Know why the agent
wrote every line

Prufs captures the full reasoning chain behind AI-generated code - from your directive through every decision, alternative, and trade-off - as a cryptographically signed, tamper-evident audit trail.

npm install @prufs/sdk

View on GitHub See pricing
Live decision trail
directive
wade: "Add user search to admin panel with typeahead"
interpretation
Implement GET /api/admin/users/search with debounced typeahead using existing Elasticsearch index 92% confidence
constraint
Must use existing API auth middleware - no new auth patterns project rule
decision
Chose: Elasticsearch over PostgreSQL FTS rejected and Fuse.js rejected
ES index already exists, supports fuzzy matching. Reusing infra avoids operational burden. 95% confidence
implementation
7 files changed · +445 -2 lines · commit a1b2c3d
verification
226 passed · 0 failed · 94% coverage on new code pass
🔒 Ed25519 signed · hash chain verifiedsigner: 1cd68501

Built for the age of agentic coding

Causal decision graph

Six node types connected by causal edges. Every line of code traces back to the human directive that spawned it, through every decision and constraint along the way.

🔒

Tamper-evident signing

Every event is Ed25519 signed and hash-chained. Modify one event and the entire chain breaks. Verification is independent of the SDK - any tool with the public key can audit the trail.

Decisions with alternatives

Not just what was chosen, but what was rejected and why. Domain-tagged, confidence-scored, and sensitivity-classified. Security decisions are automatically restricted via RBAC.

🔍

Review-via-trail

Replace diff-reading with trail-walking. Reviewers see the directive, interpretation, decisions, and trade-offs - then approve at the reasoning level, not the line level.

💻

VS Code hover-to-trail

Hover over any line of code to see its decision trail. Gutter icons show trail coverage. Right-click to trace any line back to its originating directive.

🤖

Agent self-instrumentation

Ship with a SKILL.md and agents instrument themselves. Hooks for Claude Code and Cursor. SessionObserver auto-detects decisions and constraints from natural language.

Pricing

The SDK is free and open source. Pay for cloud when your team needs shared trails.

Free
$0 forever
Full SDK, local storage, cryptographic signing
  • Unlimited trail recording
  • Ed25519 signing + hash chain
  • Claude Code + Cursor hooks
  • CLI: status, inspect, trace, replay
  • Chain verification
  • MIT licensed, no telemetry
Install SDK
Enterprise
Custom
For regulated industries needing audit-grade trails
  • SSO / SAML authentication
  • Unlimited trail retention
  • Audit export (SOC 2, ISO, OCC)
  • On-premises deployment option
  • Temporal query engine
  • Dedicated support + SLA
  • CRI compliance mapping
Contact sales

Built for regulated industries

Prufs was designed alongside the Composite Reliability Index (CRI) framework for evaluating LLM agent reliability in production banking environments. The decision trail directly implements the Explainability Performance Index (EPI) sub-index - making agent-generated code auditable at the reasoning level, not just the output level.

Request a CRI readiness assessment

Platform Status

Phase 3 in progress — cloud platform and sync engine live

Phase 1
Core SDK & cryptographic signing
Phase 2
Native VCS layer & R2 blob storage
3
Phase 3
Cloud platform, sync engine & team features
SDK on npm
Live
@prufs/sdk · MIT licensed
Cloud platform
Deployed
Cloudflare Workers + D1
R2 blob storage
Live
Commits & blobs on Cloudflare R2
Sync engine
Built
@prufs/sync · bidirectional local/cloud
GitHub packages
12 packages
Monorepo on prufs-ai/prufs
Test suite
226 passing
0 failures · Ed25519 verified