Any Model Plugin MCP server

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.

Why

Stop being locked into one AI provider. Most agent tools hardwire you to a single model vendor. Any Model Plugin is the universal remote for AI agents — delegate tasks to any engine, route through any provider, and run adversarial reviews where a different model challenges the implementation before you ship. Vendor-agnostic by design, zero dependencies, one core.

---

Table of contents

• Features
• Quick start
• Usage
  • Claude Code plugin
  • Universal CLI
  • MCP server
• Commands
• Engines
• Providers
• Configuration
• Documentation
• Roadmap
• Contributing
• Attribution
• Security
• License

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 — zero dependencies.

One-line install (Claude Code):

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

1 · 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

2 · Universal CLI (no install)

node plugins/anymodel/scripts/companion.mjs setup

3 · MCP server (Cursor, Windsurf, VS Code, Codex, Claude)

Add to your host's MCP config:

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

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

/anymodel:setup
/anymodel:models

First delegation

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

AI agents: every turn command accepts --json for machine-readable output. Use setup --json to check readiness, models --json to list available models, and delegate --json to capture results programmatically. The MCP server exposes the same surface as 8 tools — see INTEGRATIONS.md for per-host wiring.

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

Command	Description	Key flags
delegate	Delegate a task to an engine	--engine, --model, --write, --background, --resume
review	Run a code review	--engine, --model, --base, --scope
adversarial-review	Challenge implementation choices	--engine, --model, --base, --scope, focus text
choose	Interactive action/provider/model chain	--engine, --write
delegate-with	Delegate with guided model selection	--engine, --bridge, --write
review-with	Adversarial review with guided model selection	--engine, --base, --scope
models	Probe provider model lists	--provider
setup	Check engine/provider readiness	--json
status	Show active/recent jobs	--all, --json
result	Show finished job output	--json
cancel	Cancel an active job	--json

Engines

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

Engine	Harness	Sandbox	Native review	Resume	Requires
codex	codex app-server JSON-RPC	OS-level sandbox	Yes (review/start)	Yes	npm install -g @openai/codex
claude	claude -p --output-format json	Permission modes	No (schema fallback)	No	claude CLI
direct	Built-in agent loop (fetch → chat API)	Tool whitelist	No (schema fallback)	No	Nothing — zero deps

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:

Provider	Models	Env key	Quirks
zai	GLM-5.2, GLM-4.6	ZAI_API_KEY	function-tools-only, flat-assistant-content, strict-tool-adjacency
ollama	qwen3-coder, gpt-oss, minimax	OLLAMA_API_KEY	—
opencode-go	glm, kimi, qwen, deepseek	OPENCODE_API_KEY	function-tools-only, flat-assistant-content, strict-tool-adjacency, empty-choices-chunks

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:

Variable	Purpose
ZAI_API_KEY	Z.AI provider key
OLLAMA_API_KEY	Ollama Cloud provider key
OPENCODE_API_KEY	OpenCode Go provider key
ANYMODEL_ENV_FILE	Path to a .env-style file to load
ANYMODEL_BRIDGE	Bridge mode: builtin (default) or litellm
CODEX_HOME	Override the Codex CLI config directory (default: ~/.codex)

Documentation

Document	Covers
ARCHITECTURE.md	Design rationale, engine adapter interface, provider layer, config resolution, testing strategy
INTEGRATIONS.md	Per-host setup: CLI, MCP server, Claude Code native plugin
CONTRIBUTING.md	Development setup, adding providers and engines, code style, testing
CHANGELOG.md	Release history and notable changes
Project site	Landing page — features, surfaces, engines, providers
SECURITY.md	API key handling, sandboxing per engine, bridge isolation, vulnerability reporting

Roadmap

Phase	Status	Goal
Phase 0	Done	Scaffold repo; lift core orchestration; codex adapter + LiteLLM bridge; provider registry; loopback CI
Phase 1	Done	Codex + Claude adapters; per-thread config override; native + schema reviews on all engines
Phase 2	Done	Built-in Responses→Chat shim; quirk pipeline; unit test suite
Phase 3	Done	Plugin command surface + runner agent; models/setup; direct engine; marketplace install
Phase 4	In progress	Semantic /anymodel:* chains for guided provider/model selection

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
npm test   # hermetic: no API keys or network needed

Live-fire tests against real providers are opt-in: ANYMODEL_LIVE_TESTS=1 npm test (spends tokens).

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.