Transcript Sentinel

Transcript Sentinel is an interactive terminal UI and secret scanner for auditing leaked credentials in AI agent transcripts, chat history, session logs, and local agent state.

It is designed for a specific problem that normal repository secret scanners miss:

• secrets leaked into agent conversation history;
• credentials buried in history.jsonl, session files, and tool state;
• transcript archives mixed with examples, docs, plugins, and cache files;
• noisy generic regex matches that slow down incident response and credential rotation.

What Transcript Sentinel does

Transcript Sentinel helps security engineers, red teamers, AI platform teams, and builders audit:

• AI agent transcripts
• LLM chat history
• local agent session storage
• agent-side credential files
• secret exposure in coding agent workflows
• credential leakage in prompt engineering and tool use

This makes it useful for:

• secret scanning for AI agents
• transcript security reviews
• leaked API key detection
• credential rotation workflows
• incident response for agent systems
• local security audits for developer AI tools

Features

• Scan common AI agent directories in the user home folder
• Scan a custom directory with -dir
• Detect provider-specific secrets for OpenAI, Anthropic, GitHub, AWS, Slack, Stripe, Discord, Google, Heroku, and private keys
• Detect generic secret assignments with context-aware false-positive suppression
• Classify findings by confidence
• Deduplicate repeated findings in the TUI
• Show provider summaries for fast triage
• Interactive Bubble Tea TUI with:
  • target selection before scanning
  • cached scan reuse for the same target selection
  • explicit rescan and cache clearing
  • scan progress with active file activity
  • multi-pane results view
  • keyboard and basic mouse support

Why this exists

Traditional secret scanners are tuned for source code repositories and CI pipelines. AI agent environments behave differently:

• agents persist prompts, tool outputs, and credentials in local histories;
• secrets often appear in transcript artifacts instead of source files;
• examples and plugin docs create large amounts of false-positive noise;
• responders need a triage-first interface, not just a wall of regex matches.

Transcript Sentinel is built for agent transcript auditing first, repository-style scanning second.

Installation

Run directly

go run .

Build a local binary

go build -o transcript-sentinel .

Usage

Scan default AI agent directories

go run .

Scan a custom directory

go run . -dir /path/to/scan

Open the interactive terminal UI

go run . -tui

Show full secret values

go run . -show-secrets

Enable verbose scanning logs

go run . -verbose

TUI workflow

The TUI now works in three stages:

1. Select which agent directories you want to scan.
2. Run the scan with live progress, queued/completed counters, and active file activity.
3. Review results in a three-pane layout: target selection, findings list, and finding details.

The TUI also caches results per target selection:

• enter opens cached results if the exact same selection was already scanned
• r forces a fresh scan
• c clears the cache for the current selection

TUI controls

Global

• q: quit

Selection screen

• up / down or j / k: move
• space or x: toggle a target
• A: select all targets
• u: unselect all targets
• enter: open cached results or start a scan
• r: force rescan for the current selection
• c: clear cached results for the current selection
• mouse click: toggle a target
• mouse wheel: scroll target selection

Scanning screen

• esc: cancel the current scan and return to selection
• q: quit

Results screen

• tab: switch focus between targets, findings, and details
• up / down or j / k: move in focused list
• pgup / pgdown: jump through findings
• a: show all findings
• h: show high-confidence findings
• m: show medium-confidence findings
• l: show low-confidence findings
• s: toggle masked vs full secret display
• b: go back to target selection
• enter: open cached results or start a new scan from the current target selection
• r: force rescan for the current target selection
• c: clear cache for the current target selection
• mouse click: change focus or select items
• mouse wheel: scroll targets, findings, or details depending on focus

Example use cases

• Audit Claude, Codex, Gemini, or Copilot local history for leaked API keys
• Review AI coding agent workspaces before rotating credentials
• Triage leaked GitHub, OpenAI, or Anthropic tokens found in agent sessions
• Investigate whether a prompt or tool output exposed credentials in local agent state
• Run a local incident-response pass on developer AI folders after a red-team exercise

SEO and discovery terms

If you are sharing the project, these are the problem spaces it addresses:

• AI agent secret scanner
• LLM transcript security
• prompt leak detection
• leaked API key scanner
• agent history credential audit
• terminal UI secret scanner
• coding agent security tool
• transcript-based secret detection
• local AI tool security audit
• credential exposure in agent workflows

Roadmap

• richer provider and rule analytics
• stronger mouse interactions and pane-aware scrolling
• export formats for incident response and ticketing
• rotation playbooks and remediation helpers
• persistent cache between runs
• more transcript-specific heuristics and suppressions

Development

Run tests with a writable Go build cache:

env GOCACHE=/tmp/gocache go test ./...

License

Apache-2.0. See LICENSE.