Stop guessing what changed in your agent.

AgentLock creates a deterministic lockfile for the models, MCP identities, instruction files, skills, packages, and configuration names that shape agent behavior—without copying secret values.

Download package Verify SHA-256
  • Zero runtime dependencies
  • Deterministic SHA-256 integrity
  • Secret values excluded
PowerShell
> npm install -g https://htmltoolz.com/downloads/agentlock
> agentlock init .

AgentLock wrote agent.lock.json
  MCP servers        7
  instruction files  4
  AI packages        9
  integrity          sha256:9fa3…c12e

> agentlock verify .
✓ no agent dependency drift
Try before install

Generate a safe lock preview in your browser.

Paste a small MCP config, package.json, and instruction excerpt to see the sanitized fields AgentLock records before you install the CLI. The preview hashes instruction text locally and compares against an existing lockfile when supplied.

Awaiting previewLoad the sample or paste your own local config.
0MCP servers 0AI packages 0Instruction hashes 0Drift signals 
{
  "status": "Paste inputs, then generate a browser-local preview."
}
Next step Enforce the lock in pull requests.

After the preview looks right, commit the generated lockfile and add this workflow so dependency drift fails CI before it reaches main.

Generate a lock preview to create .github/workflows/agentlock.yml.

A lockfile for the part of the stack package managers cannot see.

Code can stay unchanged while an MCP schema, instruction file, client configuration, or model identifier changes underneath it. AgentLock turns those invisible dependencies into a reviewable contract.

Scan

Discover MCP identities, AI-related packages, skills, instructions, and configuration names.

Lock

Produce canonical ordering and a stable integrity hash that belongs in source control.

Verify

Fail CI with exit code 2 when the current AI stack drifts from the accepted baseline.

Diff

Explain added, removed, and changed agent-facing dependencies in human or JSON form.

Designed to be committed.

The lock records hashes and identities—not prompt contents, credential values, URL query strings, header values, or command arguments.

agentlock scan . --json
agentlock verify .
agentlock diff agent.lock.json