LLM Delegator

LLM Delegator lets a primary agent such as GPT delegate bounded work to another LLM through one read-only MCP tool. The primary agent remains responsible for planning, verification, commands, and file changes.

Provider adapters support DeepSeek's Responses API and OpenRouter's Chat Completions API without changing the MCP interface.

Capabilities

• MCP tool: delegate_task

• CLI: llm-delegator

• Explicit, relative file selection

• Configurable allowlist of workspace roots

• Per-file and aggregate input limits

• Text, Markdown, JSON, and unified-diff output modes

• Token usage returned with every successful result

• No filesystem writes or command execution

Related MCP server: local-llm

Delegation policy

The MCP tool accepts only explicit low or medium complexity classifications. GPT should delegate only routine work that is bounded and easy to verify:

• Summarization, extraction, and classification

• Mechanical content or data transformations

• First drafts of documentation or tests

• Routine code drafts and routine reviews

GPT retains architecture and system design, complex debugging, ambiguous requirements, security/privacy/auth decisions, final verification, live-state research, external actions, and destructive work. If the primary model is uncertain whether a task is routine, it should keep the task.

Model routing is automatic:

• low complexity always uses the selected provider's Flash alias.

• medium complexity uses the selected provider's Pro alias by default.

• OpenRouter maps those aliases to z-ai/glm-5.3-flash and z-ai/glm-5.3.

• High-complexity work cannot be submitted through the MCP schema.

Every handoff must be self-contained. GPT supplies:

• A detailed task describing the intended result

• One or more explicit acceptance criteria

• Relevant constraints

• An ordered plan whenever it helps the worker; medium-complexity tasks cannot run without one

This keeps the worker from guessing missing requirements or reconstructing the primary model's reasoning.

Install

git clone https://github.com/shangzhihao/llm-delegator.git
cd llm-delegator
uv sync

The DeepSeek adapter reads DEEPSEEK_API_KEY; the OpenRouter adapter reads OPENROUTER_API_KEY.

Model configuration

Edit llm-delegator.toml to control which models are active:

[models]
active = [
    { provider = "deepseek", model = "deepseek-v4-flash" },
    { provider = "deepseek", model = "deepseek-v4-pro" },
    { provider = "openrouter", model = "z-ai/glm-5.3-flash" },
    { provider = "openrouter", model = "z-ai/glm-5.3" },
]

Each active model is identified by both provider and provider-specific model ID. This keeps, for example, an OpenRouter-hosted GLM distinct from a future direct Z.ai-hosted GLM with the same model ID. Removing an entry blocks that provider/model pair before an API request. Changes take effect on the next delegation. Set LLM_DELEGATOR_CONFIG to use a file at another path. If the default llm-delegator.toml is absent, models are unrestricted for backward compatibility.

Configure Codex

Add this to ~/.codex/config.toml:

[mcp_servers.llm-delegator]
command = "/absolute/path/to/llm-delegator/.venv/bin/llm-delegator-mcp"
cwd = "/absolute/path/to/llm-delegator"
env_vars = ["DEEPSEEK_API_KEY", "OPENROUTER_API_KEY"]
enabled_tools = ["delegate_task"]
tool_timeout_sec = 240

[mcp_servers.llm-delegator.env]
LLM_DELEGATOR_ALLOWED_ROOTS = "/absolute/path/to/allowed/workspaces"

Restart Codex and use /mcp to confirm that llm-delegator is connected. Codex can then call delegate_task while GPT remains the primary model.

The MCP server advertises these operating instructions to the primary model:

• Delegate only bounded, routine, easy-to-verify work.

• Write a detailed, self-contained task with acceptance criteria and constraints.

• Supply an ordered plan for medium-complexity work and whenever it would help.

• Select only the files needed for the task.

• Treat the result as an untrusted draft and verify it.

• Keep all edits and command execution with the primary model.

CLI

The CLI exercises the same service and provider adapter as MCP:

LLM_DELEGATOR_ALLOWED_ROOTS=/path/to/allowed/workspaces \
  uv run llm-delegator \
  "Summarize the provider interface" \
  --task-kind summarize \
  --complexity low \
  --accept "Describe the interface contract and method signature" \
  --constraint "Do not propose implementation changes" \
  --workspace-root /path/to/llm-delegator \
  --file src/llm_delegator/providers/base.py \
  --json

The CLI defaults to automatic model routing. It also accepts a provider model alias or full model ID for direct adapter testing, but low-complexity work is always forced to Flash. Pass --provider openrouter to use GLM 5.3 or GLM 5.3 Flash through OpenRouter. These models require reasoning and accept only low, high, or max; the adapter maps none to low and medium to high. Reasoning effort defaults to high for MCP, CLI, and programmatic requests.

Configuration

Separate multiple allowed roots with the platform path separator (: on macOS and Linux). Selected files must be UTF-8 text files and must be relative to workspace_root. Symlinks cannot escape the selected workspace.

Development

uv run ruff format --check .
uv run ruff check .
uv run pytest