Claude for Legal shipped without a personal-injury plugin

I spent this week reading the entire claude-for-legal repository — every file in every directory, every plugin description, every skill markdown. Five hundred and sixty-five files. The reason I read all of it is that the marketing language around Claude for Legal has gotten loud, especially in the personal-injury corner of legal-tech Twitter, and the loud version of a story is often a different story than the file-level one.

Here is what the file-level story actually says.

There is no personal-injury plugin in Anthropic's claude-for-legal. Twelve practice-area plugins shipped on May 12, 2026 — commercial, corporate, employment, privacy, product, regulatory, AI governance, IP, litigation, legal clinic, law student, and legal builder hub. Plaintiff-side personal injury, the largest category of contingency-funded U.S. practice, was not one of them.

What Anthropic actually shipped on May 12, 2026

The repository at github.com/anthropics/claude-for-legal went live on May 12, 2026. By the time I finished reading it this week it had 7,767 stars, 1,319 forks, and 51 open issues — substantial early traction for a vertical open-source release. The license is Apache-2.0. The contributor list is short: four Anthropic employees, with Matt Piccolella driving the commits and Mark Pike, Anthropic's Associate General Counsel, providing the legal direction.

Pike is also the public face of the launch. He's quoted in Bloomberg Law calling Claude for Legal "a better together story" with the rest of the legal-tech ecosystem; in his Artificial Lawyer interview he describes the connectors and practice-area plugins as "available to all paid Claude customers." Plain reading: the Apache-2.0 code is the free layer, but the revenue motion is Claude for Enterprise, Claude Cowork, and Claude for Microsoft 365.

At launch, three BigLaw firms were named on the record as production users — Freshfields, Quinn Emanuel, Holland & Knight — plus a startup-focused boutique called Raines LLP. The use cases Anthropic highlighted are M&A diligence, contract redlining, privacy impact assessments, outside-counsel oversight, and regulatory monitoring. Every single one of those is BigLaw transactional or in-house counsel work. Personal injury isn't mentioned in Pike's interviews. It isn't mentioned in the May 15 deployment blog. It isn't mentioned on the claude.com/solutions/legal page.

This is not a complaint. It's an observation. Anthropic shipped what they shipped. The problem is the discourse around the release has run well ahead of the actual product, and PI lawyers are reading marketing that, on careful reading, was not built for them.

The PI-shaped hole

I ran exhaustive greps across the whole repo for the vocabulary a personal-injury practice actually uses. Here is what came back.

The litigation-legal plugin's claim-chart skill ships with a generic Negligence civil template — the four classical elements from Restatement (Second) §§ 281 and 328A plus CACI 400. That is the only PI-adjacent civil template in the repo. No Premises Liability template. No Products Liability template. No Medical Malpractice template. No Auto Negligence template. No Dram Shop template. No Wrongful Death template.

What does appear is the vocabulary of plaintiff work, scattered through the calibration files. The cold-start-interview skill defines plaintiff as a side that routes calibration vocabulary — "asserting, case value, contingency, SOL cliff." A solo-practice path exists in the practice-profile setup. PI statutes of limitations appear, but only in the legal-clinic plugin's plausibility-band reference files for California and Illinois. So Anthropic knew plaintiff PI exists. They knew the vocabulary. They built scaffolding that could support a PI plugin. They just didn't build the plugin.

The litigation-legal README is honest about its primary audience: "In-house litigation counsel support for managing a portfolio of matters." Outside-firm associate use and solo-contingency use are listed second and third. That ordering matters.

Why Anthropic skipped PI

A few hypotheses, and one of them seems more right to me than the others.

The first is funding gravity. The open-source legal-tech ecosystem in the United States is shaped largely by Legal Services Corporation grants, which target self-represented litigants and pro se access-to-justice work. The largest OSS legal projects — Docassemble, Suffolk LIT Lab's AssemblyLine, A2J Author, CourtFormsOnline — were all built for that population. Personal-injury plaintiffs are represented on contingency, which means LSC doesn't fund OSS for them, which means there's no twenty-year pipeline of templates and interview banks for Anthropic to extend. They released into a vacuum.

The second is that medical-records workflow is a research problem, not a markdown-prompt problem. A real PI demand letter ingests medical records that routinely run eight thousand to fifteen thousand pages, in OCR-scanned PDF form, with provider-specific formatting, ICD codes that need cross-referencing, and treatment narratives that need timeline reconstruction. That isn't a Claude Code skill. That's a multimodal pipeline with OCR, classification, deduplication, and chronology synthesis — closer to what DigitalOwl, Supio, and EvenUp's MedChrons product spend engineering quarters building. A markdown prompt pack is the wrong vehicle.

