Reference

Lookup tables for the DefenseClaw CLI, the sidecar API, configuration files, credentials, fail modes, and environment variables.

Lookup pages for after you know what you are doing. defenseclaw <command> --help is the authority for any flag.

To build a non-interactive defenseclaw setup guardrail command, use the command generator.

For coding agents

These pages are runbooks written to an AI agent. Pick the task, copy its prompt into your agent, and the agent reads the runbook page and does the work. Every runbook tells the agent to look first, change nothing without your yes, and report back in a fixed format.

Pick a task

TaskPrompt to pasteRunbook
Install and set up in observe modeRead https://cisco-ai-defense.github.io/defenseclaw/docs/get-started/quickstart/llms.md and follow it.Set up DefenseClaw
Find out why something is brokenRead https://cisco-ai-defense.github.io/defenseclaw/docs/reference/llms.md, open the "Troubleshoot DefenseClaw" runbook and follow it.Troubleshoot
Run an agent in a sandboxRead https://cisco-ai-defense.github.io/defenseclaw/docs/sandboxes/guide/llms.md and follow it.Sandbox an agent
Turn a sentence into a ruleRead https://cisco-ai-defense.github.io/defenseclaw/docs/policies/recipes/llms.md and follow it.Write a rule
Deploy to a fleet as an administratorRead https://cisco-ai-defense.github.io/defenseclaw/docs/enterprise/operations/llms.md and follow it.Enterprise admin

Machine-readable docs

  • https://cisco-ai-defense.github.io/defenseclaw/llms.txt lists every page with its Markdown URL.
  • https://cisco-ai-defense.github.io/defenseclaw/llms-full.txt is the whole documentation in one file.
  • Add llms.md to any page URL for that page as Markdown, for example .../docs/get-started/install/llms.md.

Rules every runbook enforces

  1. Observe first. The agent does not switch to action mode, change a fail mode, disable a connector, unblock a site or edit config.yaml by hand without your yes to that exact change.
  2. The agent never prints a key or token.
  3. On a managed computer the agent stops. A command that prints This computer is managed or This device is managed (exit 2 or 3), or "deployment_mode": "managed_enterprise" in defenseclaw status --json, means your administrator owns DefenseClaw there (Managed command availability).
  4. The agent reports in a fixed format: what it ran, what it saw, what it changed, what is left for you.

Hand this page to your agent

Paste this into your agent:

Read https://cisco-ai-defense.github.io/defenseclaw/docs/reference/llms.md, then ask me which task I want and open that runbook.

Troubleshoot DefenseClaw

Hand this page to your agent

Paste this into your agent:

Read https://cisco-ai-defense.github.io/defenseclaw/docs/reference/llms.md, open the "Troubleshoot DefenseClaw" runbook and follow it. Collect first, ask me before any fix that changes something.

You are an AI agent. Find out why DefenseClaw is not working, map each failure to a known fix, apply only the safe ones the user approves, and report.

Hard rules

  1. Observe first. Do not switch to action mode, change a fail mode, disable a connector, unblock a site or edit config.yaml by hand without the user's yes to that exact change.
  2. Never print a key, token, .env, hooks/ or device.key. Redact them from anything you quote.
  3. Stop on a managed computer: a command that prints This computer is managed or This device is managed (exit 2 or 3), or "deployment_mode": "managed_enterprise" in defenseclaw status --json, means the administrator owns DefenseClaw (Managed command availability). Tell the user to contact them.
  4. Do not open audit.db or attach it to a report.

Steps

  1. Collect health. Read-only.

    defenseclaw doctor --json

    Expected: JSON with "outcome", "exit_code", "passed", "failed", "warned", "skipped" and a checks list. Each check has check_id, status (pass, warn, fail, skip), label, detail and remediation. If the command is not found, DefenseClaw is not installed or ~/.local/bin is not on PATH; see Set up DefenseClaw.

  2. Collect state. Read-only.

    defenseclaw status --json
    defenseclaw-gateway status

    Expected: sidecar.running is true; each connector shows its mode and fail_mode.effective. If config_errors is present, run defenseclaw config validate and quote the field it names.

  3. Read the last gateway log lines, locally, and quote only the error lines.

    tail -n 40 ~/.defenseclaw/gateway.log

    Expected: start lines and no repeated errors. If it differs: redact secrets before quoting. Never paste the whole file.

  4. Map each fail and warn check by its label or message. The table below has the common ones; for any other row, run the Next step: line doctor prints under it (Read the rows). Installer and upgrade errors are on Troubleshooting. Common ones:

    Check saysSafe fix (ask first)Needs the user's decision
    gateway not runningdefenseclaw-gateway startNone
    stale generated script, hooks not founddefenseclaw setup <connector> --yes and restart the agentNone
    no gateway token configureddefenseclaw doctor --fix --dry-run, then --fix --yesRotating tokens
    config.yaml invalidnone: show the field it namesThe edit itself
    nothing is blockeddefenseclaw guardrail statusSwitching to action mode
    tools blocked while gateway downstart the gatewayChanging the fail mode
    "DefenseClaw is not checking this session"start the gateway (defenseclaw-gateway start), then defenseclaw doctor if it does not startSwitching the policy
    doctor.cowork warns, or status --json has cowork.protected: falseRun the command in cowork.step when it is a defenseclaw command (for example defenseclaw setup cowork). See Claude Cowork for each reason_codeUploading the plugin zip in Claude Desktop, or removing the plugin there: only the user can
  5. Preview repairs. Never skip this.

    defenseclaw doctor --fix --dry-run

    Expected: a list of the repairs that would run, with their risk. Nothing changes. If a repair is disruptive (it restarts the gateway), say so before asking.

  6. Apply only what the user approved, one repair at a time, then re-check.

    defenseclaw doctor --fix --yes --fix-id <REPAIR_ID>
    defenseclaw doctor --json

    Expected: the repair is applied and the check now has "status": "pass". If it differs: stop, do not try another repair, and report both outputs.

Report

Reply with:

  • Symptom the user gave, and the OS.
  • Failing checks: label, detail and the Troubleshooting entry you matched.
  • Fixes applied (command, result) and fixes you proposed but did not apply.
  • Decisions that belong to the user.
  • Whether doctor --json now exits 0.