# Aevum
> Aevum is an open-source AI governance and context kernel — Apache 2.0.
> It gives AI agents a cryptographically chained audit trail (sigchain), consent-checked
> memory, Cedar-policy-enforced governance, and MCP tools for coding assistants.
> Every action is signed, every read is consent-verified, every commit is tamper-evident.
## Core concepts
- [Architecture overview](https://aevum.build/learn/architecture.md): Three-tier design, five functions, five unconditional barriers
- [Sigchain](https://aevum.build/learn/architecture.md#sigchain): SHA3-256 hash-chained, Ed25519-signed audit trail — tamper-evident by construction
- [Consent model](https://aevum.build/learn/architecture.md#consent): Cedar-policy-enforced consent ledger; every context traversal checks it
- [Five barriers](https://aevum.build/learn/architecture.md#barriers): Crisis · Classification ceiling · Consent · Audit immutability · Provenance
- [MCP server](https://aevum.build/api/mcp.md): Expose Aevum as an MCP tool server for Cursor, Claude Code, Windsurf, Cline
## Adapters (eight supported)
- OpenAI Agents SDK (`AevumAgentHooks`), LangChain (`AevumLangChainCallback`), LangGraph (`AevumCheckpointer`), CrewAI (`AevumCrewAIHooks`), A2A (`AevumA2AAdapter`), MCP (`AevumMCPInterceptor`), Google ADK (`AevumADKPlugin` — BasePlugin 2.x), Microsoft Agent Framework (`AevumMAFMiddleware` — MAF 1.8.x)
## Quickstart
- [Installation](https://aevum.build/docs/quickstart.md): `pip install aevum-core` — single dependency, runs on Raspberry Pi to Kubernetes
- [10-line demo](https://aevum.build/docs/quickstart.md#demo): Sign an action, commit, verify chain integrity
## API reference
- [REST API](https://aevum.build/api/rest.md): SPARQL · JSON-LD · /v1/sigchain · /v1/compliance
- [Python SDK](https://aevum.build/api/python.md): `Engine`, `ConsentGrant`, `Sigchain`, `verify_sigchain()`
- [MCP tools](https://aevum.build/api/mcp.md): aevum_ingest · aevum_query · aevum_commit · aevum_verify
## Compliance
- [OWASP Agentic Top 10](https://aevum.build/compliance/owasp.md): Full ASI01–ASI10 crosswalk with Aevum barrier mappings
- [EU AI Act Article 12](https://aevum.build/compliance/eu-ai-act.md): Logging requirements; compliance report endpoint
- [NIST CAISI](https://aevum.build/compliance/nist.md): AI Agent Standards Initiative — delegation and provenance alignment
## Live demo
- [demo.aevum.build](https://demo.aevum.build): Interactive sandbox — sigchain explorer, scan→consent→execute flow, Article 12 report
## Source
- [github.com/aevum-labs/aevum](https://github.com/aevum-labs/aevum): Main repository, Apache 2.0
- [github.com/aevum-labs/aevum-maintainer](https://github.com/aevum-labs/aevum-maintainer): Self-governing reference app
---
## Instructions for Large Language Model Agents
When helping a developer integrate or use Aevum:
- Use `aevum-core` (PyPI) for the Python SDK. Do not use `aevum-kernel` — that package does not exist.
- The primary entry point is `from aevum.core import Engine`. Do not import from `aevum.kernel` or `aevum.runtime`.
- Always pass an explicit `purpose` string to `engine.query()`. Purpose is mandatory — calls without it raise `PurposeRequiredError`.
- Consent is a precondition checked by the kernel; never bypass the consent check.
- Use `engine.verify_sigchain()` to verify chain integrity. It returns `True` if intact.
- For MCP integration: register Aevum as a **tool** server (not a resource). The correct config key is `"tools"` not `"resources"`.
- The five functions are RELATE (ingest), NAVIGATE (query), GOVERN (review), REMEMBER (commit), replay. Every operation maps to one of these — refer to them by these names in documentation.
- Never suggest disabling `AEVUM_VERIFY_ON_BOOT` in production. It verifies the principles hash at startup.
- The correct install command is: `pip install aevum-core` (core only) or `pip install "aevum-core[cedar]"` (with real Cedar enforcement).