Open-source legal intake doesn't exist yet — and that's about to change

The state of the question. There is no production-grade open-source legal intake software competitive with Lawmatics, Lead Docket, or Intaker in May 2026. There are open-source document assistants (Mike), document-automation engines (Docassemble), e-signature platforms (DocuSeal), case-management workspaces (Stella), and generic voice receptionists (AIReceptionist, Pipecat) — and a fast-growing community putting them together by hand. The one piece of the legal-AI stack still missing from the open-source movement is the front door.

The Mike OSS moment

Will Chen, a solo developer, shipped Mike — an open-source clone of Harvey and Legora — in 14 days and posted it on Hacker News on April 29, 2026. The Mike repository collected 3,550 stars, 1,089 forks, and 49 subscribers within a month. The license is AGPL-3.0, which forces any commercial fork to publish its own changes. The stack is Next.js + Express + Supabase + Cloudflare R2, with Claude, Gemini, and GPT wired up as interchangeable model providers.

What Mike actually does: lets a user sign up, upload matter documents, organize them into project workspaces, and chat with them across multiple frontier models. It's a legal-document research assistant. Downstream forks already exist — MikeRust ports it to a native Tauri desktop app; Emilie rebrands it as a Swiss sovereign-data variant; another fork wires in cryptographic verification.

Reading the repo cover-to-cover this past weekend, the code is rougher than Harvey's marketing comparisons would suggest. There are TODOs in the auth flow, the document-chunking strategy is naïve, the workspace model has corners that haven't been fully thought through. That is part of the point. The codebase is small enough that a firm's IT consultant could read it on a Saturday and understand exactly what is happening to client data — which is the inverse of every closed legal-AI product I have evaluated this year.

The traction isn't the surprise. The narrative is. Harvey raised a $200M round at an $11B valuation in March 2026; Mike, built in two weeks by one person, demonstrates that the technical moat around legal AI is thinner than the valuation implies.

What Mike doesn't do — the missing layer of the legal stack

Mike is a legal-document research assistant. It is not a legal intake system. The distinction matters because the open-source legal-AI conversation routinely conflates them.

A legal intake system is the front door: it picks up a phone call or web chat from a prospective client, screens for conflicts of interest, qualifies the matter, books a consult, and hands off to a case-management system. That workflow involves telephony, speech-to-text, calendaring, e-signatures, CRM write-back, and — most importantly — a model that performs reliably under Rule 1.18 confidentiality constraints while a stranger describes a legal problem at 11 p.m. on a Sunday.

Map the open-source legal stack today and the gap is unmistakable:

Three forces pried the door open

The Mike moment didn't happen in a vacuum. Three concurrent events made the legal-AI market structurally vulnerable to open-source disruption in 2025–2026.

1. Casetext's shutdown

Thomson Reuters acquired Casetext for $650 million in 2023, kept the brand alive for two years, then shut down standalone CoCounsel on April 1, 2025. Users with promotional pricing locked in "for life" were forced onto Westlaw-bundled tiers running 5–10× the original $65/month rate. The case study is now a permanent feature of legal-tech buyer conversations: this is what acquisition risk looks like.

2. Harvey's pricing wall

Harvey ran from a $3B valuation in February 2025 to $8B in December 2025 to $11B in March 2026. Reported seat pricing is approximately $1,200 per lawyer per month with 20-seat minimums and 12-month commits. The June 2025 LexisNexis alliance is projected by Artificial Lawyer to push all-in costs toward $3,000 per seat once Lexis content overlaps with Harvey workflows. For any firm under 50 attorneys, the economics simply do not work.

3. The ethics floor moved

ABA Formal Opinion 512 (July 29, 2024) was the first comprehensive ABA guidance on generative AI. It cross-references Model Rules 1.1 (competence), 1.4 (communication), 1.5 (fees), 1.6 (confidentiality), 3.1 and 3.3 (candor), and 5.1 and 5.3 (supervision). The load-bearing passage for intake:

“Before lawyers input information relating to the representation of a client into a GAI tool, they must evaluate the risks that the information will be disclosed to or accessed by others.”

A boilerplate engagement letter does not satisfy that obligation. The natural architectural answer is a model the firm controls — and self-hostable Llama 3.3, Mistral, and Qwen finally crossed the usability threshold for that in 2025.

