Anthropic Claude Skills vs Vendor Plugins Firm Decision

Claude skills vs vendor plugins is the procurement decision facing every firm in 2026. On one side: Anthropic's open-source Cowork legal plugin — `/review-contract` and `/triage-nda` shipped February 2026, free under permissive license, configurable via YAML, running inside any paid Claude tier. On the other side: Harvey AI (quote-only, industry-estimate $1,200-$2,000+/seat/month per Artificial Lawyer reporting, not vendor-confirmed), Spellbook (quote-only, industry-estimate $180-$300/seat/month with $199/seat/month enterprise minimum at 10 seats per Artificial Lawyer / aiapps coverage, not vendor-confirmed), Thomson Reuters CoCounsel (tier prices $75-$500/user/month per Costbench March 2026 secondary-source data, not vendor-confirmed), and LexisNexis Protégé (quote-only). The structural choice isn't "better tool" — it's two different operating models. This is the operator's read on which model fits which firm.

The two operating models — and what each actually trades

Model A: Anthropic open-source skills + Claude tier. The firm runs Claude (Pro $17-$20/user/month, Team $20-$25/seat/month, or Enterprise quote-only per Anthropic pricing) and uses the open-source Cowork plugin's pre-built skills or forks them for custom workflows. Configuration is YAML the firm owns. Customization is the firm's responsibility. Skill updates come from the upstream Anthropic repo when the firm pulls them.

Model B: Vendor-managed proprietary plugin. The firm pays a vendor (Harvey, Spellbook, CoCounsel, Protégé) for a pre-built legal AI workflow with vendor-controlled UI, vendor-maintained precedent libraries, vendor-driven feature updates, and vendor-provided customer success. Configuration happens through the vendor's interface. The firm's customizations don't port to alternative tools.

The tradeoffs:

| Dimension | Model A (Anthropic) | Model B (Vendor) | |---|---|---| | Per-seat cost (industry estimates) | $17-$25/seat/month | $75-$2,000+/seat/month | | Setup time | 4-15 hours partner config | 4-12 weeks vendor implementation | | Customization speed | YAML edit, deploy in minutes | Vendor change request, days to weeks | | Workflow depth at launch | What the open-source repo ships | Vendor's full proprietary platform | | Customer success | Self-serve + community | Dedicated CSM (vendor) | | Portability if pricing rises | Re-point YAML to alt model | High switching cost | | Audit trail | Git commits | Vendor-managed | | Update cadence | Pull when convenient | Push by vendor |

The pickable side on FIT depends on three variables: firm size, available senior-counsel time for configuration, and tolerance for self-serve customer success. (deeper read on the open-source legal skills firm customization)

When Model A (Anthropic open-source) wins

Five firm types where the open-source Cowork stack wins on FIT:

1. Solo and small firms (1-15 attorneys). Vendor pricing typically prices these firms out anyway. Spellbook's industry-estimate $199/seat/month enterprise minimum at 10 seats means a 5-attorney firm can't typically buy seats. Harvey's quote-only enterprise pricing aims at AmLaw 100. Claude Team at $20/seat/month annual is the only AI workflow these firms can realistically afford with workflow depth. The 4-6 hours of senior-attorney time on YAML configuration is a one-time investment that produces ongoing leverage.

2. In-house legal teams under 30 lawyers. The procurement velocity of installing an open-source plugin (a working week) beats the 4-12 weeks of vendor RFP, MSA negotiation, security review, and onboarding. For in-house teams that need NDA triage and contract review running by month's end, the open-source path closes fast.

3. Firms with strong senior-counsel time availability. The playbook configuration is the senior-counsel investment. Firms where a senior partner can carve out 4-15 hours of configuration time across a quarter benefit from the open-source path. Firms where every senior partner is fully booked typically need the vendor's pre-built workflows even at higher cost.

4. Practice areas with idiosyncratic clause libraries. Plaintiff's contingency-fee firms (retainer agreements), public defender's offices (intake triage), niche transactional boutiques (specific deal-type structures), or firms with major clients running standardized vendor terms benefit from custom YAML playbooks more than they benefit from vendor-defined defaults.

5. Firms valuing portability over pre-built depth. The YAML playbook ports across model providers. If Anthropic's pricing changes in 2027 or if a competitor model becomes preferable, the firm migrates by re-pointing the YAML. Vendor-managed customizations don't port — switching from Harvey to Spellbook means rebuilding the customization layer.

The pickable side: for these five firm types, Model A is the cheapest competent solution in 2026. (read the procurement checklist for mid-market firms)

When Model B (vendor plugin) wins

Five firm types where the vendor-managed proprietary plugin wins on FIT:

1. AmLaw 100 deal teams running multi-billion-dollar matters. Harvey's pricing assumes the firm has BigLaw-grade workflow depth requirements: deal-room integration, DMS integration, contract repository sync, multi-matter parallel processing, dedicated 24/7 customer success. Open-source Cowork doesn't ship these out of the box. The firm would need to build them, which costs more in engineering time than Harvey's industry-estimate $1.5-2K/seat/month.

2. Firms without senior-counsel configuration time. When every senior partner is booked, the 4-15 hours of YAML configuration becomes a real bottleneck. Vendor-managed plugins ship pre-configured workflows out of the box. The firm pays for the configuration to already exist.

