Aegis

The simplest way to govern AI agent actions. No infra. No lock-in. Just Python.

pip install agent-aegis → YAML policy → governance in 5 minutes.
Works with LangChain, CrewAI, OpenAI, Anthropic, MCP, and more.

Try it Live in Your Browser • Quick Start • How It Works • Documentation • Integrations • Contributing

English • 한국어

The Problem

AI agents are getting real-world access. Without governance, a hallucinating agent can:

• Bulk-delete your CRM contacts
• Submit wrong forms to government portals
• Trigger irreversible API calls at 3am
• Run up cloud bills with infinite loops

There's no sudo for AI agents. Until now.

The Solution

Aegis is a Python middleware that sits between your AI agent and the actions it takes. It's not a separate server you have to run -- you import it directly into your agent code and it wraps every action with policy checks, approval gates, and audit logging.

Your Agent                    Aegis                         Real World
    |                           |                               |
    |-- "delete all users" ---> |                               |
    |                      [Policy check]                       |
    |                      risk=CRITICAL                        |
    |                      approval=BLOCK                       |
    |                           |--- X (blocked, logged) -----> |
    |                           |                               |
    |-- "read contacts" ------> |                               |
    |                      [Policy check]                       |
    |                      risk=LOW                             |
    |                      approval=AUTO                        |
    |                           |--- execute (logged) --------> |
    |                           |                               |
    |-- "bulk update 500" ----> |                               |
    |                      [Policy check]                       |
    |                      risk=HIGH                            |
    |                      approval=APPROVE                     |
    |                           |--- ask human (Slack/CLI) ---> |
    |                           |<-- "approved" --------------- |
    |                           |--- execute (logged) --------> |

Copy, paste, run — zero config needed:

from aegis import Action, Policy

policy = Policy.from_dict({
    "version": "1",
    "defaults": {"risk_level": "low", "approval": "auto"},
    "rules": [{"name": "block_delete", "match": {"type": "delete_*"},
               "risk_level": "critical", "approval": "block"}]
})

safe = policy.evaluate(Action(type="read_users", target="db"))
print(safe.approval)   # Approval.AUTO  ✅

danger = policy.evaluate(Action(type="delete_users", target="db"))
print(danger.approval)  # Approval.BLOCK 🚫

Or with a YAML file — 3 lines:

from aegis import Action, Policy, Runtime

runtime = Runtime(executor=your_executor, policy=Policy.from_yaml("policy.yaml"))
results = await runtime.run_one(Action("write", "salesforce", params={...}))

No servers to deploy. No Kubernetes. No vendor lock-in. One pip install, one YAML file, and your agent has policy checks, human approval gates, and a full audit trail — across any AI provider.

How It Works

Core Concepts

Aegis has 3 key components. You need to understand these to use it:

The Pipeline

Every action goes through 5 stages. This happens automatically -- you just call runtime.run_one(action):

1. EVALUATE    Your action is matched against policy rules (glob patterns).
               → PolicyDecision: risk level + approval requirement + matched rule

2. APPROVE     Based on the decision:
               - auto:    proceed immediately (low-risk actions)
               - approve: ask a human via CLI, Slack, Discord, Telegram, webhook, or email
               - block:   reject immediately (dangerous actions)

3. EXECUTE     The Executor carries out the action.
               Built-in: Playwright (browser), httpx (HTTP), LangChain, CrewAI, OpenAI, Anthropic, MCP
               Custom: extend BaseExecutor (10 lines)

4. VERIFY      Optional post-execution check (override executor.verify()).

5. AUDIT       Every decision and result is logged to SQLite automatically.
               Export: JSONL, webhook, or query via CLI/API.

Two Ways to Use

Option A: Python library (most common) -- no server needed.

