@phosra/mcp

MCP server for Phosra — exposes 76 child-safety tools to any MCP-compatible AI client (Claude Desktop, Cursor, etc.).

What's new in v0.3.0

The Continuous Loop release adds 9 tools across three pillars:

Reasoning memory (Phase A):

• list_unreviewed_agent_decisions — surface autonomous actions the agent took on your behalf that you haven't acknowledged
• mark_provenance_reviewed — companion write tool, idempotent
• query_decisions_by_law — "show me everything we did because of KOSA"

Reversibility (Phase B):

• dry_run_apply — preview what apply_brief_item would do without writing
• snapshot_policy — capture the family's policy state as a named immutable snapshot
• revert_to_snapshot — atomically restore

Streaming (Phase C):

• subscribe_brief_stream — push new Brief items as the materializer writes them
• subscribe_enforcement_failures — real-time enforcement-failure events
• subscribe_statute_registry — new statutes / status flips

The streaming pattern

The three subscribe_* tools return a stream_url plus a subscription_id. The agent must open a separate SSE connection to that URL with the same Authorization: Bearer phomcp_… header.

The stream URLs point at https://prodapi.phosra.com/api/v1/families/{family_id}/streams/{brief,enforcement-failures,statute-registry} directly — they bypass the hosted /api/mcp route on Vercel because Vercel functions have execution-time caps incompatible with long-lived SSE connections.

Installation

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "phosra": {
      "command": "npx",
      "args": ["-y", "@phosra/mcp", "--api-key", "phomcp_YOUR_TOKEN"]
    }
  }
}

Or use the hosted endpoint at https://www.phosra.com/api/mcp (JSON-RPC; streaming tools work the same).

Get a token

Generate at https://www.phosra.com/dashboard/settings/mcp.