🤖 👨‍💻 Agent Codemode

Generate programmatic tools from MCP Servers and Skills.

What is Agent Codemode?

Agent Codemode generates programmatic tools from two sources:

1. MCP Servers - Connect to any MCP server and generate typed Python bindings for its tools
2. Skills - Reusable code patterns that compose multiple tools into higher-level operations

These programmatic tools can be:

• Used directly by an agent - Import generated bindings and call tools from your agent's code
• Exposed as an MCP Server - Serve the generated tools via MCP protocol for any MCP-compatible client

Why Agent Codemode?

Traditional AI agents call tools one at a time through LLM inference. Agent Codemode enables a "Code Mode" pattern where agents write Python code that orchestrates multiple tool calls:

• More efficient: Single code generation instead of many LLM tool-call round-trips
• More reliable: Use try/except for robust error handling
• More powerful: Parallel execution with asyncio, loops, conditionals
• More composable: Save and reuse patterns as Skills

Same task, same MCP server — Code Mode uses significantly fewer tokens by composing tools in code instead of multiple LLM round-trips.

Without Code Mode	With Code Mode

Prompt: "Generate 2000 words of random text and write to a file"

Configuration Highlights

Option	Description
allow_direct_tool_calls	When False (default), call_tool is hidden; all execution flows through execute_code
max_tool_calls	Safety cap limiting tool invocations per execute_code run
sandbox_variant	Sandbox type for code execution (default: "eval"). One of eval, monty, docker, jupyter, colab, kaggle, modal, datalayer
sandbox_gpu	Optional GPU flavor / accelerator for supported variants (Modal/Datalayer examples: T4, A10G, A100, H100; Kaggle batch examples: NvidiaTeslaT4, NvidiaTeslaP100, T4, P100)
workspace_path	Working directory for sandbox execution
generated_path	Path where tool bindings are generated
skills_path	Path for saved skills

When using cloud notebook variants via code-sandboxes:

• colab reuses an already-running Colab kernel (from explicit runtime values or a channels URL).
• kaggle defaults to batch execution via Kaggle's API when no runtime URL/channels are provided; it can also create/attach interactive kernels when runtime values are provided.
• kaggle accelerator selection is available in batch mode via gpu / accelerator (for example T4, P100, NvidiaTeslaT4).

Tool Discovery Options

• list_tool_names: Fast listing with server, keywords, limit filters. include_deferred=False by default
• search_tools: Natural language search with query, server, limit. include_deferred=True by default
• get_tool_details: Get full schema and documentation for a specific tool

Tool Metadata

Tools include output_schema and input_examples to improve parameter accuracy. Tools marked with defer_loading=True are excluded from default listings but included in search results.

Installation

pip install agent-codemode

Quick Start

fromagent_codemodeimportToolRegistry, CodeModeExecutor, MCPServerConfig# Set up registry with MCP serversregistry=ToolRegistry()
# Add an MCP server (stdio transport - uses command/args)registry.add_server(MCPServerConfig(
name="filesystem",
command="npx",
args=["-y", "@anthropic/mcp-server-filesystem", "/tmp"]
))
# Or add an HTTP-based server# registry.add_server(MCPServerConfig(name="web", url="http://localhost:8001"))awaitregistry.discover_all()
# Execute code that composes toolsasyncwithCodeModeExecutor(registry) asexecutor:
result=awaitexecutor.execute(""" from generated.mcp.filesystem import read_file, write_file # Read multiple files content1 = await read_file({"path": "/tmp/file1.txt"}) content2 = await read_file({"path": "/tmp/file2.txt"}) # Process and combine combined = content1 + "\\n---\\n" + content2 # Write result await write_file({"path": "/tmp/combined.txt", "content": combined}) """)

Features

Progressive Tool Discovery

Use the Tool Search Tool to discover relevant tools without loading all definitions upfront:

# Search for tools matching a description (includes deferred tools by default)result=awaitregistry.search_tools("file operations", limit=10)
fortoolinresult.tools:
print(f"{tool.name}: {tool.description}")
# Fast listing (deferred tools excluded by default)names=registry.list_tool_names(limit=50)
# Include deferred tools explicitlynames_all=registry.list_tool_names(limit=50, include_deferred=True)

Code-Based Tool Composition

Execute Python code in an isolated sandbox with auto-generated tool bindings:

asyncwithCodeModeExecutor(registry) asexecutor:
execution=awaitexecutor.execute(""" import asyncio from generated.mcp.filesystem import ls, read_file # List all files files = await ls({"path": "/data"}) # Read all files in parallel contents = await asyncio.gather(*[ read_file({"path": f}) for f in files ]) """, timeout=30.0)
# Outputs are available on the execution objectprint(execution.stdout)
print(execution.stderr)
print(execution.text)
print(execution.success)
print(execution.execution_ok)
print(execution.execution_error)
print(execution.code_error)

🧰 Skills (Reusable Compositions)

Skills are Python files that compose tools into reusable operations. This allows agents to evolve their own toolbox by saving useful code patterns. Skills functionality is provided by the agent-skills package.

