Labs sell their own apps to their API customers, and the app layer's supplier becomes its rival
Frontier labs cross-sell first-party apps into platform accounts. App-layer startups and seat-priced SaaS are exposed.
Competitive · S-1 shows half cross-sell
Written for: Investor or analyst
Next step
Thesis: labs will capture generic apps inside their own accounts, so app-layer value sits only where workflow and proprietary data beat the model. Confirming signal: a funded app-layer startup cites lab first-party apps as the reason for a down round or pivot, or Anthropic's next filing shows the cross-sell share rising. Already building: Anthropic and OpenAI first-party apps, Hebbia on the multi-model side, and Hob as the open-source alternative.
Why now
This is the first hard number showing labs winning the app layer inside their own API accounts. Stratechery's view that competition eats the profits pushes labs further up the stack to protect margin. The survivors look like Hebbia, which reports double-digit gains from model choice in source selection and deck generation. Its moat is workflow and data, not the model. Open-source projects like Hob copy Cowork on-device, which squeezes thin wrappers from below too.
What became possible
The S-1 shows that more than half of Anthropic's $100,000-plus customers use both the Claude Platform and a first-party app. TechCrunch compares the cross-selling playbook to Amazon's. OpenAI has turned its coding tools into dots, an always-on personal assistant. Sullivan & Cromwell builds its own legal tools on Codex instead of buying a vertical app.
Two ways in
- Model-neutral vertical workflow
Back app-layer firms whose product is the regulated workflow (data rooms, legal precedent playbooks) and that switch models per task, which a single-lab app can't credibly offer.
- Customer
- Banks, law firms and investors working through large document sets
- Model
- Enterprise seats plus usage
- Private desktop agent for regulated firms
A supported, enterprise-managed version of open-source on-device agents like Hob, running open-weight models so files never leave the machine.
- Customer
- Compliance-heavy firms barred from sending documents to lab clouds
- Model
- Per-device licence plus support
Competition
The model-neutral vertical workflow idea is mostly taken in legal. Harvey already sells per-task routing across Anthropic, OpenAI and Google models with an admin model selector, and Emma Legal, Drooms and Fast.io cover AI diligence over data rooms.
Full competition notes (15 competitors)
The model-neutral vertical workflow idea is mostly taken in legal. Harvey already sells per-task routing across Anthropic, OpenAI and Google models with an admin model selector, and Emma Legal, Drooms and Fast.io cover AI diligence over data rooms. Model neutrality is now a standard feature, not an edge. A new entrant could only win by owning a narrower regulated workflow that Harvey treats as secondary, such as bank-side data-room review or credit and fund-document workflows, and pairing it with a deployment option that is not any lab's cloud. The private desktop agent idea has a clearer opening. The market is split between single-user open-source tools like Jan, AnythingLLM and LM Studio, which have no enterprise management or certifications; small appliance or consultancy sellers like LocalApps, Soverance, Mako Logics and Spectre Computers; and server-based on-premises platforms like ibl.ai and Automation Anywhere. Nobody clearly offers a supported, IT-managed, audited desktop agent running open-weight models on the device for large banks and law firms. Hob is not a good base: it appears to be closed-source paid software that runs lab-hosted coding agents. A winner would need fleet management, audit logs, protection against malicious instructions hidden in documents, certifications, and local models that keep improving. The main risk is that labs, Microsoft or PC makers add on-device or private-deployment modes.
- Harvey (multi-model)
Legal AI leader that dropped OpenAI exclusivity in 2025. It now routes each sub-task across Anthropic, OpenAI and Google models, with Auto mode and an admin-enabled model selector. This is the model-neutral legal workflow pitch, already executed at scale; its valuation was reported at about $11B in March 2026.
- Anthropic legal plugins for Claude / Claude Cowork
Proof of the lab-as-rival thesis. In May 2026 Anthropic launched 12 legal plugins (contract review, DocuSign and Thomson Reuters integrations) delivered through its first-party Cowork agent.
- Emma Legal
AI workspace for M&A legal due diligence that connects to data rooms and describes its AI as a multi-model system.
- Drooms Intelligence
Virtual data room incumbent that adds LLM search and extraction across whole M&A and real-estate data rooms. Its pages don't say which model providers it uses.
- Fast.io agent data rooms
Agent-oriented data room and storage layer that says it works with Claude, GPT-4o and any LLM. It is horizontal rather than built for legal or finance.
- Due Diligence Agents (open source)
Open-source multi-agent diligence system covering 9 domains such as legal, finance, tax and regulatory. Its quick start uses an Anthropic key.
- hob
Desktop workspace that bundles coding-agent CLIs like Claude Code, Codex and OpenCode, with local history storage. Listed as freemium SaaS at $10/month Pro, with no evidence it is open source, and it relies on lab subscriptions rather than open-weight models.
- Jan
Open-source (AGPLv3) offline desktop app for running local models. Auditable, but no enterprise management or certifications found.
- AnythingLLM / LM Studio
Popular local LLM desktop tools with RAG and agents. One roundup says LM Studio is single-user with no compliance features, and no SOC 2 or ISO certifications were found for either.
- LocalApps (LocalChat)
Dutch startup founded in 2025 selling a fully local AI chat app to solo and small law firms on subscription.
- Soverance
Sells on-premises local AI servers to law firms and healthcare providers as a one-time purchase.
- Mako Logics LocalAIBox
On-premises AI appliance for organizations handling regulated data, with browser access and no per-seat fees.
- ibl.ai air-gapped AI
Enterprise on-premises and air-gapped AI platform with local models, marketed to legal. Server deployment rather than a desktop agent.
- Automation Anywhere on-prem with Nemotron
Large automation vendor offering a fully on-premises agent automation platform on Nvidia's open-weight Nemotron model.
- Wanta
Apache-2.0 open-source framework for building desktop AI agents on OpenCode that can run against local OpenAI-compatible models.
Checked 2026-10-27
Risks
- Lab apps may stay shallow in regulated verticals, leaving room for specialists
- Enterprises may resist concentrating platform and app spend with one vendor
- On-device open models still trail frontier quality on long tasks
Strong tech readiness and team fit; open space is the weak point (5).
tech 8 · pull 7 · space 5 · fit 8 · build 6
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- Anthropic’s IPO filing reveals its strategy for keeping enterprise customers hookedTechCrunch AI
- Anthropic S-1, The Cost of Compute, Who Is Paying?Stratechery
- How OpenAI turned its coding tools into dots, an always-on personal assistantTechCrunch AI
- How Sullivan & Cromwell builds legal tools with Codex | OpenAIOpenAI
- Hebbia on Astra: Better research, better presentations | OpenAIOpenAI
- Show HN: Hob - Open-source, on device alternative to Claude Cowork and ChatGPTHacker News
Status history
- 2026-10-27 Flagged as buildable