The third hypothesis — the one I think is right — is that Anthropic shipped where their existing customers were. The named partners at launch — Freshfields, Quinn Emanuel, Holland & Knight — are AmLaw 100/200 transactional and litigation defense shops. They were already paying Anthropic for Claude Enterprise seats. They wrote the requirements that became the twelve plugins. Personal injury, by contrast, is dominated by three- to twenty-five-attorney plaintiff shops without legal-ops engineering and frequently without dedicated technology staff. They aren't Anthropic's BigLaw enterprise pipeline. So they aren't in the plugin list.

This isn't malice. It's gravity. The result is the same: the gap is real, and PI lawyers reading the launch coverage are reading about a product that wasn't built for them.

What you can actually do with claude-for-legal today as a PI lawyer

I want to be fair about this. The litigation-legal plugin has skills a plaintiff PI lawyer can use today, with some setup. They're not PI-tuned. They're not useless either.

• /litigation-legal:demand-intake — pre-drafting context gathering for a demand letter. Parties, facts, basis, leverage, BATNA. Writes a structured intake.md the demand-draft skill reads.
• /litigation-legal:demand-draft — drafts the letter from the intake, gated on a privilege / FRE 408 / waiver / admission checklist. Word document output. Honest about being a draft for attorney review.
• /litigation-legal:chronology — extract dated events from a production or matter file, de-dupe, tag by significance per the matter theory. Useful for organizing a case file once you have it.
• /litigation-legal:claim-chart — element-by-element chart for any cause of action, with pin-cited cells and gap detection. Generic civil mode only — no PI-specific templates.
• /litigation-legal:deposition-prep — build a deposition outline for a witness, organize topics around the case theory, surface impeachment material.
• /litigation-legal:matter-intake — uniform questions covering identification, conflicts, source, risk triage, materiality, owners, legal hold, key dates. Writes matter.md and history.md.

I ran the demand-draft skill on a sample case file over a weekend — fictional MVA, soft-tissue, $35,000 in medicals — to see what came out. The draft was structurally reasonable. The FRE 408 gate is genuinely useful: the skill won't draft a number unless you've confirmed in the intake that you have authorization to make a settlement demand. The privilege and waiver checklist before drafting is a thoughtful guard rail. The output is a Word document with citation placeholders and a post-send checklist. For a draft to start from, it works.

What it does not do is the thing that makes EvenUp worth what it costs. It doesn't ingest twelve thousand pages of medical records and produce a chronology with exhibit links. It doesn't pull comparable verdicts from a proprietary database. It doesn't have a damages model trained on the specific ICD codes you're submitting. It doesn't write in your firm's house voice unless you've spent twenty-plus hours loading sample demands into the cold-start interview. And it doesn't push the final document to Filevine or Litify or CASEpeer when you're done.

What's missing — and the build cost

Here is the inventory of what a PI firm running claude-for-legal would need to build themselves before the system credibly replaces an EvenUp subscription.

1. Medical-records ingestion at ten-thousand-plus-page scale. The Claude Sonnet 4.6 context window is 200K tokens — call it 400 pages of typed legal-density text. Real PI records, OCR'd from copy-of-a-copy-of-a-fax, are denser and more expensive per page. You're looking at building a chunking and retrieval pipeline on top of a vector store before the model ever sees the medical narrative.
2. Medical chronology synthesis. DigitalOwl and Supio have spent years on this specifically. The OSS substrate exists — Bio_ClinicalBERT, llm-ie for clinical extraction, LLMAnonymizer for HIPAA de-identification — but stitching them together for production litigation use is genuinely a quarter of engineering work.
3. Settlement-comparable database. EvenUp's Settlement Repository is a proprietary asset they've crowdsourced from 2,000+ customer firms. You don't get to use it when you leave them, and you can't replicate it from public verdict-research datasets (none of which are open).
4. Lien tracking and Medicare set-aside calculation. Garretson and Archer Systems own this category. No OSS exists.
5. Practice-management integration. Anthropic ships zero connectors for any PI-dominant case-management system. Filevine, Litify, CASEpeer, CloudLex, SmartAdvocate — every integration is custom code against the vendor's REST API. There is exactly one community MCP server for MyCase and four for Clio, and that's it.

The realistic one-time cost to put this together for a five-attorney plaintiff shop is somewhere between fifty thousand and one hundred fifty thousand dollars. Ongoing infrastructure and engineering maintenance is another thirty to seventy thousand a year. That's before you've drafted your first demand letter.