Note: Skills APIs are owned by agent-skills. Import skill utilities from agent_skills.

Creating Skills as Code Files

The primary pattern is skills as Python files in a skills/ directory:

# skills/batch_process.py"""Process all files in a directory."""asyncdefbatch_process(input_dir: str, output_dir: str) ->dict:
"""Process all files in a directory. Args: input_dir: Input directory path. output_dir: Output directory path. Returns: Processing statistics. """fromgenerated.mcp.filesystemimportlist_directory, read_file, write_fileentries=awaitlist_directory({"path": input_dir})
processed=0forentryinentries.get("entries", []):
content=awaitread_file({"path": f"{input_dir}/{entry}"})
# Process content...awaitwrite_file({"path": f"{output_dir}/{entry}", "content": content.upper()})
processed+=1return {"processed": processed}

Using Skills in Executed Code

Skills are imported and called like any Python module:

# In executed codefromskills.batch_processimportbatch_processresult=awaitbatch_process("/data/input", "/data/output")
print(f"Processed {result['processed']} files")

Managing Skills with SimpleSkillsManager

For programmatic skill management, use the SimpleSkillsManager:

fromagent_skillsimportSimpleSkillsManager, SimpleSkill# Create a skills managermanager=SimpleSkillsManager("./skills")
# Save a skillskill=SimpleSkill(
name="batch_process",
description="Process files in a directory",
code='''async def batch_process(input_dir, output_dir): entries = await list_directory({"path": input_dir}) for entry in entries.get("entries", []): content = await read_file({"path": f"{input_dir}/{entry}"}) await write_file({"path": f"{output_dir}/{entry}", "content": content.upper()})''',
tags=["file", "batch"],
)
manager.save_skill(skill)
# Load and use a skillloaded=manager.load_skill("batch_process")
print(loaded.code)

Examples

See the runnable examples in examples/.

Simple Examples

python examples/simple/codemode_example.py
python examples/simple/codemode_patterns_example.py

Agent CLI

Interactive CLI agent with Agent Codemode support:

# Standard mode
python examples/agent/agent_cli.py
# Codemode variant (code-first tool composition)
python examples/agent/agent_cli.py --codemode

Pydantic AI Integration

Use the CodemodeToolset for direct integration with Pydantic AI agents:

frompydantic_aiimportAgentfromagent_codemodeimportCodemodeToolset, ToolRegistry, MCPServerConfig# Set up registryregistry=ToolRegistry()
registry.add_server(MCPServerConfig(
name="filesystem",
command="npx",
args=["-y", "@anthropic/mcp-server-filesystem", "/tmp"]
))
awaitregistry.discover_all()
# Create toolsettoolset=CodemodeToolset(registry=registry)
# Use with Pydantic AI agentagent=Agent(
model='anthropic:claude-sonnet-4-5',
toolsets=[toolset],
)

MCP Server Mode

Expose the programmatic tools as an MCP server for any MCP-compatible client:

fromagent_codemodeimportcodemode_server, configure_serverfromagent_codemodeimportToolRegistry, MCPServerConfig, CodeModeConfig# Create and configure registry with MCP servers to composeregistry=ToolRegistry()
registry.add_server(MCPServerConfig(
name="filesystem",
command="npx",
args=["-y", "@anthropic/mcp-server-filesystem", "/tmp"]
))
# Configure with custom settingsconfig=CodeModeConfig(
workspace_path="./workspace",
skills_path="./skills",
generated_path="./generated",
)
configure_server(config=config, registry=registry)
codemode_server.run()

Or start with command line:

python -m agent_codemode.server --workspace ./workspace

Connect to an MCP Client

Create a launcher script to configure which MCP servers to compose:

#!/usr/bin/env python3importsysfrompathlibimportPathfromagent_codemodeimportToolRegistry, MCPServerConfig, CodeModeConfigfromagent_codemode.serverimportconfigure, run# Create registry and add MCP servers to composeregistry=ToolRegistry()
registry.add_server(MCPServerConfig(
name="my_server",
command=sys.executable,
args=["/path/to/my_mcp_server.py"]
))
# Configure pathsconfig=CodeModeConfig(
workspace_path="./workspace",
generated_path="./generated",
skills_path="./skills",
)
configure(config=config, registry=registry)
run()

Then configure your MCP client to run the launcher script.

