Cross-Harness Export

Haira agents can be exported to other agent harnesses. The --target claude-code flag generates Claude Code configuration files and an MCP server binary for custom tools.

Usage

haira build main.haira --target claude-code

Output Structure

.claude/
  agents/agent-name.md       # Agent system prompt + tools + config
  commands/workflow-name.md   # Slash commands for workflows
  mcp-servers.json            # Points to MCP binary
bin/
  haira-tools                 # MCP server binary with custom tools

Agent Files

Each agent declaration generates a .claude/agents/<name>.md file containing:

• Agent name and model
• System prompt
• Tool references (via MCP)
• Handoff targets

agent CodeReviewer {
    provider: openai
    system: "Review code for bugs, security issues, and style."
    tools: [lint_code, check_security]
}

Generates .claude/agents/code-reviewer.md:

# CodeReviewer

Model: gpt-4o

Review code for bugs, security issues, and style.

## Tools

- lint_code (via MCP: haira-tools)
- check_security (via MCP: haira-tools)

Workflow Commands

Each workflow generates a .claude/commands/<name>.md slash command:

@post("/api/deploy")
workflow Deploy(service: string, env: string) -> { status: string } {
    step "Validate" { /* ... */ }
    step "Build" { /* ... */ }
    step "Deploy" { /* ... */ }
}

Generates .claude/commands/deploy.md:

# /deploy

## Parameters

- `service` (string)
- `env` (string)

## Steps

1. **Validate**
2. **Build**
3. **Deploy**

MCP Server

Custom tools are compiled into a standalone MCP server binary (bin/haira-tools) that communicates via stdio. The .claude/mcp-servers.json file points Claude Code to this binary:

{
  "haira-tools": {
    "command": "/absolute/path/to/bin/haira-tools",
    "args": []
  }
}

Complete Example

import "http"

provider openai {
    api_key: env("OPENAI_API_KEY")
    model: "gpt-4o"
}

tool analyze_code(code: string) -> string {
    """Analyze code for potential issues and improvements"""
    resp, err = Reviewer.ask("Review this code:\n${code}")
    if err != nil { return "Analysis failed." }
    return resp
}

agent Reviewer {
    provider: openai
    system: "You are an expert code reviewer."
    tools: [analyze_code]
}

@post("/api/review")
workflow Review(code: string) -> { feedback: string } {
    feedback, err = Reviewer.ask(code)
    if err != nil { return { feedback: "Error." } }
    return { feedback: feedback }
}

haira build review.haira --target claude-code
# Exported Claude Code config:
#   1 agent(s) → .claude/agents/
#   1 command(s) → .claude/commands/
#   1 tool(s) → bin/haira-tools (MCP)

Next Steps

• Tools — define tools for export
• Agents — define agents with handoffs
• Evaluation — test agents before exporting