Any Model Plugin

Delegate tasks and run adversarial reviews from any AI agent to any provider and model.

Any Model Plugin separates the executor harness (the agent runtime that plans, runs commands, and edits files) from the model (the provider the harness thinks with). Pick the best engine for the job and the best model for the task — independently.

Table of contents

• Features

• Quick start

• Usage

  • Claude Code plugin

  • Universal CLI

  • MCP server

• Commands

• Engines

• Providers

• Configuration

• Documentation

• Roadmap

• Contributing

• Attribution

• Security

• License

Related MCP server: random-agent

Features

• Multi-engine delegation — run tasks through Codex, Claude, or the built-in direct agent loop

• Multi-provider routing — use Z.AI GLM, Ollama Cloud, OpenCode Go, or add your own via registry.toml

• Adversarial reviews — challenge your implementation with a different model before shipping

• Background jobs — delegate long-running tasks and check status / result / cancel

• Three access surfaces — Claude Code slash commands, universal CLI, or MCP server (Cursor, Windsurf, VS Code, Codex)

• Zero dependencies — pure ESM, Node.js standard library only

• Built-in wire shim — translates the engine's Responses API to any provider's chat-completions, with a quirk pipeline for provider-specific fixes

Quick start

Prerequisites

• Node.js ≥ 18.18 — node --version

• No npm install needed — the plugin has zero npm dependencies

Install as a Claude Code plugin

git clone https://github.com/pealmeida/anymodel-plugin.git
cd anymodel-plugin
claude plugin marketplace add .
claude plugin install anymodel@any-model

Set up provider API keys

export ZAI_API_KEY="your-key"
export OLLAMA_API_KEY="your-key"
export OPENCODE_API_KEY="your-key"

Or load from a file:

export ANYMODEL_ENV_FILE="$HOME/.anymodel.env"

Verify your setup

/anymodel:setup
/anymodel:models

Delegate your first task

/anymodel:delegate --engine direct --model zai/glm-5.2 "explain the architecture of this project"

Usage

Claude Code plugin

Once installed, 11 slash commands are available:

/anymodel:delegate           Delegate a task to any engine and model
/anymodel:review             Run a code review against local git state
/anymodel:adversarial-review Challenge the implementation approach and design choices
/anymodel:choose             Interactive chain: pick action, provider, model, then dispatch
/anymodel:delegate-with      Delegate with guided provider/model selection
/anymodel:review-with        Adversarial review with guided provider/model selection
/anymodel:models             Probe configured providers for available models
/anymodel:setup              Check engine and provider setup
/anymodel:status             Show active and recent jobs
/anymodel:result             Show a finished job's output
/anymodel:cancel             Cancel an active background job

Delegate a task:

/anymodel:delegate --engine codex --model zai/glm-5.2 --write "fix the flaky auth test"
/anymodel:delegate --engine claude "refactor the error handling in src/api"
/anymodel:delegate --engine direct --model ollama/qwen3-coder "add unit tests for the parser"

Run a review:

/anymodel:review --base main
/anymodel:review --engine claude --scope working-tree
/anymodel:adversarial-review "check for race conditions in the connection pool"

Guided selection chains:

/anymodel:choose delegate with zai fix the flaky test
/anymodel:delegate-with ollama implement the parser guard
/anymodel:review-with opencode-go --base main challenge the architecture

Universal CLI

The same CLI powers every surface. Run it from any shell, CI pipeline, or agent:

COMPANION="plugins/anymodel/scripts/companion.mjs"

# Delegate a task
node "$COMPANION" delegate --engine direct --model zai/glm-5.2 "explain this codebase"

# Run a review
node "$COMPANION" review --base main

# Adversarial review with focus text
node "$COMPANION" adversarial-review "audit auth and session handling"

# Check available models
node "$COMPANION" models

# Verify setup
node "$COMPANION" setup

# Job management
node "$COMPANION" status
node "$COMPANION" result <job-id>
node "$COMPANION" cancel <job-id>

All turn commands accept --json for machine-readable output.

MCP server

Expose all capabilities as MCP tools in any MCP-compatible host:

{
  "mcpServers": {
    "anymodel": {
      "command": "node",
      "args": ["/absolute/path/to/anymodel-plugin/plugins/anymodel/scripts/mcp-server.mjs"]
    }
  }
}

Supported hosts: Claude Code, Codex CLI, Cursor, Windsurf, VS Code. See INTEGRATIONS.md for per-host configuration details.

The MCP server exposes 8 tools: delegate, review, adversarial_review, status, result, cancel, models, setup.

Commands

Engines

The executor harness — the agent runtime that plans, runs commands, and edits files.

The direct engine is the "any model, no harness installed" floor. It implements a minimal agent loop with read_file, list_dir, write_file, and exec_command tools, speaking OpenAI-compatible chat directly to any registry provider. Sandbox is weakest — use --write deliberately.

Providers

The model — what the harness thinks with. Configured declaratively in registry.toml:

Use a provider by prefixing the model: zai/glm-5.2, ollama/qwen3-coder, opencode-go/kimi.

Adding a provider

Edit plugins/anymodel/scripts/lib/providers/registry.toml:

[providers.my-provider]
base_url = "https://api.example.com/v1"
env_key  = "MY_PROVIDER_API_KEY"
wire     = "chat"
quirks   = []

Set the env var, then verify with /anymodel:models. See CONTRIBUTING.md for details on quirk flags.

Configuration

Config is resolved in order of precedence:

CLI flags  >  env vars  >  <repo>/.anymodel.toml  >  ~/.config/anymodel/config.toml  >  defaults

Environment variables:

Documentation

Roadmap

Contributing

Contributions are welcome. The project has zero npm dependencies and uses Node's built-in test runner — clone and go.

git clone https://github.com/pealmeida/anymodel-plugin.git
cd anymodel-plugin
node --test tests/*.test.mjs   # 125 tests, 0 failures

See CONTRIBUTING.md for guidelines on adding providers, engines, and tests.

Attribution

This project derives from openai/codex-plugin-cc under the Apache-2.0 license. The orchestration core under plugins/anymodel/scripts/lib/core/ began as a fork of that project and has been generalized to support any engine and provider. See LICENSE and NOTICE for details.

Security

• API keys are never stored in this repository. Providers reference keys by env var name in registry.toml. .env files are gitignored.

• Sandboxing varies by engine. codex has an OS-level sandbox; claude uses permission modes; direct has tool-level guards only. Never default --write on engines without real sandboxing.

• The bridge listens on 127.0.0.1 only — not exposed to the network.

• Report vulnerabilities privately via GitHub Security Advisories rather than public issues.

See SECURITY.md for the full policy.

License

Apache-2.0 — see LICENSE and NOTICE.