The $0.60 demand letter that costs $50,000 to ship

The token math, on the other hand, is dramatic. A 350-page medical-record input plus a 4,000-token demand draft on Claude Sonnet 4.6 — Anthropic's standard production model at $3 per million input tokens and $15 per million output — costs about $0.60. The same demand from EvenUp is reported by three competitor sources to land between $300 and $800 effective per demand. At seventeen demands per month for a 200-case-per-year firm, that's $5,100 to $13,600 a month on EvenUp versus, in raw model spend, about $10 on claude-for-legal.

Save the math for the build sheet.

The naive read says claude-for-legal wins by two orders of magnitude. The honest read is more complicated. What EvenUp is selling at $500 a demand is not "an AI demand letter." It's a medical chronology built by a team of human drafters working from a proprietary records pipeline, plus a comparable-verdicts query against a database they've built over four years, plus a malpractice insurance wrapper, plus an integration with your case-management system, plus a five-minute call with a real human when you need to push back on a draft. The AI is the marketing wrapper. The product is the operation behind it.

EvenUp isn't a villain in this story. At $300 a demand they ship a real product to PI lawyers who don't want to be in the software business. The Glassdoor reviews from their drafter team are unflattering — long hours, marketed as AI but staffed as case-manager work — but that's a labor-conditions critique, not a quality critique. Plenty of plaintiff firms genuinely love the product. The Series E at a $2 billion valuation is real.

What's true is that the $0.60 token comparison doesn't capture the actual decision. The actual decision is: do you want to build the operation EvenUp built, or do you want to rent it? At seventeen demands a month with no in-house engineering, you rent. At a hundred and seventy demands a month with a legal-ops function, the build case is real — and even then, the wedge isn't the token math. It's the absence of vendor lock-in compounding against your settlement history.

The competitor I'd actually watch in this space is Tavrn, not claude-for-legal. Tavrn ships a flat-rate monthly subscription explicitly priced against EvenUp's per-demand meter. They've raised $15 million. They aren't trying to disrupt with $0.60 token math. They're trying to disrupt with $1,500-a-month-unlimited-demands math. That's the cleaner attack vector.

The commercial PI-AI ecosystem (where Anthropic stayed out)

Anthropic's absence in PI isn't because nobody else is building. The PI-AI gold rush is loud and well-funded.

On the practice-management side, every dominant PI platform now ships its own AI. Filevine launched SideBarAI plus a DemandsAI module and a Lead Docket intake product. CloudLex has Lexee. Litify has the Agentic Case Expert. CASEpeer ships AffiniPay IQ. SmartAdvocate took the open-ecosystem route and integrates with EvenUp, Eve, Supio, and LawPro.ai natively. Most of these vendors are competing AI, not partnering with Anthropic's MCP ecosystem. None publishes a connector to claude-for-legal. The PMS vendors built their AI inside the system of record, and they have no incentive to open the door to a Claude Code skill pack sitting between their lawyers and their dashboard.

This is what Anthropic's absence in PI looks like in practice. The category is well-served by commercial vendors competing for the lawyer, not partnering with Anthropic for the lawyer.

The ethics layer most PI-AI posts skip

Here are the five points a personal-injury firm running claude-for-legal in 2026 needs to internalize.

1. Verify everything. ABA Formal Opinion 512 (July 29, 2024) named client intake among the four use cases it covers and was unambiguous: lawyers must understand the tool's data handling and verify every AI output. Morgan & Morgan attorneys have already been sanctioned for AI-hallucinated case citations. Every claude-for-legal demand letter that quotes a statute, cites a case, or makes a damages projection needs human review of those specific items before it leaves the firm.

2. HIPAA before everything. 45 CFR § 160.103 makes a PI firm receiving client PHI a Business Associate. The AI vendor processing that PHI on the firm's behalf is a downstream Business Associate and must sign a BAA. Anthropic offers BAAs on paid tiers. Without one, ingesting a medical record into any Claude model is a federal violation regardless of what your state bar says.

3. Chatbots have to identify themselves. Florida Bar Ethics Opinion 24-1, January 19, 2024 — still the cleanest U.S. opinion on AI intake — requires that any chatbot collecting facts from a prospective PI client disclose its AI status, decline to give legal advice, and refer legal questions to a licensed attorney. The bar treats the lawyer as ultimately responsible for whatever the bot says.

4. Contingency fees and AI productivity are an open question. The Virginia State Bar's LEO 1901, effective November 24, 2025, took the lawyer-friendly position: a contingency fee isn't per se unreasonable just because AI did some of the work. ABA Opinion 512 leans the other direction. The defensible posture is to document the risk basis of the fee, not the time basis, and avoid stacking AI-tool "costs" as separate line items.

