ntwine

AI agents that code together.

multiple models. one codebase. real-time collaboration with presence, attribution, and shared tools.

what is this

ntwine is a collaborative AI IDE where multiple AI models work on the same codebase simultaneously. they can see each other's cursors, call each other out on bad code, and spawn sub-agents when they need help.

the harness

the core of ntwine is its tool execution harness. 2,500+ lines of Go incorporating patterns from 20+ products:

tools:

• read_file / edit_file / write_file - file operations with read-first enforcement
• search_code / list_files - codebase navigation
• shell - command execution with dangerous pattern blocking
• think - forced reflection before critical decisions
• update_notes / pin_message - shared spec and pinning
• create_memory / recall_memory - persistent context across sessions
• use_skill - load markdown skill files
• web_search - tavily web search
• mcp:*:* - any tool from any connected MCP server

harness features:

• parallel tool execution (models call multiple tools at once)
• observation masking (old tool results replaced with [masked] to save context)
• loop detection with escalation (warn -> force text -> abort)
• git worktree isolation for parallel agents
• append-only event log with pub/sub
• hooks system (pre/post tool execution)
• project config loading (.ntwine/config.md, AGENTS.md, CLAUDE.md)
• dangerous command blocking (rm -rf /, fork bombs, etc)
• environment variable sanitization

setup

go build ./cmd/server
./server --port 8080

mcp config

create .ntwine/mcp.json or .mcp.json in your project:

{
  "servers": [
    {
      "name": "filesystem",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
    }
  ]
}

project config

create .ntwine/config.md in your project with rules for the agents:

this is a go project using gin for the api layer.
always run `go test ./...` after making changes.
prefer table-driven tests.

skills

drop markdown files in .ntwine/skills/ or ~/.ntwine/skills/:

---
name: my-skill
description: does a thing
---

instructions for the agent when this skill is loaded...

license

MIT