AI Agent via MCP

Any MCP-compatible agent connects to a gibil server through the MCP protocol. The MCP server exposes lifecycle tools, VM tools, and background job tools — giving your agent full control over remote servers. Gibil uses standard MCP — any compatible client works.

Setup

Run gibil init — it configures the MCP server automatically. For Claude Code, it writes to ~/.claude/.mcp.json. For other agents, add the MCP config manually:

{
  "mcpServers": {
    "gibil": {
      "command": "gibil",
      "args": ["mcp"]
    }
  }
}

Restart your agent to activate. It now has these tools:

What your agent can do

Your agent manages the full lifecycle — forge, work, burn — without leaving MCP:

• Forge a server — create_server({ name: "task-1", repo: "...", ttl: 30 })
• Explore the repo — vm_ls, vm_grep, vm_read to understand the codebase
• Write code — vm_write to create or modify source files
• Run anything — vm_bash to install packages, run tests, build, deploy, git push
• Run long tasks in background — vm_bash({ command: "pnpm build && pnpm test", background: true }) returns a job ID
• Poll for results — vm_job_status({ job_id: "j-abc123" }) returns exit code + output when done
• Iterate — read test output, fix code, run again — the code-test loop with typed tools
• Burn when done — destroy_server({ name: "task-1" })

Why MCP over CLI mode

Both work. MCP is better when the agent needs to read and write files frequently:

• Typed tools — vm_write({ path: "src/app.ts", content: "..." }) instead of shell escaping
• Structured results — every tool returns parsed output with exit codes
• No quoting hell — the MCP server handles SSH internally

Next steps

• Use with AI Agents (full guide) — CLI mode vs MCP mode comparison
• Agent Code-Test Loop — the iteration pattern MCP enables
• AI Agent Sandbox — patterns for autonomous agents