Import Aegis into your agent code. Everything runs in the same process.

runtime = Runtime(executor=MyExecutor(), policy=Policy.from_yaml("policy.yaml"))
result = await runtime.run_one(Action("read", "crm"))

Option B: REST API server -- for non-Python agents (Go, TypeScript, etc.).

pip install 'agent-aegis[server]'
aegis serve policy.yaml --port 8000

curl -X POST localhost:8000/api/v1/evaluate \
  -d '{"action_type": "delete", "target": "db"}'
# => {"risk_level": "CRITICAL", "approval": "block", "is_allowed": false}

Approval Handlers

When a policy rule requires approval: approve, Aegis asks a human. You choose how:

Audit Trail

Every action is automatically logged to a local SQLite database. No setup required.

aegis audit                              # View all entries
aegis audit --risk-level HIGH            # Filter by risk
aegis audit --tail                       # Live monitoring (1s poll)
aegis stats                              # Statistics per rule
aegis audit --format jsonl -o export.jsonl  # Export

Quick Start

pip install agent-aegis

1. Generate a policy

aegis init  # Creates policy.yaml with sensible defaults

# policy.yaml
version: "1"
defaults:
  risk_level: medium
  approval: approve

rules:
  - name: read_safe
    match: { type: "read*" }
    risk_level: low
    approval: auto

  - name: bulk_ops_need_approval
    match: { type: "bulk_*" }
    conditions:
      param_gt: { count: 100 }  # Only when count > 100
    risk_level: high
    approval: approve

  - name: no_deletes
    match: { type: "delete*" }
    risk_level: critical
    approval: block

2. Add to your agent

import asyncio
from aegis import Action, Policy, Runtime
from aegis.adapters.base import BaseExecutor
from aegis.core.result import Result, ResultStatus

class MyExecutor(BaseExecutor):
    async def execute(self, action):
        print(f"  Executing: {action.type} -> {action.target}")
        return Result(action=action, status=ResultStatus.SUCCESS)

async def main():
    async with Runtime(
        executor=MyExecutor(),
        policy=Policy.from_yaml("policy.yaml"),
    ) as runtime:
        plan = runtime.plan([
            Action("read", "crm", description="Fetch contacts"),
            Action("bulk_update", "crm", params={"count": 150}),
            Action("delete", "crm", description="Drop table"),
        ])
        print(plan.summary())
        results = await runtime.execute(plan)

asyncio.run(main())

3. See what happened

aegis audit

  ID  Session       Action        Target   Risk      Decision    Result
  1   a1b2c3d4...   read          crm      LOW       auto        success
  2   a1b2c3d4...   bulk_update   crm      HIGH      approved    success
  3   a1b2c3d4...   delete        crm      CRITICAL  block       blocked

Features

Real-World Use Cases

Policy Templates

Pre-built YAML policies for common industries. Copy one, customize it, deploy:

policy = Policy.from_yaml("policies/crm-agent.yaml")

Production Ready

Compliance & Audit

Aegis audit trails provide evidence for regulatory and internal compliance:

Export as JSONL, query via CLI/API, or stream to external SIEM via webhook. For defense-in-depth with container isolation, see the Security Model guide.

Integrations

Works with the agent frameworks you already use:

pip install langchain-aegis               # LangChain (standalone integration)
pip install 'agent-aegis[langchain]'      # LangChain (adapter)
pip install 'agent-aegis[crewai]'         # CrewAI
pip install 'agent-aegis[openai-agents]'  # OpenAI Agents SDK
pip install 'agent-aegis[anthropic]'      # Anthropic Claude
pip install 'agent-aegis[httpx]'          # Webhook approval/audit
pip install 'agent-aegis[playwright]'     # Browser automation
pip install 'agent-aegis[server]'         # REST API server
pip install 'agent-aegis[all]'            # Everything

pip install langchain-aegis

from langchain_aegis import govern_tools

# Add governance to existing tools — no other code changes
governed = govern_tools(tools, policy="policy.yaml")
agent = create_react_agent(model, governed)

from aegis.adapters.langchain import LangChainExecutor, AegisTool

# Wrap existing LangChain tools with governance
executor = LangChainExecutor(tools=[DuckDuckGoSearchRun()])
runtime = Runtime(executor=executor, policy=Policy.from_yaml("policy.yaml"))