5. Client consent to AI use has to be specific. NYC Bar Formal Opinion 2025-6 and the Pennsylvania Joint Formal Opinion 2024-200 both say boilerplate engagement-letter language doesn't suffice when AI use is material to the representation. PI firms taking referrals from other firms inherit the consent question — if the referring firm signed the client up under "no AI," the receiving firm can't unilaterally flip that switch.

What this means for PI firms in 2026

If you're a solo or small firm (three to ten attorneys)

Skip the claude-for-legal hype. The realistic build cost wipes out any token-economics savings, and you'll spend the saved budget on rebuilding what EvenUp or Supio or Eve already operates. Pick a vendor whose contract you can actually read, whose data-retention policy is in plain English, and whose pricing you can predict month to month. If you want to experiment with Claude itself, do it on the matter-management side, not the demand-letter side. The /litigation-legal:chronology skill is genuinely useful for organizing a case file once you have it. The demand-draft skill is useful as a first-draft engine if you're willing to bring all the structured intake yourself.

If you're a fifteen-to-fifty-attorney firm with engineering capacity

The build case is real, but plan honestly. A quarter of engineering to wire claude-for-legal into Filevine or Litify, another quarter to build the medical-chronology layer, plus an ongoing quarter-time engineer to maintain it. Year-one all-in is $200,000–$600,000 depending on scale. Year-two onward is half that. The break-even against EvenUp's per-demand model is somewhere around a hundred demands a month, and the long-term win isn't price — it's that your settlement history and case data stay yours, not the vendor's.

If you're a mass-tort or MDL firm running 1,000+ cases per matter

The absence of an OSS PI-AI stack is your problem to solve, and Anthropic just handed you the framework to do it. Build the bellwether-selection skill. Build the per-case auditable output that the steering committee will need to share with non-PSC counsel. Build the aggregate-settlement allocation explainer that ABA Opinion 06-438 requires you to disclose to each client. None of that exists yet. The wedge is the wedge.

I pinged three PI lawyers in ClaireAI's customer base about claude-for-legal this week. Two had not heard of it by name. One had read about it on LinkedIn and assumed there was a personal-injury plugin in there somewhere; when I told him there wasn't, he laughed and said "of course there isn't." He's a six-attorney plaintiff shop in Texas. He uses EvenUp. He has no plans to switch.

Where ClaireAI sits in this picture

Editorial disclosure..This guide is published by ClaireAI. We are not open-source and we are not a Claude for Legal plugin. We're a commercial AI intake receptionist with practice-area calibration for personal injury. Every named vendor, repo, statistic, and ethics opinion above is verifiable through the URLs cited in the body.

Anthropic's /legal-clinic:client-intake skill is a post-call structuring tool — it helps a paralegal organize an intake conversation that has already happened. Anthropic's /litigation-legal:matter-intake writes a matter.md for a case the firm has already accepted. Neither skill answers the phone, qualifies the caller, screens for conflicts of interest before privileged facts are collected, or dispatches a DocuSign retainer when the matter is grade-A. Those are intake-front-door functions that, in the PI context, sit before claude-for-legal in the workflow, not after it.

How ClaireAI fits into a PI firm's stack

• Per-practice calibration. ClaireAI for personal injury handles UIM/UM coverage questions, MIST-versus-significant-injury triage, and policy-limits inquiries the way a paralegal at a plaintiff shop would. We don't ship a generic chatbot.
• Rule 1.18 conflict screening before any privileged facts are collected. Fuzzy match against your CRM, fail-safe to escalate, no privilege leakage.
• 66 native case-management integrations including Filevine, Litify, CASEpeer, CloudLex, SmartAdvocate, Smokeball, MyCase, and Clio — the integrations claude-for-legal explicitly does not ship.
• SOC 2 Type II infrastructure and HIPAA-aligned BAAs signed at onboarding. No PHI ingestion without the paper in place.
• Live human escalation when the AI hits the limit of what it should be answering on its own. For PI specifically, that includes any caller who mentions an ongoing DV situation, a hospitalized loved one, or anything else where a script-driven bot would be the wrong answer.

If you'd like to see ClaireAI handle a calibrated PI intake call for your practice, the demo is the most efficient way to evaluate the fit. Pricing is published on the pricing page — no sales-gated quotes. If you're looking for the broader picture on open-source legal infrastructure beyond PI, the open-source legal intake guide covers the wider landscape.

Frequently asked questions