"AI document management your auditor can defend." A cloud-native DMS for law firms and AEC teams where AI agents are the substrate — not a feature — with a hash-chained audit trail, deny-always-wins legal holds, an AEC vertical module, DLP + anomaly detection, and 77 typed MCP tools. Built on Next.js, Neon Postgres, and content-addressable blob storage.
Mid-market legal and AEC firms run on documents — contracts, matters, drawings, specs, holds. The incumbents (iManage, NetDocuments, and other established DMS vendors) own the folder metaphor and bolt AI on top. When litigation hits, exports take weeks and audit trails are fragile.
What's broken
What we built
Kodori treats AI agents as the substrate of the DMS, not an upsell. Every action is captured as an immutable, hash-chained event. Documents are stored content-addressably so dedup is automatic. Legal holds, retention classes, collections, AEC artifacts (RFIs, submittals, drawings), DLP findings, and anomaly review are all first-class concepts the agent reads and writes through 77 typed MCP tools.
Postgres full-text and pgvector semantic embeddings (1536-dim text-embedding-3-small) run in parallel and combine via Reciprocal Rank Fusion. Plain-language queries, exact phrases, or concept-based retrieval — with the path that matched each hit.
Claude Haiku 4.5 proposes metadata after extraction — sensitivity, collection, keywords, document type. Every suggestion has Accept / Dismiss buttons; accepting writes a durable event so the audit log credits the human, not the agent.
Bind documents to a matter or litigation hold and the system refuses to delete or downgrade sensitivity. The UI disables Delete, the retention queue disables disposal, and MCP tools enforce the same gate server-side. Three independent enforcement points.
Every consequential mutation appends a row to the events table. Each row's prev_hash is the SHA-256 of the previous event — tampering is detectable. Same chain backs SOC 2, 21 CFR Part 11, GDPR, and FRCP discovery exports.
Next.js 15.5 with shadcn/ui and Tailwind CSS 4. Turborepo + pnpm monorepo with 9 packages (core, db, events, agent, mcp, workflow, evals, migration, sdk) and 2 apps (web, sync-companion).
Next.js Route Handlers on Vercel for the API surface. Inngest for durable multi-step pipelines (extract → embed → auto-classify). Cloudflare Email Workers for HMAC-signed inbound email ingress.
Every agent capability is a typed MCP tool. The agent can't do anything the same tool wouldn't let the user do via the API. Read tools never mutate; write tools always emit events.
Neon Postgres with pgvector pre-installed. Drizzle ORM 0.36. 50 tables covering tenants, users, document objects + versions + content + chunks + redactions + drawings + citations, collections, permissions, hash-chained events, API keys, webhooks, invites, legal holds + custodians, retention, metadata suggestions, saved searches + alerts, AEC (projects, RFIs, submittals, inspections, change orders), AP invoice workflow, anomalies, DLP findings, annotations, productions, privilege-log overrides, share links, migration jobs, external connectors + messages, pending deletions, tenant policies, and per-tenant KMS keys.
Plain language ("all NDAs from Q4 with Acme"), exact phrases, or concepts — results show which path matched.
Claude Haiku proposes metadata; humans accept and the audit log credits the human, not the AI.
No folder trees. Documents live once and appear in any number of saved metadata views — matter, project, quarter.
Three independent enforcement points: UI controls, MCP tool guards, retention queue. Held docs cannot be deleted.
Most mutations have an inverse. From /audit, click Revert and Kodori dispatches the inverse — the chain stays intact.
Hash-chained audit, immutable events, sensitivity badges, signed-action attribution, GDPR-ready exports.
Anthropic vision for PDFs and images; pure-JS for Office; native parsing for text/JSON/CSV/XML; Illustrator .ai sniffing.
Cloudflare Email Workers with HMAC-signed inbound API. Forward a thread; Kodori files, classifies, and emits events.
Extract → embed → classify pipeline runs as crash-safe, retryable Inngest jobs with full observability.
Define classes; when retention elapses, docs surface for human-confirmed disposition. Never auto-tombstones.
Bearer tokens, cursor-paginated listings, /me + /search + /documents endpoints, ACL-respecting, OpenAPI manifest.
Neon's preview branches enable per-tenant isolation when enterprise customers require it — without leaving multi-tenant.
First-class RFIs, submittals, inspections, change orders, drawings, and project metadata. Tools to extract drawings + citations and link response documents to RFIs / submittals.
Anomaly agent flags unusual mutation patterns for step-up approval. DLP scanner surfaces sensitivity findings; reviewers decide accept / dismiss / redact, all hash-chained.
Inline annotations with resolve / reopen lifecycle. Threads attach to documents, version-aware, ACL-trimmed, and emit events into the audit chain.
No implementation fees, no per-document charges. Early Access is free during private beta; Stripe billing wires up at GA.
Free
Design partners
$30/user/mo
10–50 seat legal & AEC teams
Custom
Regulated enterprises
Kodori is in private beta with design partners in legal and AEC. Request access to evaluate, or contact us if you'd like a similar regulated-industry AI product built for your firm.
We build production-ready AI applications and SaaS platforms. Let's discuss your next project.