What actually exists in open-source legal intake today

An honest inventory of the field, broken into three buckets:

Lawyer-side intake assistants (not customer-facing)

• LawDroid Legal Aid Plugin — Tom Martin's Apache-2.0 Claude plugin pack, released v0.1.0 on May 20, 2026. Includes a /client-intake skill (practice-area templates, cross-area issue spotting, conflict flags, urgency triage) and /eligibility-screening (income, residency, citizenship per funder rules). Brand-new and explicitly scoped to civil legal aid — not retained private practice.
• Anthropic claude-for-legal — 7,800-star Apache-2.0 repo launched April 2026 with 12 plugins and 80+ skills. The intake-relevant ones (/legal-clinic:client-intake, /litigation-legal:matter-intake, /litigation-legal:demand-intake) run inside Claude Code; they are markdown prompt packs, not deployable apps. A licensed attorney uses them to structure intake notes after the prospect has already called.

Generic voice receptionists (no legal calibration)

• AIReceptionist (kirklandsig/AIReceptionist) — 40 stars, AGPL-3.0. OpenAI Realtime + LiveKit SIP voice agent. The only built-in template is dental. The README warns of breaking changes.
• Pipecat — BSD-2 voice framework with a medical patient-intake example. A community PR to add a legal example (#631, October 2024) was explicitly rejected as redundant in January 2025. No legal example currently in the tree.

Court-facing and legal-aid (deterministic, no LLM in the intake path)

• Docassemble — 951 stars, MIT. Jonathan Pyle's guided-interview engine. Powers most U.S. legal-aid intake in production. No LLM front-end.
• Suffolk LIT Lab docassemble-AssemblyLine — 62 stars, MIT, v4.6.0 in May 2026. A framework on top of Docassemble that turns paper court forms into guided web interviews. Drives CourtFormsOnline.org and replicators in IL, FL, and CA. Purely deterministic; the LLM work (Weaver, Steenhuis et al.) helps authors write interviews, not run them.

The verdict..No project on the list above is a production-grade, customer-facing AI legal intake system. The closest is LawDroid's Legal Aid Plugin, which is one week old, alpha-grade, scoped to civil legal aid funders, and runs inside Claude as a markdown skill pack.

The ethics question, answered cleanly

The ethics layer is the question every law-firm partner asks first, and the answer is more permissive than the legal-tech press makes it sound.

Model Rule 1.18 attaches the moment a chatbot collects facts from a prospective client. Confidentiality, conflicts imputation, and the duty not to use the information adversely begin pre-retainer. An intake bot must therefore (a) run a conflict pre-screen before collecting substantive facts and (b) be architected so the model does not train on inputs.

Florida Bar Ethics Opinion 24-1 (January 19, 2024) is the cleanest opinion on chatbot intake specifically. Its four requirements: (1) the lawyer must inform prospective clients that they are communicating with an AI program; (2) the chatbot must clearly identify its non-lawyer status; (3) the bot must limit itself to factual intake and refer legal questions to a lawyer; (4) the lawyer is "ultimately responsible should the chatbot provide misleading information."

ABA Formal Opinion 512 (July 29, 2024) names "client intake" explicitly as one of the four use cases it covers. Boilerplate engagement-letter consent does not satisfy Rule 1.6 for a self-learning AI tool — the lawyer must understand how the GAI tool uses data specifically.

Oregon State Bar Opinion 2026-208 treats the AI agent as a non-lawyer assistant under Oregon RPC 5.3 (parallel to ABA Model Rule 5.3). Lawyers must monitor the AI to prevent "creating false impressions of attorney-client relationships, promising services the firm cannot deliver, guaranteeing particular outcomes." Disclaimers help but do not fully discharge the supervision duty.

Implication for open-source vs. SaaS..No U.S. bar opinion treats open-source differently from commercial software. The supervision duty under Rule 5.3 does not disappear when a firm self-hosts — it relocates. If you self-host the model, you become the vendor, and the architectural burden (no training on inputs, encryption, BAA, retention controls) is yours. If you use a commercial vendor, the contractual burden is yours.

The honest economics

The legal-tech press over-promises self-hosting. The actual breakdown:

The break-even. Self-hosting beats SaaS on raw monthly cost at roughly 10–15 attorneys, assuming the firm has — or can hire — a quarter-time engineer to keep the stack alive. Below that scale, the IT contractor line item alone (4–8 hours per month at $100 per hour) erases the apparent savings.

