# AI Layer

URL: https://mechcrate.dev/docs/ai/
Source: site/apps/site/src/content/docs/docs/ai/index.mdx

---

import AiLoop from '@/components/diagrams/AiLoop.astro';

Most "AI-native" tooling means a hosted assistant. Here it means something
narrower and more checkable: an agent gets **tools that operate your ecosystem**
and **retrieval over engineering documents you wrote**, through one server you
run yourself.

<AiLoop />

Three moving parts, and they only connect in one direction: the research
pipeline opens a pull request, a human merges it, `mx rag ingest` picks up the
delta, and agents retrieve it. Nothing writes to the corpus without a merge.

| Page | What it covers |
|---|---|
| [MCP server](/docs/ai/mcp-server/) | The 47 tools, grouped into four families; building the server and wiring a client |
| [Techniques corpus &amp; RAG](/docs/ai/rag-setup/) | `rag.toml`, Neon or a local pgvector container, embeddings, and where your text actually goes |
| [Research pipeline](/docs/ai/research-pipeline/) | How the corpus grows: weekly run, primary-source corroboration, PR gate, human merge |
| [Agent execution rules](/docs/ai/agent-rules/) | The rules agents in this repository work under |

The documents themselves are published under
[Techniques Corpus](/docs/corpus/). It is the same markdown that gets chunked,
embedded and served to agents, rendered for humans.

## What it is not

- **Not a hosted service.** There is no MechCrate API. The MCP server is a local
  binary; the corpus is a Postgres connection string you own.
- **Not a model.** mx never runs inference. Your agent supplies the model; mx
  supplies tools and retrieved text.
- **Not magic retrieval.** The corpus is 66 documents that a human merged. It
  answers well inside that surface and returns weak hits outside it, which is
  what `mx rag gaps` exists to surface.