feat(mcp2.0): add and replace tools with mcp #261
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR introduces MCP (Meta Compute Platform) integration across the TinyScientist modules, replacing or augmenting existing local tools with a centralized MCP client/server model. Key changes include:
- Added
tiny_scientist/utils/mcp_client.pyfor a generic MCP client and corresponding server implementations undertiny_scientist/mcp/. - Updated imports in core classes (
Writer,Thinker,Reviewer,Coder,Scientist) to accept anmcp_clientand fall back to legacy tools when MCP is unavailable. - Added a convenience script to launch all MCP servers and updated dependencies/config templates to support the new MCP setup.
Reviewed Changes
Copilot reviewed 13 out of 15 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| tiny_scientist/writer.py | Switched to mcp.tool imports, added mcp_client fallback logic for search & diagram tools |
| tiny_scientist/utils/mcp_client.py | New MCPClient implementation with start/stop/call methods |
| tiny_scientist/mcp/tool.py | Tool definitions updated for MCP namespace (missing os import) |
| tiny_scientist/mcp/*.py | Added MCP server modules for paper search, drawer, and code search |
| scripts/run_mcp_servers.py | New script to start and interactively manage MCP servers |
| pyproject.toml & config.template.toml | Added mcp, fastmcp, httpx deps and MCP server templates |
Comments suppressed due to low confidence (3)
tiny_scientist/utils/mcp_client.py:21
- Type annotation
subprocess.Popen[str]is invalid—Popenis not a generic class; usesubprocess.Popen[Any]or omit the type parameter.
self.servers: Dict[str, subprocess.Popen[str]] = {}
tiny_scientist/utils/mcp_client.py:1
- The new
MCPClientand MCP server code are significant features but have no accompanying tests; consider adding unit tests for methods likestart_server,call_tool, andhealth_check.
import json
tiny_scientist/mcp/tool.py:11
- The module uses
os.pathbut does not importos; addimport osat the top to avoid NameError.
import toml
| # Create tool call request | ||
| request = { | ||
| "jsonrpc": "2.0", | ||
| "id": 1, |
There was a problem hiding this comment.
Using a static JSON-RPC id for every call can cause conflicts if multiple calls happen concurrently; consider generating unique or incrementing IDs per request.
Copilot uses AI. Check for mistakes.
tiny_scientist/writer.py
Outdated
| self.drawer: BaseTool = DrawerTool(model, prompt_template_dir, temperature) | ||
| self.mcp_client = mcp_client | ||
| # Fallback to traditional tools if MCP is not available | ||
| self.searcher: Optional[BaseTool] = PaperSearchTool(s2_api_key=s2_api_key) if not mcp_client else None |
There was a problem hiding this comment.
The fallback initialization of legacy tools when mcp_client is missing is duplicated in multiple classes; consider extracting into a shared factory/helper to reduce duplication.
Copilot uses AI. Check for mistakes.
pyproject.toml
Outdated
| fastmcp = "*" | ||
| mcp = "*" | ||
| httpx = "*" |
There was a problem hiding this comment.
[nitpick] Using wildcard versions (*) for critical dependencies can lead to unpredictable breakage; pin fastmcp, mcp, and httpx to specific version ranges for reproducible builds.
| fastmcp = "*" | |
| mcp = "*" | |
| httpx = "*" | |
| fastmcp = "^1.0.0" | |
| mcp = "^2.0.0" | |
| httpx = "^0.24.0" |
Copilot uses AI. Check for mistakes.
Closes #
📑 Description
✅ Checks
type/descript(e.g.feature/add-llm-agents)ℹ Additional Information