I built a long-horizon AI harness that doesn't live in the chat

The product is the crane A long-horizon harness you run. Not a plugin pack inside someone else’s. Most things branded “harness engineering” are skills, agents, and slash commands that sit inside Claude Code or Copilot. That is a wrapper around a host you do not own. The host still holds context, tools, and whether the trajectory finishes. ai-gantry is the runtime: one Distroless Go process, one persona, one OpenAI-compat model, memory you can sqlite3, chat that dials out. Gantree is the shipping yard those cranes live in — a board for an operator, never a hop in a chat turn. gantry The rigid frame that holds and positions tools. The frame does nothing by itself; the tools and the memory do the work. gantree The shipping yard those cranes live in. Files and Docker. Pull, don’t punch. Stand up the yard The harness GitHub The difference A harness is a process. A marketplace is a costume. What this stack is A binary you run. Aims, personality, and work still exist on Thursday. The model is a socket — Ollama, Gemini, ChatGPT, Grok. Swap it. The loop stays. MCP listed in mcp.toml is the grant. Omit it and it does not exist. Outbound chat (Telegram, Discord, Slack). No open ports. Ever. Independent lookups fan out in one round; the next round uses the results. Garmin → sheet → Strava is one thread. What it is not A Claude Code or Copilot plugin marketplace with “harness” in the README. Prompts, skills, and sub-agents that still live in someone else’s context window. A chatbot that dies when the window gets expensive, then /new lobotomizes it. A dashboard or gateway in the token path so every message pays a platform tax. MkDocs around a wrapper. Documentation is not a runtime. The crane ai-gantry holds the person. The yard does not talk. One process. One persona. Chat stays Telegram. Health is an exit code, not a port. Mid-chain does not die on a typo — the loop was hardened on the 7–12B class, then pointed at Flash when you want the trajectory now. static binary + persona + mcp.toml + any OpenAI-compat LLM → outbound chat Want a bot tonight? Stay here. Open the board, build a crane — this yard pulls shotah/ai-gantry, writes the files, recreates. You do not need the other checkout to stand one up. The yard Operate named pets. Never sit in the turn. Gantree is the operator plane: board, doctor, grants, logs, spend. It reads Docker and files after the fact. If a chart would make the harness slower, the chart is wrong. Agents open zero inbound ports. Only this UI is reachable, and only through a path you chose. browser → gantree (localhost | Tailscale | tunnel) gantry gantry gantry Hello Docker on the same Linux host. Node 22. First boot is /setup. git clone https://github.com/shotah/gantree.git cd gantree cp gantree.toml.example gantree.toml npm install npm run build npm start # http://127.0.0.1:3000 Open the board. Build a crane. Grant a tool. Message it on your phone. The UI never sits in a chat turn. Home Mini vs cloud VM: install. Operator docs The board, the door, the files. InstallHome Mini or cloud VM. Same console. Zero inbound agent ports. ConsoleBoard, grants, doctor, avatars. Chat stays Telegram. OperatorsSetup, login, profile, roles on a handful of people. HeadlessAttach existing dirs, tunnels, the gotchas. ArchitectureWhy Vinext on the Docker host, and why the yard pulls. SecurityThe door: bind, passphrases, sessions. Not a reason to open WAN 3000. Custom MCPYour binary, still a [[server]] in that crane’s mcp.toml. ai-gantryDesign, features, MCP — the harness repo.