Tools exposed by the MCP server:

Tool	Description
search_tools	Progressive tool discovery
list_servers	List connected MCP servers
list_tool_names	Fast listing of tool names
get_tool_details	Get full tool schema
execute_code	Run code that composes tools
call_tool	Direct tool invocation
save_skill / run_skill	Skill management
list_skills / delete_skill	Skill management
add_mcp_server	Dynamically add servers

Recommended System Prompt

When building an agent that uses Codemode, use this system prompt pattern:

You are an AI assistant with access to MCP tools via Code Mode.
## Available Meta-Tools
- **list_tool_names** - Fast listing of tool names
- **search_tools** - AI-powered tool discovery (returns full definitions)
- **get_tool_details** - Get schema for a specific tool
- **execute_code** - Execute Python code in a sandboxed environment
## Execution Model
ALL tool execution must go through execute_code. Write Python code that imports
and uses the generated tool bindings:
from generated.mcp.filesystem import read_file
content = await read_file({"path": "/data/config.json"})
print(content)

Workflow

1. Discover tools using search_tools or list_tool_names
2. Write Python code that imports tools from generated.mcp.
3. Execute using execute_code

See the Getting Started guide for a complete system prompt example.

Key Patterns

Tool Discovery

Instead of loading all tool definitions upfront (which can overwhelm context), use the Tool Search Tool pattern for progressive discovery based on the task at hand.

Tool Composition

Compose tools through code instead of reading all data into LLM context. This is faster, more reliable (no text reproduction errors), and more efficient.

Control Flow

Code allows models to implement complex control flow: loops, conditionals, waiting patterns, and parallel execution without burning through context with repeated tool calls.

State Persistence

When running in a sandbox, state can persist between execute_code calls within the same session. Variables, functions, and imported modules remain available for subsequent code executions. Skills can also be saved to disk and loaded later for reuse across sessions.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│ Agent Codemode │
├─────────────────────────────────────────────────────────────────────┤
│ ┌─────────────────┐ ┌────────────────┐ ┌───────────────────────┐ │
│ │ Tool Registry │ │ Code Executor │ │ CodemodeToolset │ │
│ │ - Discovery │ │ - Sandbox │ │ (Pydantic AI) │ │
│ │ - Search │ │ - Bindings │ │ - search_tools │ │
│ │ - Cache │ │ - Execute │ │ - execute_code │ │
│ └─────────────────┘ └────────────────┘ └───────────────────────┘ │
│ │ │
│ ┌───────────────┴───────────────┐ │
│ │ Generated Bindings │ │
│ │ generated/mcp/<name>/ │ │
│ └───────────────────────────────┘ │
├────────────────────────────────┬────────────────────────────────────┤
│ MCP Servers │ Agent Skills (agent_skills) │
│ (filesystem, bash, web, etc. │ (SimpleSkillsManager, │
│ connected via MCP protocol) │ SkillDirectory, skills/*.py) │
└────────────────────────────────┴────────────────────────────────────┘

References

• Introducing Code Mode - Cloudflare
• Code Execution with MCP - Anthropic
• Programmatic Tool Calling - Anthropic
• Advanced Tool Use - Anthropic
• Programmatic MCP Prototype

CI Workflows

This repository uses a reusable GitHub Actions workflow at .github/workflows/reusable-python.yml.

The following workflows call it:

• .github/workflows/build.yml
• .github/workflows/py-tests.yml
• .github/workflows/py-code-style.yml
• .github/workflows/py-typing.yml

Reusable workflow inputs:

• python-version: Python version to run.
• install-system-deps: Install Linux dependencies and unlock keyring.
• install-extras: Extras from pyproject.toml (for example test,typing).
• extra-packages: Additional packages installed with uv pip install.
• run-tests: Enable test execution.
• test-command: Command used for tests.
• run-mypy: Enable mypy.
• mypy-target: Package or module passed to mypy.
• run-pre-commit: Enable pre-commit checks.

License

BSD 3-Clause License