Claude provider: empty tool_filter silently excludes all MCP tools

[franklixuefei]

Bug Description

When using the Claude provider with MCP servers configured, all MCP tools are silently excluded from the API request. Agents only receive the emit_output tool, making them unable to use any MCP server capabilities (filesystem, web search, etc.).

Root Cause

In src/conductor/providers/claude.py, the _convert_mcp_tools_to_claude() method (line ~363):

if tool_filter is not None and tool["name"] not in tool_filter:
    continue

The tool_filter parameter receives [] (empty list) from the call chain:

1. AgentExecutor.__init__ sets self.workflow_tools = workflow_tools or [] (agent.py line 108)
2. resolve_agent_tools(agent.tools=None, workflow_tools=[]) returns [] (agent.py line 61)
3. _execute_with_retry passes tools=[] to _convert_mcp_tools_to_claude

Since [] is not None evaluates to True, and every tool name not in [] is also True, the loop skips every MCP tool.

Expected Behavior

Per the docstring: "If None, all tools are included." An empty list [] should also mean "no filter / all tools", since it represents "no specific tools requested" rather than "explicitly zero tools".

Actual Behavior

• MCP servers connect successfully and report tools (e.g., 14 filesystem tools, 6 web-search tools)
• _convert_mcp_tools_to_claude returns [] — all tools filtered out
• Claude API receives only emit_output — no MCP tools
• Agent cannot read files, search, or use any MCP capability
• No error or warning is logged

Reproduction

workflow:
  name: test
  entry_point: reader
  runtime:
    provider: claude
    mcp_servers:
      filesystem:
        command: npx
        args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
        tools: ["*"]
  input:
    path: { type: string, required: true }

agents:
  - name: reader
    model: claude-sonnet-4.6
    prompt: "Use read_file to read {{ workflow.input.path }}"
    output:
      content: { type: string }
    routes:
      - to: $end

ANTHROPIC_API_KEY=sk-ant-... conductor run test.yaml --input path="/tmp/test.txt"

The agent will report it has no tools available — only emit_output.

Suggested Fix

One-line change — replace the is not None check with a truthiness check:

# Before (bug):
if tool_filter is not None and tool["name"] not in tool_filter:
    continue

# After (fix):
if tool_filter and tool["name"] not in tool_filter:
    continue

This makes tool_filter=[] equivalent to tool_filter=None (no filter, include all tools). Only a non-empty list selects specific tools.

Environment

• Conductor v0.1.0 (commit 391ddaf)
• Python 3.12.13
• mcp SDK 1.26.0
• anthropic SDK 0.84.0
• MCP server: @modelcontextprotocol/server-filesystem (confirmed 14 tools returned via list_tools)

Impact

High — This bug makes the Claude provider effectively unable to use MCP tools at all, despite MCP servers being properly configured and connected. The Copilot provider is not affected (it has built-in tools and doesn't use this code path).

[m13v]

nice catch, the truthiness fix is the right approach. we hit a similar pattern building MCP servers where empty arrays vs None have very different semantics but get conflated in Python. in our case it was tool filtering on the server side - we had to be explicit about distinguishing 'no filter specified' (None, return everything) from 'empty filter' ([], which could mean either return nothing or return everything depending on convention). the suggested one-liner fix is clean. might also be worth adding a log.warning when tool_filter is an empty list, since it's ambiguous enough that other callers might hit the same confusion.

[m13v]

our MCP server for macOS automation if it helps as a reference for tool filtering patterns: https://github.com/mediar-ai/mcp-server-macos-use/blob/main/Sources/MCPServer/main.swift

[franklixuefei]

thanks! I'm trying to get conductor to run a full cycle of implementing a plan successfully with Claude provider, and am filing issues as I find them.

[jrob5756]

thanks! I'm trying to get conductor to run a full cycle of implementing a plan successfully with Claude provider, and am filing issues as I find them.

This is great, @franklixuefei ! Thanks for doing this. The claude provider hasn't been tested nearly as much. Will try to get to these quickly.

[m13v]

filing issues as you go is the right approach. the Claude provider's tool handling has some quirks compared to OpenAI's, especially around how tool results get formatted back. if you're seeing the tool_filter issue specifically, double check that the filter array isn't being initialized as empty vs null - they behave differently.

[m13v]

yeah the Claude provider seems like it needs a few more edge case fixes before it's solid. we found similar patterns when integrating Claude's tool use - the truthiness/falsy checks trip you up because empty arrays and objects behave differently than you'd expect in the filter logic. hope you get the full cycle working, conductor looks promising