Radar / Ideas / CounselDesk: the on-device AI…

CounselDesk: the on-device AI workstation for small law and accounting firms

daily ideamoderateJEV confidence 0.59JEV monetization 0.62026-10-28

OutcomeClient-confidential AI assistance with zero client files leaving the firm's computers, backed by an audit log that proves it to a regulator or client.

The problem

Seven in ten lawyers at law firms work solo or in firms of 10 or fewer, yet face the same confidentiality duties as large firms: ABA Model Rule 1.6 and Formal Opinion 512 require weighing disclosure risk before entering client information into generative AI, and California's new AI law bars inputting confidential information into systems without restricted access. Firms respond by banning client data in public AI tools, while the enterprise AI plans and private deployments large firms rely on are largely out of small firms' reach — so small firms get the risk without the tooling.

The idea

A desktop workstation for regulated professionals that runs the AI on the firm's own computers: a local agent reads and summarizes matter files, drafts, and correspondence; session policy rules (in the style of Cisco's Quarry) flag risky actions such as a credential read followed by an external send; credentials are isolated per agent (1Password-for-Agents style) and every agent gets a verifiable identity (Alter style). Every action is logged locally into an exportable audit report a firm can show a client or bar auditor. Cloud use is possible only through explicit, per-matter opt-in with masking, never by default.

Why now

Three enablers converged in October 2026: Hob proved a local desktop agent can do real file work with open-weight models; Quarry gave the session-rule pattern for policing agent behavior; and Conseal's 23 Oct launch showed a vendor already building a business on small-firm on-device privacy. Regulators are now writing AI confidentiality duties into law, turning a nice-to-have into a compliance requirement.

What it combines

Hob's local desktop agent plus Quarry's session-level behavior rules plus 1Password for Agents' credential isolation plus Alter's zero-trust agent identity: local execution, policy enforcement on action sequences, secrets the agent never sees, and attributable identity. The mix matters because any one piece alone fails the privilege test — local without rules can still exfiltrate, rules without identity cannot attribute, and identity without isolation leaks secrets.

MVP

Weekend scope: a desktop app wrapping a local model for document summarization and a redaction checklist, with a local action log exportable as a report. Deliberately skip cloud mode, practice-management integrations, multi-user support, and legal-specific model fine-tuning.

How it makes money

Budget owner: the managing partner / firm operations lead, who already pays per-seat practice-management software. Pricing anchor: Clio Manage EasyStart at $39 per user per month, with Essentials at $79 and Complete at $139 (selecthub.com comparison). Willingness-to-pay evidence: firms already pay those seats for practice management, large firms already buy enterprise-grade AI platforms with no-training and retention controls (per MacDonald Illig's client guidance), and Conseal's launch exists precisely because small firms will pay for on-device privacy they cannot get from enterprise plans.

The moat

Cloud incumbents cannot ship local-first without cannibalizing their own model: Harvey, Clio AI, and general AI vendors monetize cloud subscriptions and retained-data ecosystems, and a local-only product removes both the subscription hook and the data flywheel. A second structural moat is privilege itself: a vendor that has ever held client data cannot retroactively claim it never did, while a local-first entrant can make that claim from day one.

Distribution

Direct to solo and small firms, plus a wedge through the legal IT consultants and practice-management resellers who already serve them. Land with one workstation per firm on document summarization and redaction checks, then expand seat by seat as the audit log becomes part of the firm's compliance story.

Why it wins

Conseal masks data before it reaches cloud tools, so the workflow still depends on the cloud; Harvey and Clio AI are cloud products whose privacy rests on vendor promises. CounselDesk is local-first with an audit log, so confidentiality holds by architecture rather than by policy, and the firm can prove it.

Risks

Biggest risk: local models are weaker than frontier models for hard legal reasoning, and lawyers will notice. The MVP de-risks this by scoping to summarization and redaction checks where local models suffice, with lawyer review as the control, and by measuring task accuracy on the firm's own documents before expanding scope.

Build it with

Hob: free, open-source desktop AI assistant that runs open-weight models on your own computerThe local desktop agent pattern: file workflows with open-weight models and nothing leaving the machine.
Cisco Project Quarry: open-source behavior detection engine for AI agentsSession-level behavior rules that flag risky action sequences like credential reads followed by external sends.
1Password for Agents: personal AI agents sign in without ever seeing the passwordCredential isolation so agents sign in without ever seeing the password.
Alter (YC S25): zero-trust identity and access control for AI agentsZero-trust identity and access control so every agent action is attributable to a verified agent.
Jev: TypeSafe AI's System One models for structured decisionsCalibrated decision layer for policy judgments such as whether an action may proceed or needs human approval.

Repo to start from

counseldesk — on-device AI workstation for small firms: local document work, session policy rules, credential isolation, and an exportable audit log.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.