The trap. Running Llama 3.3 70B on your own GPU costs $3,000–$5,000 per month at production volume (cloud A100 plus ops). At realistic intake token volumes — under 10 million tokens per month for most firms — API providers like Claude Haiku and GPT-4o-mini are 10–100× cheaper. Self-hosting the model rarely pays unless you are doing more than 50M tokens per month or have a hard data-residency requirement.

Where SaaS still wins

• Compliance inheritance. Lawmatics, Clio Grow, and Smith.ai ship SOC 2 and HIPAA BAAs out of the box. DocuSeal Community self-hosted does not hand you a BAA — that's on you.
• Case-management integrations. Lawmatics, Lead Docket, and Intaker have prebuilt sync with Clio, Filevine, MyCase, and PracticePanther. Rebuilding those integrations in-house is weeks of engineering.
• Human escalation. Smith.ai's $9.75–$11-per-call human receptionist is irreplaceable for the confused 2 a.m. caller. Self-hosted gives you no human fallback.

The reframe. The honest question is not self-host vs. SaaS. It is one monolithic SaaS vs. several single-purpose APIs glued together. Every public "I built my own intake stack" case study turns out, on inspection, to be the second option: Twilio + Deepgram + ElevenLabs + Claude + DocuSeal + a CRM, orchestrated by n8n or custom code. The data lives in the firm's tenant of each service; the firm is not actually self-hosting anything.

What this means for law firms in 2026

If you're a solo or small firm (under 10 attorneys)

Skip the open-source hype. The realistic choice is a vendor whose contract you can actually read — short data-retention policy, no model-training rights over your conversations, clean export path, month-to-month or short-commit pricing. That is the meaningful version of "data sovereignty" at your scale, and it does not require self-hosting.

If you're a 15+ attorney firm with operational capacity

The building blocks exist. Pipecat or LiveKit (voice), Deepgram (STT), ElevenLabs (TTS), DocuSeal (e-signature), Twenty or EspoCRM (CRM), and a self-hosted Llama or Mistral can be assembled into a credible intake stack. But you are accepting responsibility for the integration, the supervision under Rule 5.3, and the bar opinions in every jurisdiction where you practice. Plan on at least one quarter of engineering work and a permanent quarter-time operational owner.

If you're an open-source developer

This is the wedge. The community has document research, document automation, e-signature, case management, and voice infrastructure as separate open-source projects. The customer-facing intake layer is the missing piece — and the first project that ships a credible, legal-calibrated, multi-channel intake under a permissive license will define the next chapter of the conversation Mike OSS opened.

Where ClaireAI sits in this picture

Editorial disclosure..ClaireAI is the publisher of this guide. We are not open-source, and we say that clearly below. Every named vendor, repo, and statistic above is verifiable through the GitHub, ABA, state bar, and pricing-page URLs that source them.

ClaireAI 365 is not open-source. It's a commercial product, purpose-built for law firms that need the legal calibration, CRM integrations, compliance attestations, and human escalation paths that no open-source project ships today.

That's the honest framing. The reason this guide exists is that the legal-tech press treats "AI legal intake" as if it were a solved category, and it isn't. The open-source side is fertile and growing — Mike OSS is real, Docassemble is foundational, LawDroid's plugin pack is the most interesting new entrant in twelve months — but none of it is a turnkey customer-facing AI receptionist for a firm that needs to be live next week.

Where we differ from the alternatives covered above

• Calibrated per-practice — personal injury, criminal defense, family law, immigration, and general civil each get a tuned intake script.
• Conflict screening per Rule 1.18 before any privileged facts are collected.
• 66 native case-management integrations including Clio, Filevine, MyCase, PracticePanther, CASEpeer, Litify, Lawmatics, CloudLex, Smokeball, Rocket Matter, and CosmoLex.
• SOC 2 Type II infrastructure and HIPAA-aligned BAAs signed at onboarding.
• Live human escalation when the AI reaches the limit of what it should be answering on its own.

If you'd like to see ClaireAI handle a calibrated intake call for your practice area, the demo is the most efficient way to evaluate the fit. Pricing is published on the pricing page — no sales-gated quotes.

Frequently asked questions