hhHalfhand

Replay every action your agents took.

Halfhand records prompts, tool calls, MCP interactions, file modifications, terminal commands, and execution history so you can inspect failures, understand decisions, reproduce behavior, and audit autonomous systems.

Open source · Local-first · SQLite-backed

bash
$ hh replay 6f4b82
Session: 6f4b82
Model: claude-opus
00:00:01 AGENT Started session
00:00:02 TOOL Read README.md
00:00:04 MCP github.search_repositories()
00:00:07 TOOL Write src/main.rs
00:00:10 TOOL cargo test
00:00:13 ERROR Test failed
00:00:15 AGENT Generated fix
00:00:21 SUCCESS All tests passed

AI agents fail. Understanding why is hard.

Black Box Execution

Most agent frameworks expose logs, not reasoning. You see what happened, not why it happened.

Missing Context

Tool calls, MCP requests, prompts, and file changes are scattered across systems. No single source of truth.

No Reproducibility

Failures are difficult to replay and investigate later. Context is lost the moment the process exits.

How Halfhand Works

Three simple steps to full agent observability.

Step 01

Record

Wrap any agent execution with a single command.

  • Wraps any CLI agent transparently
  • Zero configuration required
  • Captures full execution context
bash
$ hh run claude-code
Recording session...
Agent: claude-code
Session: a1b2c3
All tool calls, MCP traffic, and file changes will be captured.
Step 02

Replay

Navigate execution timelines step-by-step.

  • Chronological step navigation
  • Jump to any timestamp
  • Keyboard-driven interface
bash
$ hh replay a1b2c3
Replaying session a1b2c3
Controls:
[n]ext — advance one step
[p]revious — go back one step
[j]ump — jump to timestamp
[q]uit — exit replay
Step 03

Understand

Inspect every detail of agent behavior.

  • Prompt history with full context
  • Tool outputs and MCP traffic
  • Unified diff for file changes
bash
$ hh inspect a1b2c3 --step 7
Step 7: TOOL Write src/main.rs
--- src/main.rs
+++ src/main.rs
@@ -1,3 +1,4 @@
fn main() {
+ println!("Hello, Halfhand!");
}

Everything you need to debug agents

Purpose-built observability for autonomous systems.

Deterministic Replay

Replay historical executions with full context. Every prompt, tool call, and state change is preserved so you can step through failures exactly as they happened.

Execution Timeline

View agent decisions chronologically. Navigate through events with millisecond precision to understand the exact sequence of actions.

MCP Visibility

Inspect Model Context Protocol requests and responses. See exactly what data was exchanged between your agent and external systems.

File Diff Inspection

See exactly what changed. Unified diffs show every file modification your agent made, with before and after context.

Agent Audit Trail

Maintain historical execution records. Build trust in autonomous systems with complete, queryable logs of every decision.

Local-First

Data remains under your control. All execution history is stored locally, never uploaded to external servers.

SQLite Storage

Portable, queryable execution history. Use SQL to analyze agent behavior patterns across thousands of runs.

Terminal Native

Built for developers who live in the terminal. Keyboard shortcuts, vim-like navigation, and zero GUI overhead.

CLI-first interface

Three core commands. Zero learning curve.

bash
$ hh replay last
Loading last session...
Session: 8d9f1a
Model: claude-sonnet-4-20250514
Duration: 00:04:32
Controls:
n — next step
p — previous step
j — jump to timestamp
q — quit replay
[step 1/42] 00:00:00 AGENT session started
hh run

Record any agent

hh replay

Step through history

hh inspect

Analyze sessions

Halfhand vs. Traditional Logs

Observability designed for agents, not applications.

FeatureHalfhandTraditional Logs
Replay execution
Inspect tool callsPartial
Track file changes
MCP visibility
Deterministic replay
Agent audit trail

The flight recorder for AI agents.

As AI systems become responsible for more production work, debugging and observability become critical. Halfhand records prompts, tool calls, MCP traffic, file modifications, terminal commands, and execution history so developers can replay failures, audit behavior, and understand exactly what happened.

100%
Local-first

Your data stays on your machine

SQLite
Portable storage

Query with SQL, backup with tools you know

Replay
Deterministic

Exact reproduction of every execution

Built for developers

Start locally. Integrate with existing workflows. Keep ownership of your execution history.

Start locally

Install via cargo, run on your machine. No cloud account, no tokens, no API keys required.

Integrates with CI/CD

Export sessions as artifacts. Replay failures in CI pipelines. Generate reports for PRs.

Keep ownership

Your execution data never leaves your machine. Full privacy and compliance from day one.

What developers are saying

“This is super cool! It's like a VCR for LLMs.”

Stop guessing why your agents failed.

Replay every action. Inspect every decision. Built for the engineers shipping AI systems.

Install Halfhand
bash
$ cargo install halfhand

Requires Rust 1.75+ · macOS, Linux, Windows · Other install methods