Agent Injector

MCP server that spawns Claude Code sub-agents powered by alternative Anthropic-compatible models — MiniMax, GLM (Z.AI), and any other provider that implements the Anthropic Messages API.

Instead of using Haiku/Sonnet as sub-agents, delegate tasks to models like MiniMax M2.5 or GLM-5 that offer competitive quality at lower cost, while retaining Claude Code's full agentic toolset (Bash, file I/O, MCP servers, skills).

How It Works

Claude Desktop / Claude Code (main session)
    │
    │  MCP tool call
    ▼
Agent Injector (this MCP server)
    │
    │  spawns claude -p ... with swapped env vars
    ▼
Claude Code (headless)  →  MiniMax / GLM / ...
    │                        (Anthropic-compatible endpoint)
    │
    ▼
Full agentic loop: Bash, Read, Edit, Write, Glob, Grep, MCP servers...

MiniMax and Z.AI (GLM models) expose native Anthropic-compatible endpoints — no gateway or translation layer needed. Agent Injector simply sets ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY to point Claude Code at the alternative provider.

Tools

Tool	Description
llm_run	Run a task synchronously (blocks until done). Best for tasks under ~2 min.
llm_start	Start a background task, get a task_id for polling.
llm_poll	Check status of a running task.
llm_result	Get the full result of a completed task.
llm_cancel	Kill a running task.
llm_batch_start	Start multiple tasks in parallel, get a batch_id.
llm_batch_poll	Check status of all tasks in a batch at once.
llm_list_tasks	List all active and recent tasks.
readme	Get full usage instructions for sub-agents.

Installation

Requires Python 3.11+, uv, and Claude Code CLI installed.

Claude Desktop

Add to your Claude Desktop config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "agent-injector": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/moon-strider/agent-injector",
        "agent-injector"
      ],
      "env": {
        "MINIMAX_API_KEY": "your-minimax-api-key",
        "ZAI_API_KEY": "your-zai-api-key",
        "DEFAULT_MODEL": "MiniMax-M2.5"
      }
    }
  }
}

Set only the API keys for the providers you use. At least one is required.

Environment Variables

Providers

Variable	Required	Default	Description
MINIMAX_API_KEY	No*	—	MiniMax API key
MINIMAX_BASE_URL	No	https://api.minimax.io/anthropic	MiniMax Anthropic-compatible endpoint
MINIMAX_MODEL	No	MiniMax-M2.5	Default model for MiniMax provider
ZAI_API_KEY	No*	—	Z.AI API key (from z.ai)
ZAI_BASE_URL	No	https://api.z.ai/api/anthropic	Z.AI Anthropic-compatible endpoint
ZAI_MODEL	No	glm-5	Default model for Z.AI provider
DEFAULT_MODEL	No	First provider's default	Override default model; also determines which provider is used

*At least one provider API key must be set.

Server

Variable	Required	Default	Description
XDG_STATE_HOME	No	~/.local/state	Base directory for log files

Usage Instructions for Claude

"Call the agent-injector:readme tool to get full usage instructions"

Supported Providers

Any provider with an Anthropic Messages API compatible endpoint works. Tested:

Provider	Base URL	Default Model	Web Search
MiniMax	https://api.minimax.io/anthropic	MiniMax-M2.5	minimax-coding-plan-mcp (stdio MCP)
Z.AI	https://api.z.ai/api/anthropic	glm-5	web_search_prime (HTTP MCP)

Each provider's web search MCP is automatically configured for sub-agents when the provider's API key is set.

Important: subscription plans required. Both MiniMax and Z.AI Anthropic-compatible endpoints are only available through their respective subscription plans — they do not work with pay-as-you-go API credits. Specifically:

• MiniMax — requires an active MiniMax subscription. The Anthropic-compatible endpoint (api.minimax.io/anthropic) is not available with standard API credits.
• Z.AI (GLM) — requires a GLM Coding Plan subscription (Lite, Pro, or Max). The Anthropic-compatible endpoint (api.z.ai/api/anthropic) does not accept regular API credit keys.

If you use a regular API key without an active subscription, requests to the Anthropic-compatible endpoints will hang indefinitely instead of returning an error.

Architecture

Agent Injector spawns Claude Code in headless mode (claude -p) with environment variables that redirect all model tiers (opus, sonnet, haiku) to the configured provider:

ANTHROPIC_BASE_URL=<provider endpoint>
ANTHROPIC_API_KEY=<provider api key>
ANTHROPIC_MODEL=<model name>
ANTHROPIC_DEFAULT_OPUS_MODEL=<model name>
ANTHROPIC_DEFAULT_SONNET_MODEL=<model name>
ANTHROPIC_DEFAULT_HAIKU_MODEL=<model name>
DISABLE_PROMPT_CACHING=1

Built-in safeguards:

• Process timeouts — configurable per task (default 1200s), SIGTERM → SIGKILL escalation
• Concurrency limits — max 10 simultaneous tasks
• Automatic cleanup — dead tasks garbage-collected every 30s, results expire after 25 min
• Graceful shutdown — all child processes terminated on server exit
• File logging — daily logs in ~/.local/state/agent-injector/, auto-rotated after 7 days

License

MIT