# Or expose governed actions AS LangChain tools
tool = AegisTool.from_runtime(runtime, name="governed_search",
    description="Policy-governed search", action_type="search", action_target="web")

from aegis.adapters.openai_agents import governed_tool

@governed_tool(runtime=runtime, action_type="write", action_target="crm")
async def update_contact(name: str, email: str) -> str:
    """Update a CRM contact -- governed by Aegis policy."""
    return await crm.update(name=name, email=email)

from aegis.adapters.crewai import AegisCrewAITool

tool = AegisCrewAITool(runtime=runtime, name="governed_search",
    description="Search with governance", action_type="search",
    action_target="web", fn=lambda query: do_search(query))

from aegis.adapters.anthropic import govern_tool_call

for block in response.content:
    if block.type == "tool_use":
        result = await govern_tool_call(
            runtime=runtime, tool_name=block.name,
            tool_input=block.input, target="my_system")

from aegis.adapters.httpx_adapter import HttpxExecutor

executor = HttpxExecutor(base_url="https://api.example.com",
    default_headers={"Authorization": "Bearer ..."})
runtime = Runtime(executor=executor, policy=Policy.from_yaml("policy.yaml"))

# Action types map to HTTP methods: get, post, put, patch, delete
plan = runtime.plan([Action("get", "/users"), Action("delete", "/users/1")])

from aegis.adapters.mcp import govern_mcp_tool_call, AegisMCPToolFilter

# Option 1: Govern individual tool calls
result = await govern_mcp_tool_call(
    runtime=runtime, tool_name="read_file",
    arguments={"path": "/data.csv"}, server_name="filesystem")

# Option 2: Filter-based governance
tool_filter = AegisMCPToolFilter(runtime=runtime)
result = await tool_filter.check(server="filesystem", tool="delete_file")
if result.ok:
    # Proceed with actual MCP call
    pass

pip install 'agent-aegis[server]'
aegis serve policy.yaml --port 8000

# Evaluate an action (dry-run)
curl -X POST http://localhost:8000/api/v1/evaluate \
    -H "Content-Type: application/json" \
    -d '{"action_type": "delete", "target": "db"}'
# => {"risk_level": "CRITICAL", "approval": "block", "is_allowed": false}

# Execute through full governance pipeline
curl -X POST http://localhost:8000/api/v1/execute \
    -H "Content-Type: application/json" \
    -d '{"action_type": "read", "target": "crm"}'

# Query audit log
curl http://localhost:8000/api/v1/audit?action_type=delete

# Hot-reload policy
curl -X PUT http://localhost:8000/api/v1/policy \
    -H "Content-Type: application/json" \
    -d '{"yaml": "rules:\n  - name: block_all\n    match: {type: \"*\"}\n    approval: block"}'

pip install 'agent-aegis[mcp]'
aegis-mcp-server --policy policy.yaml

{ "mcpServers": { "aegis": { "command": "uvx", "args": ["--from", "agent-aegis[mcp]", "aegis-mcp-server"] }}}

{ "mcpServers": { "aegis": { "command": "uvx", "args": ["--from", "agent-aegis[mcp]", "aegis-mcp-server"] }}}

{ "servers": { "aegis": { "command": "uvx", "args": ["--from", "agent-aegis[mcp]", "aegis-mcp-server"] }}}

{ "mcpServers": { "aegis": { "command": "uvx", "args": ["--from", "agent-aegis[mcp]", "aegis-mcp-server"] }}}

from aegis.adapters.base import BaseExecutor
from aegis.core.action import Action
from aegis.core.result import Result, ResultStatus

class MyAPIExecutor(BaseExecutor):
    async def execute(self, action: Action) -> Result:
        response = await my_api.call(action.type, action.target, **action.params)
        return Result(action=action, status=ResultStatus.SUCCESS, data=response)

    async def verify(self, action: Action, result: Result) -> bool:
        return result.data.get("status") == "ok"

Policy Conditions

Go beyond glob matching with smart conditions:

rules:
  # Block writes after business hours
  - name: after_hours_block
    match: { type: "write*" }
    conditions:
      time_after: "18:00"
    risk_level: critical
    approval: block

  # Escalate bulk operations over threshold
  - name: large_bulk_ops
    match: { type: "update*" }
    conditions:
      param_gt: { count: 100 }
    risk_level: high
    approval: approve

  # Only allow deploys on weekdays
  - name: weekday_deploys
    match: { type: "deploy*" }
    conditions:
      weekdays: [1, 2, 3, 4, 5]
    risk_level: medium
    approval: approve

Available: time_after, time_before, weekdays, param_eq, param_gt, param_lt, param_gte, param_lte, param_contains, param_matches (regex).

