MCP server that provides code execution capabilities using Azure Dynamic Sessions, allowing Python code execution and file management
This server has fundamental definition quality issues. While all 4 tools are explicitly registered with the mcp-go framework and have descriptions, the descriptions are often vague or incomplete. Most critically, all tools lack documented output schemas, LLMs cannot understand what fields to expect in responses. Parameter descriptions exist but are minimal. The tool composition (new_session + exec + list_files + download_file) follows a coherent pattern for code execution, but lacks idempotency documentation and error guidance. The server accepts optional session_id parameters that default to an auto-generated one, creating stateful session tracking that conflicts with the MCP protocol's stateless model. No tool provides recovery guidance or error classification.
Download a file from the session to the local computer
Execute Python code in a session. Store all files generated in the `/mnt/data/` directory
List files in the session
Create a new session, any generated files are stored in the session
No output schemas documented. LLMs cannot understand what fields responses contain. new_session returns a string (session ID), exec returns text output, list_files returns a string representation of file list, download_file returns a BlobResourceContents, but none of this is documented in the tool definitions.
Stateful session management conflicts with MCP's stateless protocol. The Server struct maintains a sessionID field initialized in NewServer() and Start(), and handlers default to this if session_id param is omitted. This violates the protocol's stateless request handling, each request should be self-contained. The MCP spec (2026-07-28) requires stateless handling with _meta in each request.
Missing error recovery guidance. Code execution can fail in many ways (syntax errors, timeouts, permission issues, file not found), but error handlers do not categorize errors or suggest recovery steps. execHandler silently returns empty string if resp is empty; no guidance on retry or user action.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 43 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 44 | - | v1 |
Parameter descriptions are minimal and lack format/constraint details. 'code' param says 'Python code to execute' but omits timeout, size limits, language specifics (e.g., available libraries), or failure modes. 'session_id' and 'reason' lack enough context for LLMs to use them confidently.
No pagination or result limits documented. list_files returns a string representation of all files with no indication of pagination, max count, or performance implications. If a session has thousands of files, the response could bloat the context window.
Idempotency not declared. exec and download_file are write/read operations, but no documentation clarifies if repeated calls with the same parameters are safe. Code execution is inherently non-idempotent (side effects accumulate), but this is never stated.
Tool naming convention partially inconsistent. 'new_session' uses verb_noun correctly, but 'exec' is vague, it could mean many things. Better names: 'execute_code' or 'run_python_code'. 'list_files' and 'download_file' follow the convention well.