Refactor agent to read tracker data via remote MCP server

Gemfile

@@ -74,3 +74,5 @@ gem "marksmith", "~> 0.4.5"
 gem "commonmarker", "~> 2.3"
 
 gem "appsignal"
+
+gem "mcp"

Gemfile.lock

@@ -221,6 +221,9 @@ GEM
       actionview (>= 5.0.0)
       activesupport (>= 5.0.0)
     json (2.12.2)
+    json-schema (6.2.0)
+      addressable (~> 2.8)
+      bigdecimal (>= 3.1, < 5)
     language_server-protocol (3.17.0.5)
     lint_roller (1.1.0)
     llhttp-ffi (0.5.1)
@@ -238,6 +241,8 @@ GEM
     marcel (1.0.4)
     marksmith (0.4.5)
       activesupport
+    mcp (0.9.1)
+      json-schema (>= 4.1)
     meta-tags (2.22.1)
       actionpack (>= 6.0.0, < 8.1)
     method_source (1.1.0)
@@ -456,6 +461,7 @@ DEPENDENCIES
   importmap-rails (~> 2.1)
   jbuilder
   marksmith (~> 0.4.5)
+  mcp
   nokogiri (~> 1.18)
   pdf-reader (~> 2.12)
   pg (~> 1.1)

agent/.env.example

@@ -1,13 +1,16 @@
 # Claude API
 ANTHROPIC_API_KEY=sk-ant-...
 
-# Read-only Postgres connection for the agent
+# Read-only Postgres connection (only needed for entry processing and batch runs)
 AGENT_DATABASE_URL=postgresql://agent_reader:password@localhost:5432/outcome_tracker_development
 
-# Rails API for write operations
+# Rails API for write operations and REST reads
 RAILS_API_URL=http://localhost:3000
 RAILS_API_KEY=agent-secret-key
 
+# Remote MCP server URL (defaults to RAILS_API_URL/mcp)
+# MCP_SERVER_URL=https://www.buildcanada.com/mcp
+
 # Agent configuration
 AGENT_MODEL=claude-opus-4-6
 AGENT_MAX_COMMITMENTS_PER_RUN=100

agent/CLAUDE.md

@@ -112,7 +112,8 @@ Base URL: provided in system prompt. Auth: `Authorization: Bearer <key>` (also i
 ## Rules
 
 - Do NOT use Read, Glob, Grep, or filesystem tools to explore the Rails codebase. Everything you need is above.
-- Use the MCP tools (get_commitment, get_bill, etc.) for reading data.
+- Use the remote MCP tools (mcp__tracker__get_commitment, mcp__tracker__list_bills, etc.) for reading tracker data.
+- Use the local MCP tools (mcp__agent__get_entry, mcp__agent__list_unprocessed_entries, mcp__agent__fetch_government_page) for entry processing and page fetching.
 - Use curl via Bash for ALL write operations.
 - Every judgement (assess_criterion, create_commitment_event, update_commitment_status) MUST include a source_url that was previously fetched via fetch_government_page.
 - Fetch pages BEFORE referencing them. The fetch auto-registers them as Sources in the DB.

agent/src/agent/evaluator.py

@@ -26,15 +26,7 @@
     SYSTEM_PROMPT,
     WEEKLY_SCAN_PROMPT,
 )
-from agent.tools.db_read import (
-    get_bill,
-    get_bills_for_parliament,
-    get_commitment,
-    get_commitment_sources,
-    get_entry,
-    list_commitments,
-    list_unprocessed_entries,
-)
+from agent.tools.db_read import get_entry, list_unprocessed_entries
 from agent.tools.web_search import fetch_government_page
 from agent.tools.rails_write import register_source
 
@@ -63,58 +55,7 @@ def _tool_error(e: Exception, tool_name: str, args: dict) -> dict[str, Any]:
     }
 
