite file.
PSLRA_API_TOKENoptionally sets a static bearer token for scripts (Authorization: Bearer ...).Without
--public-urlthe server only binds to localhost and has no login. It refuses to bind to any other address without sign-in.One password, one operator. There are no per-user accounts; anyone with the password can read and refresh the tracker.
Optional: Jev for classification
Regex is good at copying a date or a ticker out of a release. It is poor at deciding what a release is. With a TypeSafe key, two judgments go to Jev, a small model that returns a typed answer with a confidence rather than text:
uv sync --extra jev # or: pip install -e ".[jev]"
export TYPESAFE_API_KEY=...Judgment | Without a key | With Jev |
Filing, reminder, investigation, settlement or noise? | Keyword rules on the headline and body | One Choice per announcement; taken over the keyword rule at confidence ≥ 0.60 (≥ 0.80 to discard on the headline alone) |
What kind of case is this docket? | Caption keywords: SEC and derivative ruled out | One Choice on caption and cause of action; taken at ≥ 0.70, so short-swing 16(b) suits and the like are also ruled out |
Jev never extracts values. Dates, tickers and class periods still come from regex. Below the confidence floor, or on any API error, the keyword rule decides, so behaviour degrades rather than breaks. When Jev and the rule disagree, both verdicts are stored (search_announcements shows jev: investigation (0.95); regex said reminder).
Cost at the time of writing is $0.042 per million input tokens; a first run over a week of wires is a few cents. --no-jev or PSLRA_JEV=0 forces regex only.
Sources
All checked live on 5 October 2026.
Source | Kind | Default | How it is read |
PR Newswire | wire | on | Its own search page, phrase "securities class action". Highest volume, noisiest. |
Business Wire | wire | on | Its site refuses automated readers (HTTP 403), so it is reached through the Google News index. Headlines only. |
GlobeNewswire | wire | on | Its "class action" tag page. |
CourtListener | court | on | Free Law Project's archive of federal dockets, searched by filing date for nature-of-suit code 850. Works without a key; anonymous access is rate-limited, so the sweep pauses between pages. A free |
Only public distribution channels are read: the newswires every firm publishes through, and the court record. Individual law firm websites are deliberately left out.
Pick sources per run with --sources prnewswire globenewswire, skip the docket sweep with --no-courts.
How it works
Collect. Read each source's current listing.
Skip what is already read. A web address already in the record costs nothing further.
Keep real cases. Investigations, settlements and unrelated news are set aside, from the headline alone where possible. They are still recorded, with the reason.
Read the detail. Ticker, exchange, company, deadline, class period, firm, and the docket number when the release quotes one.
Match or create. Compare against existing cases, strongest rule first. Nothing below 0.85 is merged.
Rule | Confidence |
Ticker + same deadline | 0.98 |
Ticker + same class period | 0.96 |
Ticker + deadline within 10 days (recorded as a conflict) | 0.92 |
Ticker + same class period end | 0.90 |
Company name + same deadline | 0.88 |
Ticker, where one side has no deadline yet | 0.86 |
Company name alone | 0.75, never merged |
A case's facts are a majority vote over its announcements, so one firm's typo does not move the deadline, and disagreement is reported (deadline_conflict, deadline_votes, class_period_conflict).
Safeguards
Silence is not an answer. A source that cannot be reached is recorded as an error and retried next run. Its watermark does not move.
Nothing is lost. Announcements over a run's reading cap (
--max-fetch, per source) are deferred, not dropped. Bare headlines with no deadline never open a case; they are held and retried each run until their case exists.In doubt, do not merge. A weak resemblance opens a separate case. An announcement naming no company and no ticker is refused.
Impossible dates are dropped. A "deadline" inside the class period, or weeks before the release, is a misread.
One run at a time. A file lock stops two passes colliding.
Every decision is recorded.
search_announcementsshows what happened to each announcement and why.
Court dockets
A lawsuit appears on the docket the day it is filed, typically weeks before any press release. Two caveats are built in:
A filing code is not a classification. Code 850 also carries derivative suits, SEC enforcement and individual investor suits. Dockets are labelled from the caption and cover sheet only:
regulatory_enforcement,derivativeand (with Jev)not_class_actionare ruled out, and everything else stays asecurities_candidate, meaning not ruled out, not confirmed.classified_byrecords whether a keyword rule or Jev decided. Nothing reads the complaint.A docket does not state the deadline. The 60-day window runs from publication of the notice, not from filing. Where
get_caseoffers a date computed from a docket it is markedestimated.
Dockets are linked to a case by the docket number a release quotes, or by company name plus a filing date within 100 days before the deadline. Unlinked candidates stay visible through list_court_dockets.
Configuration
Variable | Purpose |
| SQLite file. Default |
| How the tracker identifies itself to sources. |
| Optional. Lifts CourtListener's anonymous rate limit. |
| Optional. Turns on Jev classification. |
| Required with |
| Same as |
| Optional static bearer token for scripts, HTTP mode only. |
| Keeps the key but forces regex only. |
Project layout
pslra_tracker/
sources.py newswire readers and the CourtListener docket search
extract.py regex extraction: ticker, company, deadline, class period, firm, docket number
judge.py optional Jev judgments, with regex fallback
pipeline.py collect, skip seen, set aside, extract, match or create
store.py SQLite record of announcements, cases, dockets and watermarks
server.py MCP server
auth.py OAuth sign-in for the HTTP transport
cli.py command line
report.py HTML and CSV report
tests/ unit tests and 19 real releases used as a benchmarkTests
uv run --extra dev pytest
uv run pslra-tracker run --from-json tests/data/real_items.json --benchmark tests/data/bench.txtThe suite runs offline with no keys (Jev is replaced by a stub). tests/data/real_items.json holds 19 real releases from September 2026; bench.txt lists the 16 true tickers. Current result: 16/16 cases, no false cases, the Pentair class-period discrepancy flagged.
