Skip to content

Neovate Code

★ New
trial
AI / ML open-source MIT free

At a Glance

Open-source CLI coding agent from Ant Group with a Vite-style plugin architecture, 30+ LLM providers, MCP integration, sub-agent orchestration, and headless mode.

Type
open-source
Pricing
free
License
MIT
Adoption fit
small, medium
Top alternatives

What It Does

Neovate Code (@neovate/code) is a TypeScript-based CLI coding agent that provides an interactive and headless agentic coding experience. It is developed by Ant Group engineers (the team behind UmiJS) under the neovateai GitHub organization, released as MIT open-source. It supports code generation, bug fixing, code review, test creation, and multi-step agentic workflows via a terminal UI (built with Ink, the React-for-terminal framework).

The defining differentiator is its Vite-style plugin system — a lifecycle hook architecture that lets teams extend every aspect of the agent without forking the core: adding LLM providers, custom tools, slash commands, system prompt modifications, sub-agent definitions, telemetry hooks, and more.

Key Features

  • 30+ LLM provider support: Anthropic, OpenAI, Google Gemini, DeepSeek, Qwen, Moonshot/Kimi, ZhipuAI/GLM, MiniMax, SiliconFlow, VolcEngine, xAI/Grok, Groq, Cerebras, Nvidia, OpenRouter, HuggingFace, GitHub Copilot, and more — via the Vercel AI SDK
  • Vite-inspired plugin architecture: Lifecycle hooks across initialization, workflow, session, and agent phases; hook execution models: First, Series, SeriesLast, SeriesMerge, Parallel
  • MCP client: Supports stdio, SSE, and HTTP transports; per-server config with timeout and header support
  • Headless / quiet mode: --quiet flag and TTY-aware shell execution for CI/CD and automation pipelines
  • Sub-agent orchestration: Plugin-registered sub-agents with per-agent model configuration; subagentStop telemetry hook
  • Session management: Persistent sessions with resume, branch (/branch), rename (/rename), export, and copy (/copy)
  • Built-in tool set: read, write, edit, bash (conditional), ls, glob, grep, fetch, todo
  • VS Code extension: Directory present in repo; beta status
  • Biome + Vitest: Modern TypeScript toolchain for code quality and testing

Use Cases

  • Provider-flexible agentic coding: Teams that cannot or will not use a single provider (cost optimization, Chinese cloud constraints, model routing)
  • Internal coding agent extensions: Plugin system enables building team-specific tools, slash commands, and context injectors without forking
  • CI/CD automation: Headless mode makes it suitable for automated code review, generation, or refactoring pipelines
  • Multi-model experimentation: Easy provider/model switching to compare outputs across DeepSeek, Qwen, Claude, and GPT variants

Adoption Level Analysis

Small teams (<20 engineers): Strong fit for teams with multi-provider needs or those building internal extensions. MIT license and npm install -g simplicity reduce friction. Documentation is sparse in English; expect to read source code.

Medium orgs (20-200 engineers): Good fit if a team wants to standardize on a customizable open-source agent. The plugin system enables centralized internal extensions distributed as npm packages. MCP integration connects to internal tools. Gap: no centralized config management or policy enforcement layer.

Enterprise (200+): Not yet ready. No security audit, no published responsible disclosure policy, sparse governance documentation, and unclear long-term Ant Group commitment. Vendor lock-in risk is low (MIT + multi-provider), but operational maturity risk is high.

Alternatives

AlternativeKey DifferencePrefer when…
Claude CodeAnthropic-only, proprietary, strong memory systemYou want the most capable single-provider agent with Auto-Dream memory
Codex CLIOpenAI-only, MIT, minimal surface areaYou want a minimal open-source OpenAI agent with sandboxed execution
Gemini CLIGoogle-only, Apache-2, 1M context windowYou need very long context and are in the Google ecosystem
Goose (Block)MCP-native, AAIF governance, open-sourceYou want vendor-neutral, community-governed open-source
AiderGit-native, Python, strong architect/editor patternYou want open-source with proven git integration and model flexibility

Evidence & Sources

Notes & Caveats

  • Primary dev activity in Chinese: Most PR discussions and commit messages from core team are in Chinese. English speakers relying on community support will find it sparse.
  • Origin opacity: Open-sourced from an internal Ant Group project (umijs/takumi). No official Ant Group announcement or roadmap commitment published.
  • VS Code extension is beta: The vscode-extension/ directory exists but is not published to the VS Code Marketplace and has no independent changelog.
  • Bash timeout recently reduced: Default was 30 minutes, reduced to 2 minutes in v0.22.8 — suggests the tool was hanging in real-world use. Something to monitor.
  • No security policy: No SECURITY.md, responsible disclosure process, or published audit. Significant concern for a tool executing arbitrary bash in your local environment.
  • MCP implementation uses experimental_createMCPClient: The AI SDK MCP client is still experimental — API stability is not guaranteed.

Related