Coding-Agent Safety Pack

Use this when you want a coding agent to onboard AgentGuard into a repo with the fewest moving parts possible.

If you want AgentGuard to generate the pack instead of copying snippets by hand, run:

agentguard skillpack --write

That writes .agentguard.json plus the supported coding-agent instruction files into agentguard_skillpack/ for review. Use --target if you only need one agent's instructions, or --output-dir . if you intentionally want to materialize them directly into a repo.

The goal is always the same:

• keep the first run local
• keep the config auditable
• stop loops, retries, and runaway spend
• avoid hidden network behavior

1. Repo-local defaults

Check in a tiny .agentguard.json file:

{
  "profile": "coding-agent",
  "service": "support-agent",
  "trace_file": ".agentguard/traces.jsonl",
  "budget_usd": 5.0
}

This gives both humans and coding agents the same local defaults without secrets or hosted coupling.

agentguard skillpack generates this file for you.

2. Codex / AGENTS.md snippet

Add this to the target repo's AGENTS.md:

Use AgentGuard in local-only mode first.

- Keep traces in `.agentguard/traces.jsonl`
- Prefer `agentguard.init(local_only=True)`
- Respect the checked-in `.agentguard.json`
- Do not add API keys or hosted settings during first integration
- Verify with:
  1. `agentguard doctor`
  2. `agentguard quickstart --framework raw --write`
  3. `python agentguard_raw_quickstart.py`
  4. `agentguard report .agentguard/traces.jsonl`

3. Claude Code / CLAUDE.md snippet

Add this to the target repo's CLAUDE.md:

When using AgentGuard:

- start local-only
- use the repo's `.agentguard.json`
- keep traces in `.agentguard/traces.jsonl`
- do not add dashboard or network settings in the first PR
- prove the integration by running `agentguard doctor`, `agentguard quickstart --framework raw --write`, and `python agentguard_raw_quickstart.py`

4. GitHub Copilot / .github/copilot-instructions.md snippet

Add this to .github/copilot-instructions.md:

If you touch AI-agent code, keep AgentGuard local-first:

- prefer `agentguard.init(local_only=True)`
- use `.agentguard.json` if present
- avoid adding secrets or hosted settings
- verify changes with `agentguard doctor`, `agentguard quickstart --framework raw --write`, and a local trace report

5. Cursor project rule snippet

Add a rule like .cursor/rules/agentguard.mdc:

---
description: AgentGuard local-first safety
globs:
  - "**/*.py"
alwaysApply: false
---

When integrating AgentGuard:
- keep the first run local
- use `.agentguard/traces.jsonl`
- honor `.agentguard.json`
- avoid API keys and hosted settings in initial setup
- verify with `agentguard doctor` and `agentguard quickstart --framework raw --write` before wiring real providers

6. MCP-capable agent snippet

Only after the local SDK path is proven, MCP-capable coding agents can read retained AgentGuard traces through the published MCP server:

{
  "mcpServers": {
    "agentguard": {
      "command": "npx",
      "args": ["-y", "@agentguard47/mcp-server"],
      "env": {
        "AGENTGUARD_API_KEY": "ag_your_key_here"
      }
    }
  }
}

Keep this out of the first-run path. MCP is for retained trace access later, not for proving local enforcement.

7. Verification sequence

The shortest credible flow is:

pip install agentguard47
agentguard doctor
agentguard quickstart --framework raw --write
python agentguard_raw_quickstart.py
agentguard report .agentguard/traces.jsonl

For a stronger proof run:

agentguard demo
python examples/demo_budget_kill.py
agentguard incident demo_traces.jsonl --format json

8. Boundary reminder

Use the AgentGuard SDK for:

• local enforcement
• local traces
• local reports
• local savings summaries

Use the hosted dashboard later for:

• retained history
• alerts
• dashboard-driven remote kill signals
• team visibility
• governance

HttpSink only mirrors trace and decision events. It does not poll or execute remote kill signals by itself.