Free for lawyers

Paste a brief or drop a .docx (read in your browser, never uploaded); every case citation in it is checked for existence against 18,123,788 citations across 989 reporters and the Texas law index, with the case name where the citation graph knows it and, where we hold later-treatment data for that library, a flag when a later opinion said it overruled the case.

Up to 50 distinct citations per run, 50,000 characters, case citations only. Existence only: unverified is never reported as fabricated. Nothing is stored. Below the verdicts: a table of authorities you can paste into Word, and form notes on reporter spelling, years and pin cites: deterministic checks, not a Bluebook review. Any citation has its own page at /cite/<citation> with what we hold on it and, where we hold later-treatment data for that library, any later opinion that said it overruled the case.

The same existence check for one citation at a time, over HTTP with no key: POST /api/public/citation-check {"citation":"384 U.S. 436"}.

One citation per call, rate-limited per IP.

One page per citation at /cite/<citation>: everything we hold on that authority: the case name where the citation graph knows it, which library it was found in, and, where we hold later-treatment data for that library, any later opinion that said it overruled the case. The page is resolved by the same function the MCP citation_lookup tool calls, so the screen and the agent surface cannot disagree about what we hold.

Free, and answered only from our own library: the page never spends a CourtListener lookup, so a verdict is "found" or "unverified" and never "nonexistent". A treatment flag means a later opinion said it overruled the case; the absence of a flag is not a clean bill. A page for a citation we cannot vouch for is kept out of search engines. The "does it support your sentence?" box appears only for a signed-in reader, and only on a host that has switched it on.

Draft time entries derived from the work you did here (research, citation checks, contract review, chat on a matter), reviewed against each client's billing guidelines, exported as CSV, LEDES 1998B or Clio's import file, or pushed to Clio.

Free. Entries are a floor on time spent on this platform, not an estimate of your day; you edit and approve before export.

Paste a client's outside-counsel guidelines, or drop the .txt or .docx (read in your browser, never uploaded), and the billing rules in it are read out: the increment time must be recorded in, a daily hours cap, a ban on lumping several tasks into one entry, the phrases that client will not pay for, a requirement that every entry name the task, the participants and the purpose, and rate caps by role. Each proposal shows the client's own sentence and where it sits in their text, and saving it makes those the rules every later pre-bill review enforces.

Free. Patterns over the client's own sentences, no model and nothing invented: a sentence that reads like a billing rule but maps to no field is listed with the reason rather than dropped, and two sentences that set the same field two different ways are refused as ambiguous rather than tie-broken. Rate caps are saved and shown but never enforced, because a time entry carries no rate or role to check them against.

A case file is pseudonymized on our hardware before any frontier model sees it; the names come back in the answer and the document restores byte-exact. The mapping is encrypted per matter and never returned by any route. Three detectors propose spans and code does every replacement: our own patterns, our own small model, and openai/privacy-filter: open weights under the Apache 2.0 licence, 1.5 billion parameters, a token classifier over eight kinds of personal data, running on a Mac mini we own with no OpenAI key, account or API call anywhere in the path. The scrambler page carries the filter's real state on the host you are reading, because it is a per-host switch. In the browser at /scramble, or by API key.

Free: it runs on our own hardware, so there is no model bill to pass on. Measured 2026-10-08 on 167 adversarial filings carrying 3,241 planted secrets: the pattern pass alone left 128.7 per 1,000 of them in the text (95% interval 117.6 to 140.6), the pattern pass plus the PII filter left 35.8 per 1,000 (29.9 to 42.8) at 2.9 seconds a filing, and with our own model in the loop as well the same set measured 7.8 per 1,000 (5.3 to 11.5); no citation was destroyed and no filing refused in either new arm. On a 264-span set the pattern pass finds 75.4% of names and identifiers, the filter 65.9%, the two together 90.2%. Organisations are the weakest: the filter has no label for one. Placeholder substitution is not anonymity, because residual facts can re-identify people, and the filter's own authors call it “a redaction and data minimization aid, not an anonymization, compliance, or a safety guarantee”. Both sentences are on the page.

Paste a contract or a clause and pick a state; each clause is read against that state's verified contract-law rules, each rule quoted from the statute or opinion it rests on, and the review says what the rule is, what the trap is, and what it could not verify. In the browser at /contracts, or by API key.

Free: part of the free package by decision, not a trial. The review's own model spend is under a cent and counts toward a key's spend cap like any call. States are added only after their rules pass the audit bar.

Write your own position on a clause family (required, forbidden, or must say this phrase) and every clause the review found is read against it: meets, conflicts, or not addressed. Inline for one request, or saved to your account and applied to every review after it. The verified rule and the trap for that family stay beside your own verdict, unchanged.

Free. A verdict is decided by exactly two things, the rule you wrote and the text of the one clause it is checked against, with no model consulted, so at most 40 rules, a phrase of at most 80 characters, matched whole-word. A family the review never found is "not addressed" rather than guessed at, and "conflicts" means your phrase did not match: it says nothing about whether the clause is enforceable.

Every dated obligation the agreement sets for itself, taken out of the text you pasted ("within 30 days after termination", "at least 60 days before the end of the Term", business days), each with the clause's own sentence and the words that fixed the period marked in it. Type the date a period runs from and that line becomes a date, with the counting rule written out beside it rather than hidden behind a disclosure.

