GitHub - Diplomat-ai/diplomat-agent: What can your AI agent do to the real world? Scan your code. See which tool calls have zero checks

You deployed a Python AI agent. Do you know every function it can call that writes to a database, sends an email, charges a card, or deletes data — and which ones have zero checks?

diplomat-agent runs a static AST scan and tells you exactly that. Zero dependencies. 2 seconds on a 1,000-file repo.

pip install diplomat-agent
diplomat-agent scan .

What it looks like

diplomat-agent — governance scan

Scanned: ./my-agent
Tool calls with side effects: 12

⚠ process_refund(amount, customer_id)
  Write protection:       NONE
  Rate limit:             NONE
  → stripe.Refund.create() with no amount limit
  Governance: ❌ UNGUARDED

⚠ delete_user_data(user_id)
  Confirmation step:      NONE
  Batch protection:       NONE
  → session.delete() with no confirmation
  Governance: ❌ UNGUARDED

✓ update_order(order_id)
  Governance: ✅ GUARDED

────────────────────────────────────────────
RESULT: 8 unguarded · 3 partial · 1 guarded (12 total)

Why this matters for AI agents

In a web app, a human clicks a button. The UI has validation, confirmation dialogs, rate limits per session.

In an agent, an LLM decides which functions to call, with what arguments, how many times. It doesn't know your business rules. It can loop, hallucinate arguments, or get prompt-injected.

Without guards in the code, there's nothing between the LLM's decision and the real-world consequence.

We scanned 16 open-source agent repos. 70.9% of analyzable tool calls have no guard, with a 1.7% opacity rate — measured with inter-procedural tracing across 7,552 tool calls.

What it detects

40+ patterns across 8 categories:

What counts as a guard: input validation, rate limiting, auth checks, confirmation steps, idempotency keys, retry bounds. Full list →

Integrate everywhere

CI — block unguarded PRs

- name: Diplomat governance scan
  run: |
    pip install diplomat-agent
    diplomat-agent scan . --fail-on-unchecked

IDE — review what the copilot wrote

Works in your IDE with zero extension to install:

Pre-commit hook

repos:
  - repo: https://github.com/Diplomat-ai/diplomat-agent
    rev: v0.5.3
    hooks:
      - id: diplomat-agent

SARIF — native VS Code Problems panel

diplomat-agent scan . --format sarif --output results.sarif

Open with SARIF Viewer. Or upload to GitHub Code Scanning.

Scan only changed files

diplomat-agent scan . --diff-only

Generate your agent's SBOM

diplomat-agent scan . --format registry --output-registry toolcalls.yaml

Like requirements.txt — but for what your agent can do, not what it depends on. Commit it. Diff it in PRs. When your agent gains a new capability, the change shows up in review.

What is a Behavioral BOM →

Benchmarks

Application layer: ~62% unguarded across 2,943 tool calls in 9 repos (weighted, v0.5.0 with inter-procedural tracing). Frameworks sit higher — absence of guards there is by design. We scan both identically. Large repos (>400 tool calls) take longer with inter-procedural tracing (e.g. CrewAI ~38s).

Full results on 16 repos →

Verdicts

OPAQUE is honest, not alarming. It means the scanner reached the boundary of what static analysis can see (callable passed to an executor, remote MCP call, unresolvable dispatcher branch). Use # checked:ok — [reason] once reviewed.

Output formats

Acknowledge a tool call

If a function is intentionally unguarded or protected elsewhere:

def send_alert(message):  # checked:ok — protected by API gateway
    requests.post(ALERT_URL, json={"msg": message})

From scanning to runtime

diplomat-agent finds what your agent can do. diplomat-gate stops it from doing the dangerous parts at runtime.

# Step 1 — find what your agent can do
pip install diplomat-agent
diplomat-agent scan .
# → 12 unguarded tool calls (8 payments, 4 emails)

# Step 2 — protect them at runtime
pip install "diplomat-gate[yaml]"
# → write gate.yaml, wrap your tools with @gate

from diplomat_gate import Gate

gate = Gate.from_yaml("gate.yaml")
verdict = gate.evaluate({"action": "charge_card", "amount": 15000})
# verdict.decision  → STOP
# verdict.violations → [{"policy": "amount_limit", "message": "Amount 15000 exceeds limit of 10000"}]

15+ pre-built policies (payments, emails, shell commands). CONTINUE / REVIEW / STOP in < 1ms. Zero dependencies.

diplomat-gate → · diplomat.run → (hosted control plane with hash-chained audit trail)

Standards alignment

• OWASP Agentic Top 10 mapping →
• EU AI Act / NIST / DORA compliance →
• CSAF 2.0 advisory generation →

Known limitations

• Static analysis only — no runtime detection
• Python only — TypeScript on the roadmap
• Inter-procedural tracing: same-package top-level functions (depth 2). Class methods, cross-package chains, and depth > 2 are not resolved — use # checked:ok for guards in those paths or external packages
• MCP scanning: Python only (FastMCP / official SDK) — TypeScript/Node MCP servers are out of scope
• MCP scanning: transport-layer auth (OAuth, token gateway) is invisible — "unguarded" means no guard inside the tool function, independent of transport
• MCP scanning: @mcp.tool attribute decorator only — bare @tool (from direct import) is not detected
• MCP scanning: @server.call_tool() low-level dispatcher is resolved when handler branches are in scan scope; unresolved/out-of-scope branches are surfaced as OPAQUE
• Full limitations →

Roadmap

• Python AST scanner (40+ patterns)
• toolcalls.yaml behavioral SBOM
• CSAF 2.0 + SARIF 2.1.0 output
• CI integration (--fail-on-unchecked)
• IDE agents (Copilot Chat, Claude Code, Cursor)
• Pre-commit hook
• --diff-only and --file modes
• Inter-procedural tracing: decorators + same-package call chains (depth 2)
• MCP server scanning
• TypeScript support
• VS Code extension (inline diagnostics on save)
• PR comment integration

Requirements

• Python 3.9+
• Zero dependencies (stdlib ast only)
• Optional: rich (colored output), pyyaml (registry)

Contributing

Detection patterns are data, not logic — adding a new pattern means adding an entry to a list in scanner/patterns.py. No AST knowledge required.

See CONTRIBUTING.md for how to add patterns, report false positives, and run the test suite.

License

Apache 2.0