 
-# ── Read-only DB tools (via MCP) ───────────────────────────────────────────
-
-@tool(
-    "get_commitment",
-    "Fetch a commitment with its criteria, matches, events, linked bills, departments, and source documents.",
-    {"commitment_id": int},
-    annotations=ToolAnnotations(readOnlyHint=True),
-)
-async def get_commitment_tool(args: dict[str, Any]) -> dict[str, Any]:
-    _tool_log("get_commitment", f"Loading commitment {args['commitment_id']}")
-    try:
-        return _tool_result(get_commitment(args["commitment_id"]), "get_commitment")
-    except Exception as e:
-        return _tool_error(e, "get_commitment", args)
-
-
-@tool(
-    "list_commitments",
-    "List commitments with optional filters. Params: status, policy_area, commitment_type, stale_days, limit.",
-    {
-        "type": "object",
-        "properties": {
-            "status": {"type": "string", "enum": ["not_started", "in_progress", "completed", "broken"]},
-            "policy_area": {"type": "string", "description": "Policy area slug"},
-            "commitment_type": {"type": "string"},
-            "stale_days": {"type": "integer"},
-            "limit": {"type": "integer"},
-        },
-    },
-    annotations=ToolAnnotations(readOnlyHint=True),
-)
-async def list_commitments_tool(args: dict[str, Any]) -> dict[str, Any]:
-    _tool_log("list_commitments", f"filters: {args}")
-    try:
-        return _tool_result(list_commitments(**args), "list_commitments")
-    except Exception as e:
-        return _tool_error(e, "list_commitments", args)
-
-
-@tool(
-    "get_bill",
-    "Fetch a bill with all stage dates (House/Senate readings, Royal Assent) and linked commitments.",
-    {"bill_id": int},
-    annotations=ToolAnnotations(readOnlyHint=True),
-)
-async def get_bill_tool(args: dict[str, Any]) -> dict[str, Any]:
-    _tool_log("get_bill", f"Loading bill {args['bill_id']}")
-    try:
-        return _tool_result(get_bill(args["bill_id"]), "get_bill")
-    except Exception as e:
-        return _tool_error(e, "get_bill", args)
-
+# ── Agent-local tools (not served by the remote MCP server) ────────────────
 
 @tool(
     "get_entry",
@@ -144,35 +85,6 @@ async def list_unprocessed_entries_tool(args: dict[str, Any]) -> dict[str, Any]:
         return _tool_error(e, "list_unprocessed_entries", args)
 
 
-@tool(
-    "get_commitment_sources",
-    "Get the source documents (platform, Speech from the Throne, budget) for a commitment. Use this to determine where a commitment originated for the budget evidence rule.",
-    {"commitment_id": int},
-    annotations=ToolAnnotations(readOnlyHint=True),
-)
-async def get_commitment_sources_tool(args: dict[str, Any]) -> dict[str, Any]:
-    _tool_log("get_commitment_sources", f"commitment {args['commitment_id']}")
-    try:
-        return _tool_result(get_commitment_sources(args["commitment_id"]), "get_commitment_sources")
-    except Exception as e:
-        return _tool_error(e, "get_commitment_sources", args)
-
-
-@tool(
-    "get_bills_for_parliament",
-    "Get all government bills for a parliament session with their stage dates.",
-    {"type": "object", "properties": {"parliament_number": {"type": "integer"}}},
-    annotations=ToolAnnotations(readOnlyHint=True),
-)
-async def get_bills_for_parliament_tool(args: dict[str, Any]) -> dict[str, Any]:
-    pn = args.get("parliament_number", 45)
-    _tool_log("get_bills_for_parliament", f"parliament {pn}")
-    try:
-        return _tool_result(get_bills_for_parliament(pn), "get_bills_for_parliament")
-    except Exception as e:
-        return _tool_error(e, "get_bills_for_parliament", args)
-
-
 @tool(
     "fetch_government_page",
     "Fetch and parse content from an official Canadian government webpage (*.canada.ca / *.gc.ca only). "
@@ -225,27 +137,22 @@ async def fetch_government_page_tool(args: dict[str, Any]) -> dict[str, Any]:
         return _tool_error(e, "fetch_government_page", args)
 
 
-# ── MCP Server (read tools + fetch only) ───────────────────────────────────
+# ── MCP Servers ────────────────────────────────────────────────────────────
 