Semantic Conditions

Go beyond keyword matching with the two-tier semantic conditions engine:

rules:
  - name: block_harmful_content
    match: { type: "generate*" }
    conditions:
      semantic: "contains harmful, violent, or illegal content"
    risk_level: critical
    approval: block

Tier 1 uses fast built-in keyword matching. Tier 2 plugs in any LLM evaluator via the SemanticEvaluator protocol -- bring your own model for nuanced content analysis.

Deep Features

Advanced capabilities for production-grade agent governance.

Behavioral Anomaly Detection

Aegis learns per-agent behavior profiles and automatically detects anomalies -- no manual threshold tuning required.

from aegis.core.anomaly import AnomalyDetector

detector = AnomalyDetector()

# Feed observed actions to build per-agent behavior profiles
detector.observe(agent_id="agent-1", action_type="read", target="crm")
detector.observe(agent_id="agent-1", action_type="read", target="crm")
detector.observe(agent_id="agent-1", action_type="read", target="crm")

# Detect anomalies: rate spikes, bursts, new actions, unusual targets, high block rates
alerts = detector.check(agent_id="agent-1", action_type="delete", target="prod_db")
# => [Anomaly(type=NEW_ACTION, detail="action 'delete' never seen for agent-1")]

# Auto-generate a policy from observed behavior
learned_policy = detector.generate_policy(agent_id="agent-1")

Detects: rate spikes | burst patterns | never-seen actions | unusual targets | high block rates

Compliance Report Generator

Generate audit-ready compliance reports from your existing audit logs. No additional tooling needed.

aegis compliance --type soc2 --output report.json
aegis compliance --type gdpr --output gdpr-report.json
aegis compliance --type governance --days 30

from aegis.core.compliance import ComplianceReporter

reporter = ComplianceReporter(audit_store=runtime.audit_store)
report = await reporter.generate(report_type="soc2", days=90)

print(report.score)        # 87.5
print(report.findings)     # List of findings with severity
print(report.evidence)     # Linked audit log entries

Supported report types: SOC2 | GDPR | Governance -- each with scoring, findings, and evidence links.

Policy Diff & Impact Analysis

Compare two policy files and understand exactly what changed and what impact it will have.

# Show added/removed/modified rules between two policies
aegis diff policy-v1.yaml policy-v2.yaml

# Replay historical actions against the new policy to see impact
aegis diff policy-v1.yaml policy-v2.yaml --replay audit.db

 Rules: 2 added, 1 removed, 3 modified

 + bulk_write_block     CRITICAL/block   (new)
 + pii_access_approve   HIGH/approve     (new)
 - legacy_allow_all     LOW/auto         (removed)
 ~ read_safe            LOW/auto → LOW/auto  conditions changed
 ~ deploy_prod          HIGH/approve → CRITICAL/block  risk escalated
 ~ bulk_ops             MEDIUM/approve   param_gt.count: 100 → 50

 Impact (replayed 1,247 actions):
   23 actions would change from AUTO → BLOCK
    7 actions would change from APPROVE → BLOCK

Agent Trust Chain

Hierarchical agent identity with delegation and capability-scoped trust.

from aegis.core.trust import TrustChain, AgentIdentity, Capability

# Create a root agent with full capabilities
root = AgentIdentity(
    agent_id="orchestrator",
    capabilities=[Capability("*")],  # glob matching
)

# Delegate a subset of capabilities (intersection semantics)
worker = root.delegate(
    agent_id="data-worker",
    capabilities=[Capability("read:*"), Capability("write:staging_*")],
)

# Worker can only do what both root AND delegation allow
chain = TrustChain()
chain.register(root)
chain.register(worker, parent=root)

# Verify capability at runtime
chain.can(worker, "read:crm")           # True
chain.can(worker, "delete:prod_db")     # False -- not in delegation

# Cascade revocation: revoking parent revokes all children
chain.revoke(root)
chain.can(worker, "read:crm")           # False

aegis scan -- Static Analysis

AST-based scanner that detects ungoverned AI tool calls in your Python codebase.

aegis scan ./src/

# Output:
# src/agents/mailer.py:42  openai.ChatCompletion.create()  -- ungoverned
# src/agents/writer.py:18  anthropic.messages.create()     -- ungoverned
# src/tools/search.py:7    langchain tool "web_search"     -- ungoverned
#
# 3 ungoverned calls found. Run `aegis score` for governance coverage.

