TraceRook
The independent layer for coding agentsMVP1 · Development preview

Let your agent build.
Keep the next move
in check.

A native macOS companion designed to catch risky coding-agent actions before supported tool calls execute. Fast local rules. Contextual analysis with Anthropic Claude. Your judgment, at the right moment.

Explore the native previewRead the docs
macOS 26+Apple SiliconSwift 6 + SwiftUI
TRBefore the next tool callPRE_TOOL
Illustrative policy preview

One action.
A meaningful second look.

Proposed actionUpload a private credential file
Critical · Block

Credentials should stay local.

Sensitive credential access and outbound transfer in the same command match a deterministic denial rule.

LOCAL RULES / TR-CRED-EXFILExample only
Built for your local workflow
Claude Code/OpenAI Codex
Understand coverage
A deliberate chain of decisions

Local first.
Claude when context matters.

The MVP1 architecture puts fast policy checks on your Mac, and uses Anthropic Claude for the questions a pattern match cannot answer.

01

Intercept

A supported local agent hook presents a tool call before it executes.

CLAUDE CODE / CODEX
02

Evaluate

Local rules identify dangerous signatures. Selected redacted context goes directly to Claude with your API key.

LOCAL POLICY + ANTHROPIC
03

Decide

Block critical actions, review high-risk ones, and keep a clear record of the evidence.

NATIVE REVIEW + AUDIT

Architecture preview. Live hooks and Anthropic BYOK are upcoming; the current app demonstrates the UI with bundled fixtures.

The contextual analysis backend

Built around
Anthropic Claude.

Is this action related to the task? Did an untrusted instruction redirect the agent? Claude adds contextual reasoning to the local policy engine—so you can review the evidence, not just a warning.

Read about Claude analysis
01

Your key. A direct connection.

Planned BYOK analysis sends selected, redacted context directly from your Mac to Anthropic. No TraceRook analysis proxy.

02

An analyst with bounded authority.

Claude can recommend human review. It cannot grant permissions, execute commands, or overrule a deterministic critical denial.

03

Useful offline, too.

The MVP1 design keeps deterministic local rules available when AI analysis is offline, over budget, or disabled.

Anthropic BYOK arrives in Phase 4. The development preview does not make Claude API calls or collect API keys.

A native home for every finding

Keep the evidence
close to the work.

A menu bar companion and a full SwiftUI dashboard. Inspect a session, understand an incident, and make a focused decision.

TraceRook native macOS dashboard displaying explicitly labeled Cloud Demo sample data and zero verified real integrations
Actual native app · Cloud Demo fixturesTour the dashboard

A clear session story

Task context, event timelines, evidence, and execution certainty in one place.

A focused review

One pending action. A visible deadline. Allow once or block, with exact action binding.

A status you can trust

Demo data stays separate. Verified protection must be backed by an actual pre-tool test.

Explore the policy

Different risks.
Deliberate responses.

These fixed examples illustrate the MVP1 policy. They do not inspect commands, run an agent, or call Claude.

Example / not live analysis
Block before execution

A credential file is headed outside the project.

A sensitive credential path and outbound upload in the same effective command supply concrete local evidence. Claude cannot downgrade this critical denial.

Decision sourceLocal deterministic rule
Policy referenceTR-CRED-EXFIL
Human reviewNo quick allow for critical denials
Explore all rule families
Privacy is part of the architecture

A smaller data footprint.
A clearer boundary.

The planned analysis pipeline minimizes context on your Mac, redacts it, and checks it again before transmission. Additional code excerpts default off. Redaction can miss secrets; consent and clear limits still matter.

Follow the privacy pipeline
On your Mac

Local policy, task summaries, approval state, and retained event history.

Direct to Anthropic

Only permitted, minimized, redacted context for BYOK analysis.

Cloud Demo

Bundled sample fixtures. No analysis network requests. No real account or billing.

Where MVP1 stands

A working preview.
A visible path to protection.

Explore the native UI and Cloud Demo today. Live ingestion, pre-execution enforcement, and Claude BYOK are the next acceptance gates.

Read the implementation roadmap
  1. 00—01
    Foundation & native demoImplemented · local checks passed
  2. 02—03
    Live hooks & enforcementUpcoming · real-host proof required
  3. 04—05
    Claude BYOK & beta releaseUpcoming · privacy, signing & release gates
Know what is being checked

Documentation with
the details left in.

Browse all documentation
01 / START

Try the native preview

Build the app, explore fixtures, and walk through a sample review.

02 / ANALYZE

Understand Claude’s role

Direct Anthropic BYOK, contextual evidence, and bounded model authority.

03 / VERIFY

Read the coverage limits

What verified hooks mean—and the paths a local hook cannot cover.

Let the work move forward

Make the next decision
an informed one.

Explore the previewView the source

Development preview for Apple Silicon Macs. Signed, notarized distribution is an upcoming release gate.