Allsource mcp setup

AllSource MCP Setup

Configure the AllSource MCP server for local debugging in Claude Code.

Triggers on: "configure allsource mcp", "setup mcp", "allsource mcp config", "add allsource to claude code", "mcp server setup".

---

Option 1: Rust stdio MCP server (recommended for local debugging)

Reads directly from WAL + Parquet files. No running server needed.

Build

cd <chronos-monorepo>
cargo build --release -p allsource-mcp

Binary at: target/release/allsource-mcp

Claude Code Configuration

Add to ~/.claude/settings.json or project .claude/settings.json:

{
  "mcpServers": {
    "allsource": {
      "command": "<path-to>/allsource-mcp",
      "args": ["--data-dir", "<path-to-data>"]
    }
  }
}

Longhand on macOS

{
  "mcpServers": {
    "allsource": {
      "command": "/Users/<you>/Projects/chronos/chronos-monorepo/target/release/allsource-mcp",
      "args": ["--data-dir", "/Users/<you>/Library/Application Support/Longhand/allsource"]
    }
  }
}

Using environment variable

{
  "mcpServers": {
    "allsource": {
      "command": "/path/to/allsource-mcp",
      "env": {
        "ALLSOURCE_DATA_DIR": "/path/to/data"
      }
    }
  }
}

Available Tools

After configuration, these tools appear in Claude Code:

Tool	Description
query_events	Query events (entity_id, event_type, time range, limit)
sample_events	Sample recent events across all entities
quick_stats	Store summary: counts, types, date range, durability
get_snapshot	Projection/snapshot state for an entity
event_timeline	Chronological event list for an entity
explain_entity	Human-readable entity lifecycle summary
reconstruct_state	Fold events to rebuild current state
analyze_changes	Changes within a time window

Verification

After adding the config, restart Claude Code and check:

1. Run /mcp — allsource should appear in the server list
2. Ask Claude to "use quick_stats" — should return event counts
3. Ask Claude to "query events for entity X" — should return events

---

Option 2: Elixir MCP server (remote mode)

Connects to a running Core server over HTTP. Use when Core is running in Docker.

Configuration

{
  "mcpServers": {
    "allsource": {
      "command": "mix",
      "args": ["phx.server"],
      "cwd": "<chronos-monorepo>/apps/mcp-server-elixir",
      "env": {
        "CORE_URL": "http://localhost:3900",
        "CORE_MODE": "remote",
        "PORT": "3904"
      }
    }
  }
}

With Docker stack running

If the Docker compose stack is up (allsource-core-leader on port 3280):

{
  "mcpServers": {
    "allsource": {
      "command": "mix",
      "args": ["phx.server"],
      "cwd": "<chronos-monorepo>/apps/mcp-server-elixir",
      "env": {
        "CORE_URL": "http://localhost:3280",
        "CORE_MODE": "remote"
      }
    }
  }
}

---

Option 3: Elixir MCP server (embedded mode)

Runs Core in-process via Rustler NIF. No separate Core needed, but requires Rust + Elixir toolchain.

{
  "mcpServers": {
    "allsource": {
      "command": "mix",
      "args": ["phx.server"],
      "cwd": "<chronos-monorepo>/apps/mcp-server-elixir",
      "env": {
        "CORE_MODE": "embedded",
        "CORE_DATA_DIR": "/path/to/data"
      }
    }
  }
}

---

Troubleshooting

Server not starting

• "No such file or directory": Check the binary path. Build with cargo build --release -p allsource-mcp.
• "data_dir is required": Pass --data-dir arg or set ALLSOURCE_DATA_DIR env var.
• Permission denied: Check file permissions on the data directory.

No events found

• Verify the data directory has files: ls <data-dir>/storage/ should show .parquet files and/or ls <data-dir>/wal/ should show .log files.
• If empty, the application may not have written events yet, or the path is wrong.
• Check if the data dir has a __system/ subdirectory — if yes, it's an AllSource data dir.

Tools not appearing

• Run /mcp in Claude Code to check server status.
• Check stderr output: set RUST_LOG=debug in the env to see MCP protocol messages.
• Ensure the JSON in settings is valid (no trailing commas).

Slow startup

• Large Parquet collections take time to scan on first open.
• WAL replay happens at startup — many WAL entries = slower start.
• After first open, subsequent queries are fast (data cached in DashMap).

Wrong data / stale results

• The MCP server opens the data directory at startup and replays WAL.
• If the application is still running and writing, new events won't appear until MCP server restarts.
• For live debugging, use the Elixir MCP server in remote mode connected to the running Core.