Free, with no model and no network: it answers even on a host where the review itself is dark, though reaching it still means signing in, because it sits on the review page. Nothing is dated until you supply the trigger, and only by its own trigger. A period in months or years is recognised and deliberately not computed, because the end-of-month convention is something parties litigate. "Promptly", "within a reasonable time" and a notice period that fixes no end are refused with a reason. No court rule is applied: these are the agreement's own days.

For each clause family (non-competes, liquidated damages, choice of law, warranty disclaimers and the rest), every state we answer, with the rule, the trap and each authority's verbatim quotation and source link.

Free. Only rules that passed verification are shown, and each one says whether it has been re-read at the current audit bar.

Two to four states side by side on one clause family, each column carrying that state's rule, the trap and every authority quoted word for word: the same verified entries the full survey and the per-state pages show, rendered through the same components, so the three views cannot drift apart.

Free. Two to four states in a view, because a wider table stops being readable on a phone. A state you name that has no verified rule for that clause gets a column saying exactly that rather than being dropped from the comparison, and a state code that is not a real jurisdiction is a 404 instead of an empty column.

Drop or paste two versions of an agreement (.docx is read in your browser, never uploaded) and see every changed passage, deletions struck and insertions marked, each labelled with the clause family it sits in, under a summary of what changed by family. The comparison is arithmetic over the two texts, run in your browser; no model reads them and nothing is sent to a server.

Free. It is a text comparison, not a review: it says what changed and where, not whether the change is good for you. Run the new version through contract review for that. A changed passage the clause classifier cannot place is reported as unlabelled rather than hidden, and a moved paragraph is shown as moved only when it is long enough to be sure of.

Pick a rule set we hold verbatim (Fed. R. Civ. P. 6(a), Fed. R. App. P. 26(a), Tex. R. Civ. P. 4, Wash. CR 6(a), Wash. RAP 18.6(a)), enter the trigger date and the period, and read the computed date with the rule's own words beside each step.

Free. Weekends are applied as the quoted rule says; court holidays are not, because we do not hold the holiday statutes, and the page says so. It does not know local rules, standing orders or service additions.

Paste federal docket entries, one per line or as a JSON array, and the six period rules we hold word for word (Fed. R. Civ. P. 12(a), 59(b) and 59(e), Fed. R. App. P. 4(a)(1)(A), Tex. R. Civ. P. 99(b) and 329b) are applied to the entries that match them, each deadline carrying the rule's own words and the entry it was derived from.

Free, and the paste path runs entirely in your browser: nothing is uploaded and no account is needed, and it is the path that works on this host. An entry it cannot classify is listed with the reason, never dropped silently. Where a period runs from service the service date is yours to type, because a docket entry records a filing date and not service. Holidays are not applied. Texas state deadlines only ever come from pasted events, because the federal archive holds no Texas state dockets. Reading a docket by its CourtListener id is off here until the data agreement is in place.

Pick which model answers, from the list this deployment allows. It is the same allow-list the chat route itself enforces, so a model the picker offers is a model the route will actually call, and every answer names the model that produced it.

Choosing is free; the model call it makes is metered on the rate card like any other. The picker offers only what this host allows, so the list is shorter on some hosts than others, and a model that is not on it comes back as a plain refusal rather than a quiet substitution. No GPT or OpenAI model is ever offered, and one would be refused even if a configuration named it.

Your own firm's numbers on the workspace: hours and entries by matter over 7, 30 and 90 days, the entries the pre-bill review flagged, time already exported against time not yet exported, matters with no activity for 30 days, and active docket watches with the filings they received. Every figure is a count or a sum over your account's own time entries, matters and watches, with the window it covers stated beside it.

Free, and nothing here is projected, averaged into a forecast or written up as an insight. "Billed" means one measurable thing only, that the entry was exported to CSV, LEDES or Clio, not that an invoice was sent or paid. A tool that is switched off on this host contributes nothing and says so by name, rather than a zero that would read as a real measurement.

Four choices kept on the server instead of in one browser: which jurisdiction grounds chat, which model answers, which matter is open, and which rule set the deadline calculator starts on. They follow your account across sessions and devices, and the browser's own copy is only a cache the stored value overwrites.

Free. There is no settings screen: each choice is saved by the surface that makes it. Exactly those four keys are stored and nothing else, an unrecognised key is a 400 that names it rather than being quietly kept, and a GPT or OpenAI model id is refused here the same way the chat route refuses it, so this can never hold a model the route would not call.

Share one matter with the organisation active in your session, explicitly, one matter at a time, read-only unless you grant edit. Members of that organisation see the matters shared with them and nothing else, time entries on a shared matter are visible as part of it and writable only by whoever created them, and revoking a share removes the visibility at the next read.

Free. Sharing is done from the workspace page, where a matter shared with you is listed under "Shared with your firm" with its time entries, or through the API. A share to any organisation other than your session's own is refused, and an API key has no organisation at all, so a key never reaches a shared matter.

Every model and its price, the MCP server's tool list, and the published open datasets of the law we serve: 1,134,330 statute sections across 25 states and 21,062 court rules across 28 states as of 2026-09-11, downloadable from https://huggingface.co/docketx.

Free. Model calls themselves are metered; discovery is not.