callguard 0.3.0

Runtime safety for AI agents. Stop agents before they break things.

CallGuard

Runtime contracts for AI agents.

AI agents make tool calls. Tool calls have side effects. Nobody governs what happens between "agent decides" and "tool executes." CallGuard is that governance layer — preconditions, postconditions, session limits, and a full audit trail, enforced at the point where decision becomes action.

Show Me

contracts.yaml

apiVersion: callguard/v1
kind: ContractBundle

metadata:
  name: my-policy

defaults:
  mode: enforce

contracts:
  - id: block-sensitive-reads
    type: pre
    tool: read_file
    when:
      args.path:
        contains_any: [".env", ".secret", "credentials", ".pem", "id_rsa"]
    then:
      effect: deny
      message: "Sensitive file '{args.path}' blocked."
      tags: [secrets, dlp]

Python

import asyncio
from callguard import CallGuard, CallGuardDenied

async def main():
    guard = CallGuard.from_yaml("contracts.yaml")

    try:
        result = await guard.run("read_file", {"path": "/app/config.json"}, read_file_fn)
        print(result)
    except CallGuardDenied as e:
        print(f"Denied: {e}")

asyncio.run(main())

CLI

$ callguard validate contracts.yaml
✓ contracts.yaml — 1 contract (1 pre)

$ callguard check contracts.yaml --tool read_file --args '{"path": ".env"}'
⛔ DENIED by block-sensitive-reads
   Message: Sensitive file '.env' blocked.
   Tags: secrets, dlp
   Rules evaluated: 1

Framework integration (one adapter, same guard)

from callguard.adapters.langchain import CallGuardMiddleware

middleware = CallGuardMiddleware(guard)
# Wraps any LangChain tool — preconditions, audit, and session limits apply automatically

Features

• YAML contracts — Preconditions, postconditions, and session limits declared in version-controlled YAML files
• 6 framework adapters — LangChain, CrewAI, Agno, Semantic Kernel, OpenAI Agents SDK, Claude Agent SDK
• Audit trail — Structured JSON events with automatic redaction of secrets (OpenAI keys, AWS creds, JWTs, GitHub tokens)
• Observe mode — Shadow-deploy contracts without blocking; review CALL_WOULD_DENY events before enforcing
• CLI tooling — validate, check, diff, and replay commands for CI/CD integration
• Principal context — Role, ticket ref, and claims propagated through every decision and audit event
• Session limits — Cap total calls, attempts, and per-tool executions to catch runaway agents
• Zero runtime deps — Pure Python 3.11+. OTel, sinks, and adapters are optional extras

How It Compares

Approach	Scope	Runtime enforcement	Audit trail
Prompt/output guardrails	Input/output text	No — advisory only	No
API gateways / MCP proxies	Network transport	Yes — at the proxy	Partial
Security scanners	Post-hoc analysis	No — detection only	Yes
Manual if-statements	Per-tool, ad hoc	Yes — scattered logic	No
CallGuard	Tool call contracts	Yes — deterministic pipeline	Yes — structured + redacted

Install

pip install callguard              # core (zero deps)
pip install callguard[yaml]        # + YAML contract engine
pip install callguard[sinks]       # + webhook, Splunk, Datadog sinks
pip install callguard[cli]         # + validate/check/diff/replay CLI
pip install callguard[all]         # everything

Built-in Templates

guard = CallGuard.from_template("file-agent")      # secret file protection, destructive cmd blocking
guard = CallGuard.from_template("research-agent")   # output PII detection, session limits
guard = CallGuard.from_template("devops-agent")     # role gates, ticket requirements, bash safety

Links

• Documentation
• GitHub
• PyPI
• Changelog
• License (MIT)