ACP

Record, Replay, and Test AI Agent Behavior

A local-first developer tool for recording, replaying, and testing AI agent behavior. Make agent execution deterministic, observable, and testable.

Read the DocsView on GitHubWatch Demo
By Hyper Grey / 2026
01 / What It Does

CORE WORKFLOW

No. 01

Instrument

Decorator-based instrumentation captures every LLM call, tool invocation, and state change. Zero-overhead recording with automatic redaction.

Learn More→
No. 02

Debug

Timeline visualization, state inspector, and diff viewer for forensic debugging. Read-only replay from recorded artifacts.

Learn More→
No. 03

Analyze

Inspect, replay, test, and analyze commands. YAML-based behavioral tests and automated pattern detection.

Learn More→
No. 04

Replay

Same trace produces identical execution every time. No API calls during replay—purely from recorded data.

Learn More→
No. 05

Inspect

Every input, output, and state transition is captured. Step through the agent's reasoning at any decision point.

Learn More→
02 / Components

THE ACP TOOLKIT

Instrumentation

Python SDK

Decorator-based recording. @acp.tool() and @acp.llm_wrapper capture every execution automatically.

01
Visual Debugging

VS Code Extension

Timeline view, state inspector, diff viewer, and run comparison for forensic debugging.

02
YAML Tests

Behavioral Testing

Assert on patterns, tool usage, and outcomes. Tests catch regressions without depending on exact text.

03
Command Line

CLI Tools

acp inspect, replay, test, and analyze. All commands work on recorded artifacts—no agent execution.

04
Zero Side Effects

Deterministic Replay

Same trace = identical execution. Replay never calls external APIs. Read-only, side-effect free.

05
Pattern Detection

Automated Analysis

Detects high step count, memory growth, repeated calls, and error rates automatically.

06
03 / Success Metrics

WHAT WE PROVE

01 / DETERMINISTIC

DETERMINISTIC REPLAY

Same trace produces same replay behavior. Same final state after replay. Every time, guaranteed.

02 / FULL

FULL INSPECTION

Any step can be inspected. Inputs, outputs, and state visible at every decision point.

03 / BEHAVIORAL

BEHAVIORAL TESTS

Tests catch logic regressions. Tests don't depend on exact text. YAML-based definitions.

04 / FAILURE

FAILURE ANALYSIS

Tool highlights inefficiencies. Broken agent scenario exists. Tool explains why it broke.

04 / Get Started

QUICK START

1. Run Agent

  • npm start
  • or --broken

2. Inspect

  • npm run inspect
  • traces/file.json

3. Test

  • npm run test
  • tests/basic.yaml

4. Analyze

  • npm run analyze
  • traces/file.json

Links

License

  • MIT
  • Open Source

© 2026 Hyper Grey. All rights reserved.

Agent Control Plane — Built with precision.