aegis

<!-- mcp-name: io.github.Acacian/aegis --> <p align="center"> <h1 align="center">Agent-Aegis</h1> <p align="center"> <strong>The governance layer for AI agents. One API, 12 frameworks, every governance primitive.</strong> </p> <p align="center"> Aegis is to agent governance what Redis is to data structures — one runtime that unifies prompt-injection blocking, PII masking, policy enforcement, trust delegation, and tamper-evident audit across every agent framework. No code changes.<br/> <code>pip install agent-aegis</code> → <code>aegis.auto_instrument()</code> → 12 frameworks are now governed. </p> </p> <p align="center"> <a href="https://www.bestpractices.dev/projects/12253"><img src="https://www.bestpractices.dev/projects/12253/badge" alt="OpenSSF Best Practices"></a> <a href="https://github.com/Acacian/aegis/actions/workflows/ci.yml"><img src="https://github.com/Acacian/aegis/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://pypi.org/project/agent-aegis/"><img src="https://img.shields.io/pypi/v/agent-aegis?color=blue&cacheSeconds=3600" alt="PyPI"></a> <a href="https://pypi.org/project/agent-aegis/"><img src="https://img.shields.io/pypi/pyversions/agent-aegis?cacheSeconds=3600" alt="Python"></a> <a href="https://github.com/Acacian/aegis/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="License"></a> <br/> <a href="https://github.com/Acacian/aegis/actions/workflows/ci.yml"><img src="https://img.shields.io/badge/tests-6400%2B_passed-brightgreen" alt="Tests"></a> <a href="https://github.com/Acacian/aegis/actions/workflows/ci.yml"><img src="https://img.shields.io/badge/coverage-92%25-brightgreen" alt="Coverage"></a> <a href="https://acacian.github.io/aegis/playground/"><img src="https://img.shields.io/badge/playground-Try_it_Live-ff6b6b" alt="Playground"></a> </p> <p align="center"> <a href="#what-is-aegis"><strong>What is Aegis</strong></a> &bull; <a href="#primitives">Primitives</a> &bull; <a href="#frameworks">Frameworks</a> &bull; <a href="#use-cases">Use Cases</a> &bull; <a href="#30-second-start"><strong>30-Second Start</strong></a> &bull; <a href="#research">Research</a> &bull; <a href="https://acacian.github.io/aegis/">Docs</a> &bull; <a href="https://acacian.github.io/aegis/playground/"><strong>Playground</strong></a> </p> <p align="center"> <b>English</b> &bull; <a href="./README.ko.md">한국어</a> </p>

---

<p align="center"> <img src="docs/assets/demo.gif?v=2" alt="Aegis Demo" width="880"> </p>

---

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.

| Layer | What it does | Examples | |-------|-------------|----------| | 1. Primitives | A universal contract for every tool call | Action, ActionClaim, Policy, Result, DelegationChain, AuditEvent | | 2. Adapters | Auto-instrument any framework through its own hooks | LangChain callbacks, CrewAI BeforeToolCallHook, OpenAI Agents tracing, Google ADK BasePlugin, MCP transport, DSPy modules, httpx middleware, Playwright context | | 3. Governance | Declarative primitives you compose into policy | Prompt injection / PII / leak / toxicity guardrails, RBAC, rate limit, cost budget, drift detection, anomaly scoring, trust delegation, justification gap, selection audit, Merkle audit chain | | 4. Lifecycle | One runtime, every stage of agent ops | Scan → Instrument → Policy CI/CD → Runtime → Proxy → Audit |

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.

| Primitive | Purpose | Module | |-----------|---------|--------| | Action | Unified representation of any tool / LLM / HTTP / MCP call across all frameworks | aegis.core.action | | ActionClaim | Tripartite structure — Declared (agent-authored) / Assessed (Aegis-computed) / Chain (delegation) | aegis.core.action_claim | | Policy | Declarative YAML rules: match → risk → approval (auto / approve / block) | aegis.core.policy | | ClaimPolicy | Policy layer that evaluates 6-dimensional impact vectors, not just tool names | aegis.core.claim_policy | | Guardrails | Deterministic regex checks for injection, PII, prompt leak, toxicity — 2.65ms cold / <1µs warm | aegis.guardrails | | DelegationChain | Multi-agent hand-off tracking with monotone trust constraint (non-increasing) | aegis.core.agent_identity | | AuditEvent | Tamper-evident append-only log, Merkle-chained, SQLite + JSONL + webhook sinks | aegis.core.merkle_audit | | SelectionAudit | Audits what an agent excludes, not just what it picks — detects cosmetic alignment | aegis.core.selection_audit | | JustificationGap | 6D asymmetric scoring: agents declare impact, Aegis independently assesses, gap triggers escalation | aegis.core.justification_gap | | CryptoAuditChain | Ed25519-signed chain for long-term compliance evidence | aegis.core.crypto_audit |

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.

