maestro-nerve

maestro-nerve

Shared, evidence-grounded memory for AI agents.

One workspace memory plane for Claude Code, Codex, ChatGPT, Claude Desktop, and custom MCP clients.

maestro-nerve is the public client distribution for Maestro Nerve. It ships:

• mnerve, a production CLI for login, workspace inspection, and MCP config generation.
• Claude Code and Codex plugin manifests.
• A portable agent skill that explains the hosted MCP tool contract.
• A thin HTTP client against https://nerve-api.maestro.onl.

The backend is hosted by Maestro and remains closed source. This repository does not contain retrieval, ranking, extraction, normalization, storage schema, or reconciliation logic.

Quick Start

Use uv or pipx for the CLI. They install Python command-line applications in isolated tool environments and avoid modifying a project virtualenv.

uv tool install maestro-nerve
mnerve login
mnerve access --client claude-code --format mcp-json

Other good install paths:

# Run once without a persistent install.
uvx maestro-nerve login

# Persistent CLI install with pipx.
pipx install maestro-nerve

# Fallback when your environment already manages Python packages.
python -m pip install --user maestro-nerve

mnerve login opens the hosted Maestro login page and completes through a local loopback callback on 127.0.0.1. It does not use localhost:3000 in production. For SSH or browser-restricted sessions:

mnerve login --no-browser
mnerve login --manual

Development builds can explicitly target a local web app:

mnerve login --app-url http://localhost:3000

Agent Setup

Install the CLI once, then ask the backend for the exact MCP config shape your host expects:

mnerve access --client claude-code --format mcp-json
mnerve access --client codex --format mcp-json
mnerve access --client chatgpt --format envelope

Supported formats:

Format	Use
envelope	Script-friendly response with client, transport, workspace, endpoint, config, and note.
mcp-json	Pasteable MCP server JSON.
settings-json	JSON shaped for settings files that merge under mcpServers.
env	KEY=value lines for shell launchers and automation.

Human hints are written to stderr. Stdout remains parseable.

Claude Code Plugin

/plugin marketplace add maestro-ai-stack/maestro-nerve
/plugin install maestro-nerve@maestro-nerve
/reload-plugins

The plugin brings the skill and metadata. The hosted MCP connection still needs CLI auth:

uv tool install maestro-nerve
mnerve login
mnerve access --client claude-code --format mcp-json

Codex Plugin

codex plugin add maestro-ai-stack/maestro-nerve
uv tool install maestro-nerve
mnerve login
mnerve access --client codex --format mcp-json

If your agent/plugin tooling is installed with npm, pnpm, or Bun, keep using that tool for the host. mnerve itself is distributed as a Python CLI on PyPI, so uv or pipx is the cleaner install path for the command.

Skill Only

Use this when the host already has an MCP connection and only needs agent instructions:

npx skills add maestro-ai-stack/maestro-nerve -y -g

Equivalent package runners such as pnpm dlx or bunx are fine if your local skill manager supports them.

CLI Commands

mnerve login                              # browser login; stores local credentials
mnerve login --manual                     # paste an API key
mnerve logout                             # remove local credentials
mnerve whoami                             # print backend identity + primary workspace
mnerve workspace list                     # list reachable workspaces
mnerve status                             # one-line reachability probe
mnerve access --client <c> --format <f>   # print MCP config for a host
mnerve serve-mcp                          # optional stdio MCP dev fallback
mnerve version                            # installed package version

Authentication