Meridian

The vault is the single source of truth. Any agent reads from it, plans against it, executes, and hands control back through it.

Meridian is a framework for Human ↔ Agent collaboration. It uses an Obsidian vault as shared state between humans and any AI agent — Claude Code, Cursor, a GPT, a CLI script. Agents don't share memory or APIs. They share the vault.

→ Full concepts and philosophy

How It Works

Vault (Projects + Tasks)  ←────────────────────────►  Any Agent (Execution)
        ▲                                                       │
        │  status fields are the only shared state              │
        │                                                       ▼
   Human reviews                                   Reads · Plans · Builds
   & approves                                      Creates review checkpoints
        ▲                                                       │
        └───────────────── vault write ─────────────────────────┘

1. Define — Human writes a task in a Project note (owner::agent)
2. Pick up — Agent finds the next agent/backlog task for project X
3. Plan — A planning artifact (PRD, ADR, RFC…) is produced or loaded via /mdn-load
4. Checkpoint — Meridian inserts a owner::me status::review task pointing to the plan
5. Approve — Human reads, edits if needed, runs /mdn-approve
6. Execute — Agent picks up the approved task and executes against the plan
7. Review — Agent creates a verification checkpoint; human confirms or requests changes
8. Done — Loop continues from step 2 (or ends here)

Installation

# Default (Claude Code)
bash <(curl -fsSL https://raw.githubusercontent.com/user/meridian/main/install.sh)

# For Codex / OpenAI
bash <(curl -fsSL https://raw.githubusercontent.com/user/meridian/main/install.sh) --tool codex

Run /mdn-init in any Claude Code session to complete first-time setup.

Quick Start

/mdn-init name:my-app       # initialize a new project
/mdn-add project:my-app     # add a task
/mdn-load project:my-app path:/path/to/plan.md type:prd   # load a planning artifact
/mdn-approve project:my-app # review, then approve + execute in one step
/mdn-status                 # check status at any time

Skills

Command	Description
/mdn-init name:<slug>	First-run setup and project folder structure creation
/mdn-add project:<slug>	Add a human or agent task to a project
/mdn-load project:<slug> path:<path> type:<type>	Register a planning artifact and create a review checkpoint
/mdn-approve project:<slug>	Approve a pending plan and immediately trigger execution
/mdn-run project:<slug>	Execute the next approved agent task in a project
/mdn-plan project:<slug>	Auto-generate plan stubs for all unplanned backlog tasks
/mdn-status	Dashboard: all active projects with task counts by status
/mdn-daily	Daily brief: everything needing human attention today
/mdn-sync project:<slug>	Ingest .md files from registered source directories as plan notes
/mdn-archive project:<slug>	Move completed tasks to TASK_HISTORY.md

→ Full skills reference

Documentation

Document	Contents
concepts.md	Vision, philosophy, state machine, agent vs human roles
workflow.md	Step-by-step guide through the workflow phases
protocol.md	Agent-agnostic protocol spec — implement your own adapter
skills.md	Full Claude Code skills reference
configuration.md	Config format, note structure, task syntax, plan format