aegis score -- Governance Score

Quantify your governance coverage with a 0-100 score and generate a shields.io badge.

aegis score ./src/ --policy policy.yaml

# Governance Score: 84/100
#   Governed calls:   21/25 (84%)
#   Policy coverage:  18 rules covering 6 action types
#   Anomaly detection: enabled
#   Audit trail:       enabled
#
# Badge: https://img.shields.io/badge/aegis_score-84-brightgreen

Add the badge to your repo:

![Aegis Score](https://img.shields.io/badge/aegis_score-84-brightgreen)

Architecture

aegis/
  core/        Action, Policy engine, Conditions, Risk levels, Retry, JSON Schema
  core/anomaly     Behavioral anomaly detection -- per-agent profiling, auto-policy generation
  core/compliance  Compliance report generator -- SOC2, GDPR, governance scoring
  core/trust       Agent trust chain -- hierarchical identity, delegation, revocation
  core/semantic    Semantic conditions engine -- keyword matching + LLM evaluator protocol
  core/diff        Policy diff & impact analysis -- rule comparison, action replay
  adapters/    BaseExecutor, Playwright, httpx, LangChain, CrewAI, OpenAI, Anthropic, MCP
  runtime/     Runtime engine, ApprovalHandler, AuditLogger (SQLite/JSONL/webhook/logging)
  server/      REST API (Starlette ASGI) -- evaluate, execute, audit, policy endpoints
  cli/         aegis validate | audit | schema | init | simulate | serve | stats | scan | score | diff | compliance

Why Aegis?

There are many ways to add governance to AI agents. Here's how they compare:

vs. Writing Your Own

vs. Platform-Native Guardrails

OpenAI, Google, and Anthropic each ship built-in guardrails — but they only govern their own ecosystem. If your agent calls OpenAI and Anthropic, or uses LangChain and MCP tools, you need one governance layer that works across all of them. That's Aegis.

vs. Enterprise Governance Platforms

Enterprise platforms like centralized control planes need Kubernetes clusters, cloud infrastructure, and procurement cycles. Aegis is a library — pip install and you have governance in 5 minutes. Start with a library, graduate to a platform when you need to.

CLI

aegis init                              # Generate starter policy
aegis validate policy.yaml              # Validate policy syntax
aegis schema                            # Print JSON Schema (for editor autocomplete)
aegis simulate policy.yaml read:crm delete:db  # Test policies without executing
aegis audit                             # View audit log
aegis audit --session abc --format json # Filter + format
aegis audit --tail                      # Live monitoring
aegis audit --format jsonl -o export.jsonl  # Export
aegis stats                             # Policy rule statistics
aegis serve policy.yaml --port 8000     # Start REST API server
aegis scan ./src/                       # Detect ungoverned AI tool calls (AST-based)
aegis score ./src/ --policy policy.yaml # Governance score (0-100) + badge
aegis diff policy-v1.yaml policy-v2.yaml           # Compare policies
aegis diff policy-v1.yaml policy-v2.yaml --replay  # Impact analysis with action replay
aegis compliance --type soc2 --output report.json  # Generate compliance report

Roadmap

Contributing

We welcome contributions! Check out:

• Good First Issues -- great starting points
• Contributing Guide -- setup, code style, PR process
• Architecture -- how the codebase is structured

git clone https://github.com/Acacian/aegis.git && cd aegis
make dev      # Install deps + hooks
make test     # Run tests
make lint     # Lint + format check
make coverage # Coverage report

Or jump straight into a cloud environment:

Badge

Using Aegis? Add a badge to your project:

[![Governed by Aegis](https://img.shields.io/badge/governed%20by-aegis-blue?logo=data:image/svg%2bxml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHZpZXdCb3g9IjAgMCAxMDAgMTAwIj48dGV4dCB5PSIuOWVtIiBmb250LXNpemU9IjkwIj7wn5uh77iPPC90ZXh0Pjwvc3ZnPg==)](https://github.com/Acacian/aegis)

License

MIT -- see LICENSE for details.

Copyright (c) 2026 구동하 (Dongha Koo, @Acacian). Created March 21, 2026.

_{Built for the era of autonomous AI agents.
If Aegis helps you, consider giving it a star -- it helps others find it too.}