Persistent memory, task coordination, and a human-in-the-loop queue for every Claude / Cursor / Windsurf session. Open source. Local-first. MCP-native.
Free tier: 1 project · 30 days · no credit card required · self-host for free, forever
Every session, decision, and human-in-the-loop request — coordinated live. This is the real dashboard, not a mockup.
File claims and sprint-item touches_files hints warn agents before they copy work that overlaps with an active session.
Your AI tools don't talk to each other. Meridian is the layer that fixes it.
Every Claude tab boots blind. Run two in parallel and they duplicate work or contradict each other because neither sees the other's plan.
Hit the token limit mid-task and the whole working state evaporates. Re-explaining what you've already shipped is a tax on every session.
Two sessions editing the same dashboard, server, or database file can silently overwrite each other unless the work is claimed and serialized.
Meridian tracks active sessions, sprint-item file intent, and explicit file claims so parallel work stays coordinated.
Executors call claim_file before editing. New sessions see warnings when a live session already owns the same file.
High-contention work such as dashboard.js, server.py, and db/__init__.py can be serialized instead of raced.
touches_files intentSprint items can declare expected files so handoffs warn before another agent copies overlapping work.
Claim a single class or function instead of a whole file — Meridian parses it (ast / tree-sitter) and only blocks on overlapping line ranges, so two sessions can edit the same file safely.
A planner can inject sprint items mid-run; executors pick them up at the next item boundary via a board_change signal — no interruption.
All seven features ship today. Open source, MSL-1.0.
North star, version goal, session focus — every tab reads the same. One save, everyone aligns.
claim_task locks work atomically. No two sessions pick up the same task twice.
When your AI gets stuck, it pauses and asks you — not us.
Sessions surface coordination questions to a single dashboard panel. Autonomous until it needs you — then back to work.
Editable constitution alongside the append-only log. The "current truth" cold sessions need before doing anything.
L0/L1/L2 compression. generate_handoff writes a file a fresh session can paste and resume in seconds.
Claude Code, Codex CLI, Claude Desktop, Cursor, Windsurf, LangGraph webhook — any MCP client, any OS.
Connect your repo once in Settings. Planning sessions get read_file, search_code, git_log and more — no Copilot, no extra connector, no local install.
Wire Claude Code and Codex to Meridian in 30 seconds. SessionStart injects your project context automatically. Stop fires a delta checkpoint before the session closes. Zero manual calls.
# one-liner install
curl -fsSL https://usemeridian.us/hooks.sh | bash
# one-liner install
irm https://usemeridian.us/hooks.ps1 | iex
Injects your project state (goal, sprint, recent tasks, pinned decisions) into every new Claude Code / Codex session automatically.
Before the session closes: buckets completed tasks into an auto-capture note and writes a delta handoff so the next session can resume in seconds.
Get all 20+ Meridian tools directly in your claude.ai planning chat — no copy-pasting session output, no local install.
Works with Claude Code, Cursor, Windsurf, Claude Desktop, ChatGPT, and any MCP client. Self-host free forever or use the hosted tier.
Local SQLite by default. Connect your own Postgres for shared team state. Free forever for personal use.
# macOS / Linux $ curl -fsSL https://pixi.sh/install.sh | bash $ git clone https://github.com/meridianmcp/Meridian $ cd Meridian && pixi run start # Windows PowerShell > iwr -useb https://pixi.sh/install.ps1 | iex > git clone https://github.com/meridianmcp/Meridian > cd Meridian; pixi run start # dashboard → http://localhost:7878/dashboard # MCP endpoint → claude-code, cursor, windsurf all auto-discover via stdio
Or download a single binary — no Python, no git clone:
No install? Use Meridian directly from claude.ai in your browser →
Start free — self-host forever, or try the hosted tier for 30 days with no credit card.
CLAUDE.md is static context — good for standing instructions, bad for tracking state. Meridian is live: sessions read and write to a shared DB in real time, coordinate sprint items atomically, queue HITL requests, and generate handoffs across context resets.
start_session, log_task, generate_handoff, etc.) and any compatible AI client can call them as built-in capabilities.