Use the public OSS package today. Start each task with a task contract, let the agent edit within scope, and run the review check before you trust the change.
npx agent-guardrails plan --task "..."
Usage guide
Use it as the release companion for the AI agent you already like.
Use it alongside Codex, Claude Code, OpenCode, Cursor, or your usual terminal. Agent Guardrails keeps the release side clear.
Start here
Two paths, one workflow.
The core workflow stays the same. OSS gives you the bounded review check. Pro adds the local Workbench, the next check to run, and a readable report when it has been enabled for you.
Follow the setup steps you received, then open the local Workbench after each guarded check. The Workbench should answer the release question before you dig into details.
npx agent-guardrails pro workbench --open
Who does what
The user decides. The agent does the mechanical work.
The coding agent runs the commands and checks. The human reviews the result and decides whether to ship.
Define the task, review the Workbench or report, decide whether the evidence is good enough, and choose whether to merge or deploy.
Stay in scope, run the proof command, save evidence under `.agent-guardrails/evidence/`, rerun the check, and report whether the verdict changed.
The real flow
Five steps from prompt to release decision.
The agent handles the repetitive part. The human reviews the result and the remaining risk.
Bound the task
Create a task contract before the AI agent starts editing. This records intended files, allowed paths, required commands, and evidence paths.
npx agent-guardrails plan --task "..."
Let the agent work
Use Codex, Claude Code, OpenCode, Cursor, or another agent normally. Agent Guardrails works alongside that workflow.
Keep unrelated edits out of scope
Run the check
Run the check after code changes. OSS catches scope and validation issues. Pro adds a clearer ship answer when enabled.
npx agent-guardrails check --review
Open Workbench
Read the local Workbench first. It shows "Can I ship?", the one next action, and an agent handoff prompt.
npx agent-guardrails pro workbench --open
Capture proof
Ask the agent to run or watch the proof, save evidence, rerun the check, and report whether the verdict changed.
Receipt stays with the repo
Command load
Today, the human usually types two or three commands.
Most sessions only need two or three commands from the human. The repeated work can stay with the agent.
Create the task contract with npx agent-guardrails plan --task "...".
Run npx agent-guardrails check --review after the agent finishes the scoped change.
Open npx agent-guardrails pro workbench --open, read the next action, then hand the mechanical work back to the agent.
Product views
What opens after the check.
These are the two views most people look at first.
Agent handoff
Give the agent one concrete next step.
The prompt tells the agent what to run, where to save evidence, and how to report back after the rerun.
The repo keeps the decision and the evidence together.
The Workbench and report explain what changed, why the release is still blocked, and what to do next.
If your team automates around this, the same information is available in structured output.
MCP integration
When the agent supports MCP, the loop gets smoother.
If your agent supports MCP, it can read the current status, save evidence notes, and record the result without manual copy-paste.
`pro_read_workbench` returns the current release decision, proof queue, and agent handoff.
`pro_capture_evidence_note` appends structured proof notes and artifact paths into the repo evidence file.
`pro_record_proof_outcome` records whether the proof actually helped, so future suggestions become more useful over time.