Known limits
Extraction is regex with or without Jev. Company names are heuristic and sometimes truncated.
Business Wire items are headlines only, so they rarely carry a class period.
Press releases trail the filing; the docket sweep narrows that gap but its candidates need a human or a complaint-reading step before they count as cases.
Securities cases only. Other practice areas would need their own nature-of-suit codes and classifiers.
Sources change their pages. A layout change surfaces as a source error in
source_status, not as an empty result.Be polite: the tracker identifies itself (
PSLRA_USER_AGENT) and pauses between requests. Keep it that way.
Licence
MIT
Available Tools
6 toolsget_caseARead-only
Full detail for one case by id, or every case for a ticker: the deadline and how many announcements voted for each date, the class period, every firm that announced it, the linked court docket when one was found, and every source announcement with how it was matched.
Inputs: ticker, case_id.
list_casesARead-only
List tracked securities class actions, soonest lead plaintiff deadline first.
status: open (deadline today or later), expired, no_deadline_found, or all. due_within_days: only cases whose deadline falls within this many days. ticker: exact stock ticker. query: substring of the company name. first_seen_since: ISO date; only cases first seen on or after it (what is new).
Inputs: limit, query, status, ticker, +2 more.
list_court_docketsARead-only
Federal dockets filed under nature-of-suit code 850 (Securities/Commodities), from CourtListener. These appear the day a case is filed, typically weeks before any press release.
A filing code is not a classification: code 850 also carries derivative suits, SEC enforcement
and individual investor suits. classification is read from the caption and cover sheet only
(classified_by says whether a keyword rule or the Jev model decided), so securities_candidate
means "not ruled out", and even securities_class_action is unconfirmed until the complaint is read.
unmatched_only: dockets not yet linked to a tracked case (possible cases no firm has announced).
Inputs: limit, classification, unmatched_only, filed_within_days.
refresh_casesAIdempotent
Run one pass: read the newswires, set aside what is not a filed case, extract ticker / class period / deadline, and merge each announcement into its lawsuit.
sources: any of prnewswire, businesswire, globenewswire (default: all three). days: look-back window. max_fetch: cap on article bodies read per source this run; anything over the cap is deferred to the next run, not dropped. include_courts: also sweep CourtListener for securities dockets (nature of suit 850) and link them to cases. A source that cannot be reached is reported under its name with an error and retried next time.
Inputs: days, sources, max_fetch, include_courts.
search_announcementsBRead-only
Search every announcement the tracker has read, including those set aside, with the decision made about each (why it was set aside, or which case it joined and by which rule). kind: filing, reminder, investigation, settlement or other. source: e.g. prnewswire.
Inputs: kind, limit, query, source.
source_statusARead-only
Whether the Jev classifier is active, and every source with whether it is switched on, when it last succeeded, its last error, and the newest announcement reached, plus record counts. A source whose newest announcement has not moved for days while others are current has quietly stalled.
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Most tools have clearly distinct purposes: ingestion, case lookup, case listing, announcement search, docket listing, and source health. The main ambiguity is between get_case and list_cases, since get_case with a ticker returns every case for that ticker while list_cases also lists cases with ticker/filter options.
Five of six tools follow a clear verb_noun snake_case pattern: refresh_cases, get_case, list_cases, search_announcements, list_court_dockets. The only deviation is source_status, which is a noun phrase rather than a verb-led action, but the overall convention remains readable and consistent.
Six tools is well-scoped for a focused PSLRA case-tracking server. Each tool covers a distinct part of the workflow: ingestion, case detail, listing, announcement search, docket discovery, and source health.
The surface covers ingestion, lookup, listing, announcement search, court docket discovery, and source diagnostics, which is strong for the domain. Minor gaps remain, such as no direct tool to manually create or update a case, but the automated refresh and query tools cover the core lifecycle.
Maintenance
Related MCP Connectors
Search U.S. case law, fetch opinions, and ask matter-aware legal questions over your documents.
Judge-level filing rules, court holidays and enforcement data for US federal and state courts.
SEC filings — 13F, Form 4, 8-K, 13D, IPO and short-interest data — as 17 agent tools.
Search public U.S. federal litigation: companies, cases, dockets and document metadata.
Related MCP Servers
AlicenseAqualityDmaintenanceSearch and retrieve US federal court cases, dockets, claims, and documents via PACER — directly from Claude and other MCP-compatible AI assistants.1046 npmMIT- AlicenseNot gradedqualityCmaintenanceConnects Claude to MyCase legal practice management, enabling natural language queries for cases, contacts, calendar, billing, and more.101 npm6MIT
- AlicenseAqualityBmaintenanceEnables Claude to retrieve read-only public market data, including SEC EDGAR filings, FRED macro series, and end-of-day prices, with source and retrieval dates.11MIT
- AlicenseAqualityBmaintenanceEnables MCP clients to query Brazil's DataJud (CNJ) public court API with typed, validated, retry-safe tools: fetch a case by its CNJ number with check-digit validation and automatic court resolution, list a case's movements with plain-English hints, search cases by class, subject, court unit or filing dates, and list all 91 court indices. Responses are returned as structured content, making court case metadata usable from Claude Desktop, Claude Code, or a custom agent.41MIT