What is Aegis
Every AI agent framework reinvents the same governance primitives — and each one does it slightly differently. Aegis is the abstraction layer that unifies them.
import aegis
aegis.auto_instrument() # 12 frameworks governed. No other code changes.
You don't write a LangChain guardrail and a CrewAI guardrail and an OpenAI guardrail — you write one Policy and every framework inherits it.
How this differs from the guardrail libraries
They solve a different problem, and mostly a text-shaped one. Guardrails AI validates model output against a hub of validators; NeMo Guardrails scripts conversational and tool policy in the Colang DSL; Snyk Agent Scan — Invariant Labs' mcp-scan, since the Snyk acquisition — scans MCP servers and agent skills for known risk patterns and can proxy them at runtime; LLM Guard chained input/output scanners until it was archived in July 2026. Each one you wire in yourself, at a call site you choose. Aegis starts from the other end: auto_instrument() finds the frameworks already installed and instruments them in place, so one Policy covers all of them without a line of agent code changing. What it enforces is agent-shaped rather than prompt-shaped — delegation chains under a monotone trust constraint, audits of what an agent excluded rather than what it picked, the distance between an agent's declared intent and its measured impact, and a tamper-evident audit chain. All of it is deterministic, so there is no second model sitting in the request path. These are not exclusive choices: a semantic or model-based detector drops into GuardrailEngine.add() alongside the built-ins.
Primitives
The contract every adapter maps into. Framework-agnostic by design.
Every governance feature in Aegis — anomaly detection, cost budgets, drift, cascade guards, kill switches — is a composition of these primitives. Read the Concepts guide to see how they fit together.
Frameworks
One API. 12 agent frameworks + 3 protocol-level adapters.
auto_instrument() detects what's installed and patches only those — no hard dependencies. Custom adapters use the same BaseAdapter interface. Every adapter above is exercised against the current upstream release daily by the integration workflow, which drives each framework's real entrypoint and asserts a guardrail fires — the unit suite fakes these frameworks, so it cannot see upstream drift on its own.
Integration policy. Native extension points win wherever one exists that can actually block — Google ADK's BasePlugin and CrewAI's BeforeToolCallHook both do, and there auto_instrument() patches only enough to install the native object. Pydantic AI and the OpenAI Agents SDK ship native implementations you opt into (AbstractCapability, tool_input_guardrail) alongside a patch-based path for the zero-code case. Everything else is patched because no blocking hook exists: the raw OpenAI/Anthropic SDKs and DSPy expose none, and LlamaIndex's instrumentation dispatcher emits events but swallows handler exceptions, so it can observe and not enforce.
The Pydantic AI integration is the reason this policy is written down. It was monkey-patched until core maintainer DouweM reviewed it — "It doesn't look like those features are actually exposed as Pydantic AI capabilities?" — and was rebuilt on the native extension API in response (src/aegis/contrib/pydantic_ai.py; the review is pydantic-ai#4888).
Patching is the fallback, not the preference — it is the part most exposed to upstream change, which is why the integration workflow exists.
Default Guardrails
Deterministic regex — no LLM calls, no network. 2.65ms cold / <1µs warm per check.
Use Cases
The same primitives, five different entry points. Pick whichever matches your workflow.
1. Runtime protection (most common)
One line. Any framework.
import aegis
aegis.auto_instrument()
Or zero code changes — AEGIS_INSTRUMENT=1 python my_agent.py. Injection blocking, PII masking, prompt-leak warnings, audit trail, and policy enforcement become active for every LangChain / CrewAI / OpenAI / Anthropic / LiteLLM / ADK / DSPy / LlamaIndex / Pydantic AI call.
Pydantic AI native capability — no monkey-patching, explicit per-agent control:
from pydantic_ai import Agent
from aegis.contrib.pydantic_ai import AegisCapability
agent = Agent(
"openai:gpt-4o-mini",
capabilities=[AegisCapability.default()], # injection, PII, toxicity, prompt-leak, hallucination
)
result = await agent.run("What is AI governance?")
Full Pydantic AI integration guide →
2. Pre-production scanning
Find ungoverned AI calls before they ship.
pip install agent-aegis
aegis scan .
Aegis Governance Scan
=====================
Scanned: 47 files in ./src
Found 5 ungoverned tool call(s):
agent.py:12 OpenAI function call with tools= — no governance wrapper [ASI02]
tools.py:8 LangChain @tool "search_db" — no policy check [ASI02]
llm.py:21 LiteLLM litellm.completion() — no governance wrapper [ASI02]
run.py:5 subprocess subprocess.run — direct shell execution [ASI08]
api.py:14 HTTP requests.post — raw HTTP in agent code [ASI07]
Governance Score: D (5 ungoverned call(s))
Supports --format json|sarif|suggest, --threshold A-F, .aegisscanignore, and inline # aegis: ignore pragmas. Auto-fix with aegis scan --fix.
3. Policy CI/CD
Security tools protect at runtime. Aegis also manages the policy lifecycle — the same way you test and ship code.
aegis plan current.yaml proposed.yaml --audit-db aegis_audit.db
# Policy Impact Analysis
# Rules: 2 added, 1 removed, 3 modified
# Impact (replayed 1,247 actions):
# 23 actions would change from AUTO → BLOCK
aegis test policy.yaml tests.yaml # Run in CI
aegis test policy.yaml --generate # Auto-generate test suite
aegis test new.yaml tests.yaml --regression old.yaml # Regression check
# .github/workflows/policy-check.yml
- uses: Acacian/aegis@main
with:
policy: aegis.yaml
tests: tests.yaml
fail-on-regression: true
Or block ungoverned calls at PR time:
- uses: Acacian/[email protected]
with:
command: scan
fail-on-ungoverned: true
4. Audit & compliance
Every call is logged to a tamper-evident Merkle chain, with mappings to EU AI Act / NIST AI RMF / SOC2 built in.
