mirror of
https://github.com/PrefectHQ/fastmcp.git
synced 2026-08-10 07:39:10 +02:00
b9153404f4* code mode * update uv.lock for monty optional dep 🤖 Generated with Claude Code * retry CI * Address PR review comments on CodeMode transform 🤖 Generated with Claude Code Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Fix ty unresolved-attribute error on search_helper 🤖 Generated with Claude Code Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * more idiomacy * harden * fix docs * harden * fix red CI * Refactor CodeMode to use CatalogTransform base class Removes the duplicate ContextVar bypass pattern in favor of the shared CatalogTransform machinery. Also fixes a pre-existing bug where `from __future__ import annotations` caused NameError for Annotated in nested function scopes at runtime. * Remove redundant _get_visible_tools wrapper in CodeMode * Rewrite CodeMode docs with proper motivation and structure * Fix type narrowing in collision test * Stop unwrapping tool results in CodeMode's call_tool call_tool() inside execute blocks now returns structured content as-is, preserving the {"result": value} wrapping. This means the output schema shown in search results accurately describes what call_tool() returns, so LLMs can trust the schema when writing code. Also adds examples/code_mode/ with a server and narrated client demo. * Simplify call_tool return type: dict | str * Fix example client to unwrap structured results * Let server resolve tool versions instead of pinning first match * Rewrite CodeMode docs to match current behavior * Rename optional extra from monty to code-mode --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Jeremiah Lowin <153965+jlowin@users.noreply.github.com>
37 lines
1.8 KiB
Markdown
37 lines
1.8 KiB
Markdown
# Code Mode
|
|
|
|
CodeMode collapses an entire tool catalog into two meta-tools: `search` (keyword-based discovery) and `execute` (run Python scripts that chain tool calls in a sandbox). Instead of burning context tokens on every intermediate result, the LLM writes a script that runs server-side and returns only the final answer.
|
|
|
|
## Run
|
|
|
|
```bash
|
|
uv run python server.py # in one terminal
|
|
uv run python client.py # in another
|
|
```
|
|
|
|
## Example Output
|
|
|
|
```
|
|
══════════════════ CodeMode Transform ══════════════════
|
|
|
|
┌────────────── list_tools() ──────────────┐
|
|
│ Tool Description │
|
|
│ search Search for available tools ... │
|
|
│ execute Chain `await call_tool(...)` ... │
|
|
└── 8 backend tools collapsed into 2 ──────┘
|
|
|
|
┌──── search(query="math arithmetic") ─────┐
|
|
│ # Tool Description │
|
|
│ 1 add Add two numbers together. │
|
|
│ 2 multiply Multiply two numbers. │
|
|
│ 3 fibonacci Generate the first n ... │
|
|
└── 3 results ─────────────────────────────┘
|
|
|
|
┌────────────── execute ───────────────────┐
|
|
│ a = await call_tool("add", {"a": 3 ... │
|
|
│ b = await call_tool("multiply", ... │
|
|
│ return b │
|
|
└── result: 14.0 ──────────────────────────┘
|
|
```
|
|
|
|
The key insight: with standard MCP, each `call_tool` is a round-trip through the LLM. With CodeMode, the LLM writes one script and all the tool calls happen server-side. Intermediate data never touches the context window.
|