-ALL_TOOLS = [
-    get_commitment_tool,
-    list_commitments_tool,
-    get_bill_tool,
+# Agent-local tools that need direct DB access or are side-effecting.
+# Read-only tracker tools are served by the remote MCP server (POST /mcp).
+LOCAL_TOOLS = [
     get_entry_tool,
     list_unprocessed_entries_tool,
-    get_commitment_sources_tool,
-    get_bills_for_parliament_tool,
     fetch_government_page_tool,
 ]
 
-tracker_server = create_sdk_mcp_server(
-    name="tracker",
+agent_server = create_sdk_mcp_server(
+    name="agent",
     version="1.0.0",
-    tools=ALL_TOOLS,
+    tools=LOCAL_TOOLS,
 )
 
-ALLOWED_TOOLS = [f"mcp__tracker__{t.name}" for t in ALL_TOOLS] + ["Bash", "WebSearch"]
-
 
 # ── Agent runner ────────────────────────────────────────────────────────────
 
@@ -263,11 +170,31 @@ def _build_options() -> ClaudeAgentOptions:
 
     model = os.environ.get("AGENT_MODEL", "claude-sonnet-4-6")
 
+    # Remote MCP server for read-only tracker tools (commitments, departments,
+    # bills, promises, ministers, feed items, dashboard, burndown).
+    # Agent-local MCP server for entry tools and fetch_government_page.
+    tracker_url = os.environ.get("MCP_SERVER_URL", f"{rails_url}/mcp")
+
+    remote_tools = [
+        "mcp__tracker__list_policy_areas",
+        "mcp__tracker__list_commitments", "mcp__tracker__get_commitment",
+        "mcp__tracker__list_departments", "mcp__tracker__get_department",
+        "mcp__tracker__list_promises", "mcp__tracker__get_promise",
+        "mcp__tracker__list_bills", "mcp__tracker__get_bill",
+        "mcp__tracker__list_ministers", "mcp__tracker__list_activity",
+        "mcp__tracker__get_commitment_summary", "mcp__tracker__get_commitment_progress",
+    ]
+    local_tools = [f"mcp__agent__{t.name}" for t in LOCAL_TOOLS]
+    allowed_tools = remote_tools + local_tools + ["Bash", "WebSearch"]
+
     return ClaudeAgentOptions(
         model=model,
         system_prompt=SYSTEM_PROMPT + api_context,
-        mcp_servers={"tracker": tracker_server},
-        allowed_tools=ALLOWED_TOOLS,
+        mcp_servers={
+            "tracker": {"type": "url", "url": tracker_url},
+            "agent": agent_server,
+        },
+        allowed_tools=allowed_tools,
         permission_mode="bypassPermissions",
         cwd=str(pathlib.Path(__file__).resolve().parent.parent.parent),  # agent/ dir where CLAUDE.md lives
         setting_sources=["project"],

agent/src/agent/main.py

@@ -1,18 +1,32 @@
 """CLI entry point for the commitment evaluation agent."""
 
-import sys
+import os
 import time
 import uuid
 
 import click
+import httpx
 
 from agent.evaluator import (
     evaluate_commitment,
     process_bill_change,
     process_entry,
     weekly_scan_commitment,
 )
-from agent.tools.db_read import list_commitments
+
+
+def _fetch_commitments(status=None, policy_area=None, limit=100):
+    """Fetch commitments from the Rails REST API, oldest-assessed first."""
+    rails_url = os.environ.get("RAILS_API_URL", "http://localhost:3000")
+    params = {"per_page": limit, "sort": "last_assessed_at", "direction": "asc"}
+    if status:
+        params["status"] = status
+    if policy_area:
+        params["policy_area"] = policy_area
+
+    resp = httpx.get(f"{rails_url}/commitments", params=params, timeout=30.0)
+    resp.raise_for_status()
+    return resp.json().get("commitments", [])
 
 
 @click.group()
@@ -70,7 +84,7 @@ def scan_all(limit: int, status: str | None, policy_area: str | None, as_of: str
     click.echo("Starting weekly scan...")
     run_id = str(uuid.uuid4())
 
-    commitments = list_commitments(
+    commitments = _fetch_commitments(
         status=status,
         policy_area=policy_area,
         limit=limit,