AIDEMD
Code is ephemeral. Intent is law.
Ship code conversationally. A team of specialized agents translates your intent into production quality software through the AIDE cascading intent pipeline.
Paste this into your project's .mcp.json
{
"mcpServers": {
"aide": {
"command": "npx",
"args": [
"@aidemd-mcp/server@latest"
]
}
}
}- 1
Paste this into your project's `.mcp.json`.
- 2
Restart Claude Code so it loads the new MCP server.
- 3
Call `aide_init` from the tool palette.
aide_init
The cost of not knowing
A six-month fractional specialist engagement runs $30,000–$60,000 at the low end (Geisheker Group, 2026). AIDEMD captures the same domain knowledge for free via npx.
How it works
Intent is the contract. Code is ephemeral.
brain (Obsidian vault) — shared infrastructure
Persistent Obsidian vault shared across every agent and every session. Not a pipeline stage — a knowledge layer the pipeline reads and writes.
- domain expert writes findings to brain
- strategist reads domain research from brain
- architect reads coding playbook and brain
- auditor reads coding playbook and brain
Mechanisms
- Fix loop: QA produces a todo, implementor fixes one item per session in clean context, QA re-validates.
- Resume mid-pipeline: Orchestrator detects state from file existence and picks up where the last session left off.
Not a rules pack
CLAUDE.md, AGENTS.md, .cursorrules — and why AIDE is mechanically different.
Rules packs
- —Flat files appended to every context window
- —Static instructions — drift silently as code evolves
- —No structure: conventions, warnings, and philosophy in one lump
- —Agent cannot verify it followed the rules
- —Examples: CLAUDE.md, AGENTS.md, .cursorrules
AIDE
- +Intent specs cascade from research through the full eight-agent pipeline
- +Cross-project brain persists domain knowledge in an Obsidian vault filed by domain — not trapped in a chat window
- +Auto-learning: when the fix loop closes, retro findings promote to the brain — rules packs cannot learn from their own failures
- +QA checks every output against the intent spec — drift is caught at the spec boundary, not in review
- +Coding playbook: architect navigates a structured convention hierarchy via a dedicated skill — rules packs are flat unstructured text
The mechanical difference: rules packs are static lumps — they cannot update themselves, cannot verify compliance, and cannot learn. AIDE cascades research into outcome-checked intent specs, and when the fix loop closes, retro findings promote to the brain so every future session starts smarter.
Intent cascades through the project tree
Agents read from root to leaf — each spec narrows the parent's intent, so every module carries the full context of the project above it.
content-pipeline/├── .aide/│ └── intent.aide ← project-wide intent├── src/│ ├── .aide ← narrows to content domain│ └── service/│ ├── ingest/│ │ ├── .aide ← narrows to ingestion│ │ ├── index.ts│ │ ├── parseFeed/│ │ │ ├── .aide ← narrows to feed parsing│ │ │ └── index.ts│ │ └── normalize/│ │ └── index.ts (helper — no spec)│ ├── transform/│ │ ├── .aide ← narrows to transformation│ │ └── index.ts│ └── publish/│ ├── .aide ← narrows to publishing│ └── index.ts