3. Litigation teams using CoCounsel for Westlaw integration. Thomson Reuters' rebuilt CoCounsel embeds Westlaw + Practical Law + Anthropic models. For litigation teams already on Westlaw, the integrated research-plus-AI workflow is a workflow gain Cowork can't match — Cowork doesn't have a research database baked in. (read the Anthropic × Thomson Reuters CoCounsel rebuild analysis)

4. Firms where procurement requires vendor MSA + insurance + SOC 2. Some firm procurement processes require traditional vendor relationships — MSA, professional liability insurance, SOC 2 Type II certification, indemnification provisions. Open-source plugins don't carry vendor liability. For firms whose procurement process won't accept open-source deployment, the vendor path is the only viable path.

5. Firms wanting pre-built precedent libraries. Spellbook ships the Spellbook Library precedent learning feature. Harvey ships pre-built workflows for common BigLaw matter types. Open-source Cowork ships skill primitives, not precedent depth. Firms that want a pre-built precedent corpus from day one need a vendor.

The pickable side: for these five firm types, paying the vendor premium is operationally rational. The cost differential against open-source isn't waste — it's purchasing pre-built workflows, customer success, and procurement-process compatibility.

The hybrid pattern — most BigLaw will run both

The cleanest answer for AmLaw 100 firms in 2026 is hybrid deployment. Run vendor stacks (Harvey, CoCounsel) for BigLaw-grade matter management, deal-room integration, and pre-built workflow depth. Run open-source Cowork plugins for in-house legal ops, mid-tier matter triage, and custom workflows the vendors don't ship.

The public reference is Freshfields. Per the April 23, 2026 announcement, Freshfields runs:

- Direct Anthropic deployment — Claude across 5,700 employees in 33 offices via the firm's proprietary AI platform, with co-development for legal-focused workflows and plans to expand to Cowork. - Thomson Reuters CoCounsel — early adopter of the rebuilt CoCounsel Legal (Anthropic-powered, Westlaw + Practical Law embedded).

The Freshfields pattern signals what most BigLaw firms will run by year-end 2026: Anthropic-direct for custom workflows and firm-wide AI access, vendor stacks for pre-built integrations and matter-specific depth. The two layers don't compete — they complement.

For mid-market firms (10-50 attorneys), the hybrid pattern is rare because the vendor pricing prices them out of Harvey or full CoCounsel All Access ($500/user/month per Costbench secondary-source data, not vendor-confirmed). Most mid-market firms run pure Model A through 2026 and revisit if a vendor lands at mid-market pricing.

For in-house legal teams, the hybrid pattern is also rare — most teams pick one model and stick. The exception is in-house teams at companies with active vendor MSAs already in place (e.g., the company's outside counsel runs Harvey and the in-house team gets Harvey access through the firm's seat allocation).

The second-order effect: vendors selling against Anthropic-direct deployment will increasingly position themselves as the "BigLaw-grade depth" layer rather than the foundational AI vendor. CoCounsel's rebuild around Anthropic models is the explicit version of this positioning. (read the Anthropic disintermediation vs CoCounsel augmentation analysis)

Decision framework: 4 questions to pick your side

Translation of the FIT analysis into a decision framework. Answer four questions:

Q1: How many attorneys will use the tool? - 1-15: Model A (Anthropic open-source). Vendor pricing prices you out anyway. - 15-50: Model A unless specific BigLaw integration requirement. The cost savings (9-100x at industry-estimate pricing) typically justify configuration time. - 50+: Hybrid. Run Anthropic-direct for breadth + vendor stacks for depth where needed.

Q2: Do you have 4-15 hours of senior-counsel time available for playbook configuration? - Yes: Model A is viable. The configuration is a one-time investment. - No: Model B (vendor plugin) — the pre-built workflows are what you're paying for.

Q3: Does your procurement process require vendor MSA + insurance + SOC 2? - Yes: Model B. Open-source deployment doesn't fit traditional procurement. - No: Model A is viable. The cost savings flow through.

Q4: Do you need pre-built precedent libraries or just skill primitives? - Pre-built precedents required: Model B (Spellbook for transactional, Harvey for BigLaw-grade matter mgmt, CoCounsel for Westlaw-integrated research). - Skill primitives are enough: Model A. Configure YAML against your firm's actual deal flow over 6-12 months.

Apply the framework: a 25-attorney mid-market firm with a senior partner who can carve out 6 hours for YAML configuration, no specific vendor MSA requirement, and willingness to build precedent depth over time = clear Model A. The firm saves $48,000-$84,000/year against Spellbook industry estimates (not vendor-confirmed) and $354,000-$594,000/year against Harvey industry estimates (not vendor-confirmed) at 25-attorney scale.

A 250-attorney AmLaw firm with no senior-counsel configuration time, mandatory vendor MSA requirement, and BigLaw-grade matter management depth required = clear hybrid. Run Claude Enterprise via Microsoft Foundry for breadth, layer Harvey or CoCounsel for matter-specific depth.

The pickable side: for the median legal-AI buyer in 2026 — a mid-market firm of 15-75 attorneys — Model A wins on FIT. The cost savings, configuration flexibility, and portability outweigh the pre-built workflow depth a vendor would ship. (read the in-house counsel deployment checklist)

The Bottom Line: My take: for solos through mid-market firms (1-50 attorneys), Anthropic's open-source Cowork skills plus Claude Team are the better procurement decision in 2026 — 9-100x cheaper than vendor plugins at industry-estimate pricing, with portable YAML playbooks the firm owns. For BigLaw and AmLaw 100, the hybrid pattern wins: Anthropic-direct for breadth and custom workflows, vendor stacks (Harvey, CoCounsel, Spellbook) for BigLaw-grade depth where pre-built workflows or research-database integration justify the premium. Freshfields is the public template. Most firms haven't run the math; the ones that do are migrating their mid-tier workflows to open-source skills and keeping vendor stacks for the workflows that genuinely need them.