| Framework | Hook | Integration | |-----------|------|-------------| | Google ADK | BasePlugin lifecycle (tool calls, agent routing, sessions) | Native — the patch only installs the plugin | | CrewAI | global BeforeToolCallHook, Crew.kickoff/kickoff_async | Hybrid — native hook for tool calls, patch for crew entry | | Pydantic AI | AbstractCapability · Agent.run/run_sync | Native (opt-in) · Patch (auto) | | OpenAI Agents SDK | tool_input_guardrail/tool_output_guardrail · Runner.run/run_sync | Native (opt-in) · Patch (auto) | | LangChain | BaseChatModel.invoke/ainvoke, BaseTool.invoke/ainvoke | Patch | | OpenAI API | Completions.create (chat & completions) | Patch | | Anthropic API | Messages.create | Patch | | LiteLLM | completion, acompletion | Patch | | Google GenAI | Models.generate_content (new + legacy) | Patch | | LlamaIndex | LLM.chat/achat/complete/acomplete on every concrete subclass, BaseQueryEngine.query/aquery | Patch | | Instructor | Instructor.create, AsyncInstructor.create | Patch | | DSPy | Module.__call__, LM.forward/aforward | Patch | | MCP | Transport-layer proxy for any MCP server (stdio / HTTP) | Proxy — no patching | | httpx | HttpxExecutor for raw HTTP egress (REST agents, webhooks) | Wrapper — no patching | | Playwright | PlaywrightExecutor for browsing agents | Wrapper — no patching |

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

| Guardrail | Default | What it catches | |-----------|---------|-----------------| | Prompt injection | Block | 13 attack categories, 109 patterns, 9 languages (EN/KO/ZH/JA/ES/DE/FR/TH/VI) | | PII detection | Warn | 13 categories (email, credit card, SSN, IBAN, API keys, etc.) | | Prompt leak | Warn | System prompt extraction attempts | | Toxicity | Warn | Harmful, violent, or abusive content | | MCP STDIO injection | Block | JSON-RPC injection, frame concatenation, unicode escape bypass (OX Security advisory) |

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:

| API Group | Endpoints | Purpose | |-----------|-----------|---------| | Core | evaluate, execute, audit, policy | Policy evaluation + execution pipeline | | Agents | register, heartbeat, list, status | Agent lifecycle management | | Guardrails | check, list | Content safety checks | | Policy Versioning | commit, diff, rollback, tag | Git-like policy change management | | Crypto Audit | verify, entries, evidence | Tamper-proof audit chain verification | | Trust & Drift | trust score, drift detection | Per-agent behavioral analysis | | Cost | budget check, reports | LLM cost governance | | Compliance | reports, regulatory gaps | SOC2 / GDPR / EU AI Act reports | | Sessions | list, replay | Session recording + forensic replay |

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?

| | Writing your own | Platform guardrails | Enterprise platforms | Aegis | |---|---|---|---|---| | Abstraction level | Per-framework if/else | Single-vendor SDK | Proprietary gateway | Universal primitives across 12 frameworks | | Setup | Days of if/else | Vendor-specific config | Kubernetes + procurement | pip install + one line | | Code changes | Wrap every call | SDK-specific | Months of integration | Zero — auto-instruments | | Policy portability | Rewrite per framework | Locked to ecosystem | Usually single-vendor | One YAML policy, every framework | | Governance primitives | Build from scratch | Subset, vendor-defined | Proprietary | 10+ composable primitives | | Policy CI/CD | None | None | None | aegis plan + aegis test | | Audit trail | printf debugging | Platform logs only | Cloud dashboard | SQLite + JSONL + webhooks + Merkle chain | | Compliance | Manual docs | None | Enterprise sales cycle | EU AI Act, NIST, SOC2 built-in | | Cost | Engineering time | Free-to-$$$ | $$$$ + infra | Free (MIT). Forever. |

What Only Aegis Does

Other tools check inputs and outputs. Aegis governs the decision itself — with primitives no other governance runtime exposes.

| Capability | What it means | Based on | |---|---|---| | Tripartite ActionClaim | Every tool call splits into Declared (agent-authored, untrusted), Assessed (Aegis-computed), and Chain (delegation) fields. The structural separation is what makes cosmetic alignment detectable. | Justification Gap measurement on 14,285 tau-bench calls | | Justification Gap | 6-dimensional asymmetric scoring: agents declare impact, Aegis independently assesses it, and per_dim = max(0, assessed − declared). Under-reporting triggers escalate (>0.15) or block (>0.40). | Name "ActionClaim" from COA-MAS (Carvalho); 6D metric + runtime form original | | Selection Governance | Audits what agents exclude, not just what they choose. A model that "helpfully" omits risky options is exerting selection power — Aegis detects this. | Santander et al., arXiv:2602.14606 | | Monotone Trust Constraint | Delegated agents cannot escalate their own authority. Trust levels must be non-increasing along the chain — violations auto-block. | Lattice-based access control | | Full Lifecycle | Scan (detect) → Instrument (protect) → Policy CI/CD (test) → Runtime (govern) → Proxy (gateway) → Audit (trace). One library, one pip install. | — |

---

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.

---

<p align="center"> _{The governance layer for AI agents. One API, 12 frameworks, every governance primitive.}<br/> _{If Aegis helps you, consider giving it a star -- it helps others find it too.} </p>