English | 中文版
A production-focused MCP service that lets agents/LLMs run SymPy safely and efficiently. It combines AST policy checks, runtime resource limits, and prewarmed workers to deliver low-noise, parse-friendly results.
- Single tool:
sympy(input only requirescode) - Prewarmed worker pool to avoid repeated
import sympy - Two-layer safety: AST guard + runtime resource limits
- Compact structured JSON output for low token overhead
- Standardized error codes for reliable auto-retry workflows
- Symbolic algebra, differentiation, integration, equation solving
- MCP tool integration for Codex / Cursor / Claude Desktop / custom MCP clients
- Agent workflows that need controllable failures and clean error signals
Call example:
fastmcp call \
--command 'python -m sym_mcp.server' \
--target sympy \
--input-json '{"code":"import sympy as sp\\nx=sp.Symbol(\"x\")\\nprint(sp.factor(x**2-1))"}'Client config (python -m, recommended):
{
"mcpServers": {
"sympy-sandbox": {
"command": "python",
"args": ["-m", "sym_mcp.server"]
}
}
}Client config (installed as sym-mcp):
{
"mcpServers": {
"sympy-sandbox": {
"command": "sym-mcp",
"args": []
}
}
}Client config (uvx):
{
"mcpServers": {
"sympy-sandbox": {
"command": "uvx",
"args": ["sym-mcp"]
}
}
}- Python 3.11+
- Linux / macOS (Linux recommended for production)
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -e .
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -e ".[dev]"python -m sym_mcp.serverfastmcp list --command 'python -m sym_mcp.server'sympy
code: str
Notes:
- You must
print()final outputs. - If nothing is printed,
outmay be empty.
Success:
{"out":"x**2/2"}Failure:
{"code":"E_RUNTIME","line":3,"err":"ZeroDivisionError: division by zero","hint":"Runtime error. Check variable types, division-by-zero, or undefined names near the reported line."}Field definitions:
out: stdout text on successcode: error codeline: user code error line, ornullerr: compact error message (traceback noise removed)hint: fix hint (based on configured hint level)- If
out/err/hintis too long, it will be truncated with...[truncated]
E_AST_BLOCK: blocked by AST safety policyE_SYNTAX: syntax errorE_TIMEOUT: timeoutE_MEMORY: memory limit triggeredE_RUNTIME: general runtime errorE_WORKER: worker communication/state failureE_INTERNAL: internal server error
- Use math-only Python code.
- Only import
sympyormath. - Always
print()final answers. - For multiple outputs, use multiple
print()lines. - On failure, patch minimally near
lineand retry. - For
E_TIMEOUT, reduce scale first; forE_MEMORY, reduce object size/dimension; forE_AST_BLOCK, remove unsafe statements.
Example:
import sympy as sp
x = sp.Symbol("x")
expr = (x + 1)**5
print(sp.expand(expr))- Only
sympy/mathimports are allowed - Dangerous capabilities are blocked (
eval,exec,open,__import__, etc.) - Dunder attribute traversal is blocked (e.g.
__class__)
- Per-task CPU time limit + timeout kill
- Per-worker memory limit via
setrlimit - Worker auto-rebuild on failure to keep server healthy
src/sym_mcp/server.py: MCP entrypoint and tool registrationsrc/sym_mcp/security/ast_guard.py: AST validationsrc/sym_mcp/executor/worker_main.py: worker loopsrc/sym_mcp/executor/pool.py: async prewarmed process poolsrc/sym_mcp/executor/sandbox.py: restricted execution and stdout capturesrc/sym_mcp/errors/parser.py: error normalization and code mappingsrc/sym_mcp/config.py: runtime configuration
SYMMCP_POOL_SIZE: worker pool size, default10SYMMCP_EXEC_TIMEOUT_SEC: per execution timeout (sec), default3SYMMCP_MEMORY_LIMIT_MB: memory cap per worker (MB), default150SYMMCP_QUEUE_WAIT_SEC: queue wait timeout (sec), default2SYMMCP_LOG_LEVEL: log level, defaultINFOSYMMCP_MAX_OUTPUT_CHARS: output truncation threshold, default1200SYMMCP_HINT_LEVEL: hint level (none/short/medium), defaultmedium
Most likely the code does not print() the final result.
It is easier for agents to parse reliably and reduces token cost.
setrlimit behavior differs by OS. Linux is preferred for production.
Current primary delivery is stdio. HTTP/SSE can be added later via FastMCP transport extensions.
- This is restricted Python execution, not VM/container-grade isolation
- Memory limit behavior is OS-dependent
- Output is truncated at threshold, with
...[truncated]suffix
PYTHONPATH=src pytest -qPYTHONPATH=src python scripts/benchmark.py --concurrency 100 --total 500- Run
PYTHONPATH=src pytest -qbefore submitting PRs - When adding new capabilities, update:
- error code docs
- README examples
- related unit/integration tests
- Publishing process: PUBLISHING.md