aionai

A shared working-state layer for your AI coding tools. Claude, Cursor, and other MCP clients read and write one small notebook, so they stay on the same page — and you stop being the copy-paste bus between them.

One-click install for Cursor (activates once aionai is published to PyPI). Remember to set AIONAI_SOURCE=cursor in the server's env afterwards.

---

The problem

You use more than one AI assistant on a project. They don't know what each other did: you tell Claude a decision, switch to Cursor, and Cursor has no idea. You end up re-explaining and copy-pasting context by hand.

What aionai does

aionai is a tiny local MCP server that keeps a shared, persistent notebook every tool can read and write. It does three jobs:

• Remember — decisions, changes, questions, and notes, organized by project.
• Track — a lightweight roadmap/to-do list with progress rollup.
• Hand off — post a task to another tool's inbox (with an optional "doorbell").

It does not replace your repo, docs, or git — those stay the source of truth. aionai just holds the live working state and hands each tool the slice it needs.

Install

pipx install aionai        # or:  uv tool install aionai

No install step at all, if you use uv:

uvx aionai --help

Connect your tools

Point each client at the aionai command. Give each client a distinct AIONAI_SOURCE (so handoffs route correctly) and set AIONAI_SEGMENT to your project name.

Claude Code

claude mcp add aionai --env AIONAI_SOURCE=claude-code --env AIONAI_SEGMENT=myproject -- aionai
# or with uv:  claude mcp add aionai --env AIONAI_SOURCE=claude-code -- uvx aionai

Cursor — .cursor/mcp.json:

{
  "mcpServers": {
    "aionai": { "command": "aionai", "env": { "AIONAI_SOURCE": "cursor", "AIONAI_SEGMENT": "myproject" } }
  }
}

Claude Desktop — Settings → Developer → Edit Config:

{
  "mcpServers": {
    "aionai": { "command": "aionai", "env": { "AIONAI_SOURCE": "claude-desktop", "AIONAI_SEGMENT": "myproject" } }
  }
}

Restart the client after editing its config — MCP servers are launched (and their env read) when the client connects.

Use it

Add this to each tool's rules (CLAUDE.md, Cursor rules) so they do it reflexively:

Before working, call context_pull(segment="myproject") and treat the result as the current truth. As you work, context_log(...) your decisions/changes/questions/tasks. When something is settled, context_resolve(id).

That's the whole loop: pull first, write back. Now open Cursor and it already knows what you and Claude decided — no paste.

Segments

State is organized by a dotted segment path whose first element is the project:

myproject                     # the whole project
myproject/backend             # a layer
myproject/backend/auth        # a workstream

Pulling a parent includes all descendants. That one mechanism keeps an always-on space from turning into an undifferentiated blob.

The tools

tool	purpose
context_pull(segment)	current working state + your inbox
context_log(segment, type, content, refs)	append a decision/change/question/task/note
context_resolve(id)	close a question or task
context_search(query, segment)	full-text recall over history
context_verify(segment)	check change entries against git (merged vs. bare claim)
context_handoff(segment, content, to)	post a handoff to another tool's inbox
roadmap_add_node / roadmap_update / roadmap_block	build & manage the roadmap
roadmap_view / roadmap_progress	see the tree / how far along
constraints_for_task(segment)	decisions + open questions that constrain a task
project_lookup(query, project)	reuse an approach from another project

They also surface as slash commands (/mcp__aionai__pull, …log, …resolve, …search, …handoff) in clients that support MCP prompts.

Optional extras

• Auto-ingest commits — copy hooks/post-commit into a repo's .git/hooks/ (set AIONAI_SEGMENT) and every commit logs itself as a change.
• Doorbell — set AIONAI_DELIVERY=1 in a sender's env and a context_handoff to Cursor also summons Cursor via its deeplink (you confirm before it runs). Without it, handoffs are inbox-only. Only cursor has a verified deeplink today.

How it works

One SQLite database, one append-only table. Every decision, task, handoff, and roadmap node is a row tagged with a segment and a type. History is free because nothing is overwritten. Full-text search uses FTS5 with a LIKE fallback. The MCP tools are thin wrappers over a plain-Python storage layer (src/aionai/store.py).

Configuration

var	default	purpose
AIONAI_DB	~/.aionai/aionai.db	database path (share explicitly if clients don't hit the default)
AIONAI_SOURCE	agent	who is writing (cursor / claude-code / claude-desktop); routes the inbox
AIONAI_SEGMENT	project	default project segment
AIONAI_DELIVERY	(unset)	1 enables the doorbell in the sender

Development

git clone https://github.com/imsm/aionai && cd aionai
pip install -e ".[dev]"
pytest
ruff check .

License

Apache-2.0 © Ismail Saleh.