H·AI·K·U works with any MCP-compatible AI coding tool. The MCP server detects the harness at startup and adapts its behavior accordingly -- adjusting subagent spawning, skill invocation, quality gate enforcement, and instruction language to match each harness's capabilities.

Prerequisites

Node.js 22+ (required to run the MCP server)
git (H·AI·K·U uses git for persistence and branching)
Clone the repository or download a release:

git clone https://github.com/gigsmart/haiku-method.git

Per-Harness Setup

Each harness has its own configuration location and format. In all cases, update /path/to/haiku-method to the actual path where you cloned or extracted the repository.

Cursor

Config file: .cursor/mcp.json (in your project root)

{
  "mcpServers": {
    "haiku": {
      "command": "/path/to/haiku-method/plugin/bin/haiku",
      "args": ["mcp", "--harness", "cursor"]
    }
  }
}

Notes:

Cursor has a ~40 tool limit. H·AI·K·U automatically removes browser UI tools to stay within this budget.
Subagents are supported with parallel spawning.
MCP prompts are available but do not surface as native slash commands -- invoke them through the prompt picker.

Windsurf

Config file: ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "haiku": {
      "command": "/path/to/haiku-method/plugin/bin/haiku",
      "args": ["mcp", "--harness", "windsurf"]
    }
  }
}

Notes:

No subagent support. Units execute sequentially rather than in parallel.
~100 tool limit.
No MCP elicitation support. Review gates are FSM-enforced rather than interactive.

Gemini CLI

Config file: ~/.gemini/settings.json

{
  "mcpServers": {
    "haiku": {
      "command": "/path/to/haiku-method/plugin/bin/haiku",
      "args": ["mcp", "--harness", "gemini-cli"]
    }
  }
}

Notes:

MCP prompts surface as slash commands (/haiku:start, /haiku:pickup).
Experimental subagent support via @subagent with parallel spawning, but no isolation.
Model tiers map differently: H·AI·K·U's haiku tier maps to flash, and both sonnet and opus map to pro.

OpenCode

Config file: Global config (YAML format)

mcpServers:
  haiku:
    command: /path/to/haiku-method/plugin/bin/haiku
    args:
      - mcp
      - --harness
      - opencode

Notes:

MCP prompts support is still maturing. Prompts are not yet exposed as slash commands.
Subagents execute sequentially (no parallel spawning).
No elicitation support. Review gates are FSM-enforced.

Kiro

Config file: .kiro/agents/*.yaml or the Settings UI

mcpServers:
  haiku:
    command: /path/to/haiku-method/plugin/bin/haiku
    args:
      - mcp
      - --harness
      - kiro

Notes:

Hook system supported (like Claude Code). Quality gates can be hook-enforced.
MCP prompts surface as slash commands.
Subagents supported with parallel spawning and isolation.
MCP elicitation supported for interactive review gates.

Environment Variable Alternative

Instead of passing --harness as a CLI argument, you can set the HAIKU_HARNESS environment variable in your MCP server configuration's env block:

{
  "mcpServers": {
    "haiku": {
      "command": "/path/to/haiku-method/plugin/bin/haiku",
      "args": ["mcp"],
      "env": {
        "HAIKU_HARNESS": "cursor"
      }
    }
  }
}

Valid values: claude-code, cursor, windsurf, gemini-cli, opencode, kiro.

Feature Comparison

Feature	Claude Code	Cursor	Windsurf	Gemini CLI	OpenCode	Kiro
Skills as slash commands	Native	Via prompts	Via prompts	Slash commands	Limited	Slash commands
Parallel subagents	Yes	Yes	No	Experimental	No	Yes
Hook system	Yes	No	No	No	No	Yes
Quality gates	Hook-based	FSM-enforced	FSM-enforced	FSM-enforced	FSM-enforced	Hook-based
FSM tamper detection	Hook-based	Checksum	Checksum	Checksum	Checksum	Hook-based
Browser review UI	Yes	No	No	No	No	No
MCP elicitation	Yes	Yes	No	No	No	Yes

Getting Started

After configuring your harness:

Open your project in the harness.
Use the haiku:status prompt to check for any active work in the project.
Use haiku:start to begin a new intent -- the elaboration flow will guide you through defining what you want to build.

Known Limitations

When running H·AI·K·U outside Claude Code, the following limitations apply:

No auto-context injection on session start. Claude Code injects H·AI·K·U context automatically via hooks. Other harnesses must use the haiku:status prompt at the start of each session to load active intent state.
No automatic output tracking. The agent must explicitly register outputs via haiku_unit_set. Claude Code's hook system handles this transparently.
No context exhaustion warnings. Claude Code detects when the context window is running low and triggers a graceful checkpoint. Other harnesses do not provide this signal.
Browser-based review UI unavailable. The haiku-review-server serves a local web UI for reviewing stage gates. In other harnesses, review gates use MCP elicitation where available, or fall back to FSM-enforced advancement.