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
SQLite + JSONL + webhook sinks. Ed25519 signing for long-term evidence. See the Compliance guide.
5. Governance server (multi-agent)
Centralized governance for multiple agents. Each agent connects via SDK, server handles policy, guardrails, audit, and compliance.
pip install 'agent-aegis[server]'
aegis-server
37 REST endpoints + WebSocket audit streaming + web dashboard. Agents auto-register, send heartbeats, and query policy over HTTP. See Governance Framework Server.
30-Second Start
pip install agent-aegis
import aegis
aegis.auto_instrument()
# All 12 frameworks now governed with default guardrails.
Or use a YAML policy for full control:
aegis init # Creates aegis.yaml
# aegis.yaml
guardrails:
pii: { enabled: true, action: mask }
injection: { enabled: true, action: block, sensitivity: medium }
policy:
version: "1"
defaults:
risk_level: medium
approval: approve
rules:
- name: read_safe
match: { type: "read*" }
risk_level: low
approval: auto
- name: no_deletes
match: { type: "delete*" }
risk_level: critical
approval: block
Install Options
pip install agent-aegis # Core (includes auto_instrument for all frameworks)
pip install langchain-aegis # LangChain standalone integration
pip install 'agent-aegis[mcp]' # MCP server + proxy
pip install 'agent-aegis[server]' # REST API + dashboard
pip install 'agent-aegis[all]' # Everything
MCP Proxy — govern any MCP server with zero code changes
{
"mcpServers": {
"filesystem": {
"command": "uvx",
"args": ["--from", "agent-aegis[mcp]", "aegis-mcp-proxy",
"--wrap", "npx", "-y",
"@modelcontextprotocol/server-filesystem", "/home"]
}
}
}
Works with Claude Desktop, Cursor, VS Code, Windsurf. STDIO injection protection, tool poisoning detection, rug-pull detection, argument sanitization, policy evaluation, full audit trail.
Governance Framework Server
Run Aegis as a dedicated governance server with REST API, WebSocket streaming, and web dashboard.
pip install 'agent-aegis[server]'
aegis-server --init # Generate aegis-server.yaml
aegis-server # Start server on :8000
37 REST endpoints covering the full governance lifecycle:
Connect with the Python SDK (sync or async):
from aegis import AegisClient
with AegisClient("http://localhost:8000", agent_id="my-agent") as client:
result = client.evaluate("delete", "user_data")
# result["risk_level"] == "CRITICAL", result["is_allowed"] == False
from aegis import AsyncAegisClient
async with AsyncAegisClient("http://localhost:8000", agent_id="my-agent") as client:
result = await client.evaluate("read", "reports")
Config-driven via aegis-server.yaml — guardrails, webhooks (Slack/PagerDuty), rate limiting, cost budgets, and auth all declarative. See aegis-server.example.yaml.
Why Aegis?
What Only Aegis Does
Other tools check inputs and outputs. Aegis governs the decision itself — with primitives no other governance runtime exposes.
CLI
aegis scan ./src/ # Detect ungoverned AI calls
aegis score ./src/ --policy policy.yaml # Governance score (0-100)
aegis init # Generate starter policy
aegis validate policy.yaml # Validate syntax
aegis plan current.yaml proposed.yaml # Preview policy changes
aegis test policy.yaml tests.yaml # Policy regression testing
aegis check policy policy.yaml read:crm # Policy decision per action (CI gate)
aegis audit # View audit log
aegis serve policy.yaml # REST API + dashboard
aegis probe policy.yaml # Adversarial policy testing
aegis autopolicy "block deletes" # Natural language → YAML
Research
Original measurements on public agent trace datasets. Stdlib-only, reproducible in 30 seconds.
- The Justification Gap in 14,285 Tau-Bench Tool Calls — Formal definition of the Tripartite ActionClaim with a silent-baseline empirical study. 90.3% approve / 9.7% escalate / 0% block across four model:domain groups. Airline domain exposes ~2× the mean gap of retail. Includes soundness sketches for the three structural invariants and an honest note on the
max-only override limitation discovered during the study. - Tool Distribution Drift in 1,960 Tau-Bench Trajectories — Shannon entropy on tool name sequences across GPT-4o and Sonnet 3.5 New. 39.8% of scored trajectories collapse onto one or two tools by the end. Bimodal distribution, 1.7× cross-model gap. All scripts and raw data included.
Run the same signal on your own trace:
aegis check drift --trace path/to/trace.jsonl
The CLI reads only the tool_name field — never args, CoT, or prompts — so enterprise users can score prod traces without exfiltrating PII.
We also ran aegis scan across 39 public agent repositories and graded their governance posture: 92% scored an F. That is a finding about the ecosystem, not about any one project — most agent code has no tool-call policy at all.
Documentation
Full documentation at acacian.github.io/aegis:
- Integration guides — LangChain, CrewAI, OpenAI, MCP, and more
- Policy reference — conditions, templates, best practices
- Security features — guardrails, anomaly detection, compliance
- API stability — what 1.x guarantees, and what counts as a breaking change
- Architecture — how the codebase is structured
- Interactive playground — try in browser, no install
Contributing
git clone https://github.com/Acacian/aegis.git && cd aegis
make dev # Install deps + hooks
make test # Run tests
make lint # Lint + format check
Contributing Guide • Good First Issues •
License
MIT -- see LICENSE for details.
Copyright (c) 2026 구동하 (Dongha Koo, @Acacian). Created March 21, 2026.