Tekt — The Utility Belt for Your AI Harness

Easy AI at the edge. Tekt sets up an AI sandbox on your Linux or Windows machine so agents can do real work for you. It is an open-source utility belt for AI engineers: a hand-curated catalog of tools, agents, MCP servers and skills. It brings the best of open source and SaaS, already picked and vetted, so you don't have to choose. One script installs it on infrastructure you own, and you bring your own intelligence. Builders get a workbench ready to go. Operators get a stack they can run and share. Managers get a clear view of what the agents are doing. Sovereignty · Security · Privacy.

Browse every shelf at tekt.md/catalog, or read the source of truth: tekt.catalog.yaml.

flowchart LR
  Bootstrap["Bootstrap<br/>install.sh / install.ps1"] --> Dev["Tekt.Dev<br/>Git · Go · Python · Node · VS Code · Docker"]
  Dev --> Base["Tekt.Base<br/>rclone · AWS CLI · MinIO"]
  Base --> Edge["Tekt.Edge<br/>Tailscale · ngrok"]
  Edge --> Iris["Tekt.Iris<br/>Ollama · Agent runtimes · MCP servers"]
  Iris --> Cloud["Tekt.Cloud<br/>MCPHub · LibreChat · n8n · Sovrant"]

Tekt is a catalog, not a product: every entry is someone else's excellent work, credited explicitly. What Tekt adds is the curation (what earns a place on a shelf), the composition (the arkitype), and a bootstrap that takes you from a blank machine to a working edge/cloud stack in one sitting.

Who it's for

Why it's different

Tekt isn't an agent, a model, or a platform, and it doesn't want your prompts or your data. It's the utility belt around them. It helps you manage the four things every AI harness has, gets you started, and keeps different agents in sync through shared communications and knowledge tools. Every tool is pre-vetted and runs on your own infrastructure.

Bring your own intelligence

Tekt ships no model and no keys. Point your agents at whichever provider you choose, by API or through the provider's own client:

Wire any of them into LibreChat, n8n, or an agent runtime's config. Tekt only connects them.

Share with your people

Git is too technical, and S3 is too much work. Your people already use Google Drive, OneDrive, Dropbox, Box, Nextcloud or a shared folder, so a Tekt Space uses that. It's a folder on your computer that stays in sync with the storage you pick, laid out the same way for everyone: docs/ for documents, knowledge/ for notes worth keeping, and skills/ for AI skills.

tekt space add team drive     # sign in to Google Drive in your browser, and that's it
tekt space autosync on        # sync every 10 minutes in the background
tekt connect                  # let your AI apps (Claude Code, Claude Desktop, Codex) use it

Share the Tekt/team folder from Google Drive the way you share anything, then run tekt space invite team. It writes the invitation (how to accept the folder, install Tekt and join) and copies it to your clipboard. From then on you both, and both your AIs, work from the same files. Skills travel the same way. Drop a skill folder into skills/, or run tekt skill new team summarize, and after the next sync it shows up in everyone's Claude Code. Full guide: tekt.md/spaces.

Roadmap

Where the belt is headed: a modular framework of providers, knowledge, agents, communication and cloud. You plug in the models, then orchestrate, build and empower your own workflows. The iteration-by-iteration plan is in .manage/2026.09.10.iterations.plan.md.

Quick Start

Curl (macOS / Linux / WSL2):

curl -fsSL https://tekt.md/install.sh | bash

PowerShell (Windows):

irm https://tekt.md/install.ps1 | iex

The PowerShell installer refreshes PATH as it goes, so you should not need to pause and restart between steps. If a newly installed command still is not visible after the script finishes, open a new PowerShell window and run .\install.ps1 status.

Git:

git clone https://github.com/xingh/tekt.md
cd tekt.md
bash install.sh          # or: .\install.ps1

VirtualBox (pre-built VM):

An OEM Linux Mint distribution by Rahul Singh with most prerequisites pre-installed. It boots in VirtualBox (or any compliant VM system) and walks you through setup like a brand-new computer.

• Install Oracle VirtualBox
• Download the TEKT VM here
• After initial setup:

cd ~/Tools/tekt.md
git pull
bash install.sh

Supported platforms: macOS (Intel + Apple Silicon), Ubuntu/Debian, Fedora/RHEL, Arch Linux, WSL2, Windows (PowerShell).

After the install

tekt space add team drive    # share a folder with your AI and your people (Drive, OneDrive, Dropbox…)
tekt space autosync on       # keep your Spaces in sync every 10 minutes
tekt connect                 # let Claude Code, Claude Desktop and Codex use your Spaces
tekt space invite team       # write an invitation for your colleagues (copied to the clipboard)
tekt tool add memory team    # give your AI apps a memory that lives in the team Space
bash install.sh status       # what landed, layer by layer
bash install.sh mcp          # MCPHub + 4 curated MCP servers → :3000
bash install.sh ui           # LibreChat → :3080, n8n → :5678
bash install.sh share 3080   # HTTPS-share a UI (Tailscale Serve, else ngrok)
bash install.sh catalog      # print the tool catalog

The installer also puts a tekt command on your PATH, so the same subcommands read tekt status, tekt space add, tekt mcp, tekt ui, tekt share. Type tekt help to see them all.

The Documentation Path

Five documents take you from composition to a shared, working home-lab stack:

What Gets Installed

Tekt.Dev — Development Environment

Tekt.Base — Data, Storage & Sync

Tekt.Edge — Network & Exposure

Tekt.Iris — Intelligence

OpenClaw onboarding is intentionally deferred during bootstrap to keep install non-interruptive: run openclaw onboard --install-daemon when ready.
On Windows, restart PowerShell after installer runs, then verify with .\install.ps1 status.

Tekt.Cloud — Chat, Workflows & Command Center

Scaffolded by the installer, started on demand (mcp / ui):

The curated MCP servers

bash install.sh mcp brings up MCPHub with four local servers — filesystem, fetch, memory, github — working in your instance workspace (~/Tekt/Instances/<host>/workspace) and exposed at http://localhost:3000/mcp. Four is the default, not the limit: swap or add servers in mcp_settings.json (hot-reloaded), and put HTTPS in front with bash install.sh share 3000. Details: 03 — Software.

Prerequisites

• macOS: Xcode Command Line Tools (xcode-select --install)
• Linux: curl, sudo access
• Windows: winget (App Installer); WSL2 recommended for full parity
• All: ~10 GB free disk, a reliable connection, and (optionally) API keys for hosted models — or none at all if you go all-local with Ollama

Customizing

Version pins and repo URLs live in tekt.catalog.yaml — edit it next to the script and both installers pick up the pins. To skip a tool, comment out its call in main(). The catalog is also the honest ledger of upstream creators and licenses; if you redistribute Tekt, keep it intact.

Architecture Reference

Tekt/
├── Global/
│   └── Workspaces/              ← synced from S3 via rclone / s5cmd
└── Instances/
    └── <hostname>/
        ├── workspace/           ← what local agents & MCP servers see
        ├── mcp/                 ← MCPHub config + compose
        ├── cloud/               ← LibreChat / n8n / MinIO / Postgres stacks
        ├── agents/              ← NanoClaw and other source-staged agents
        └── sovrant/             ← command center (.NET 10 source build)

Full topology, archetypes, and signal flow: 00 — Arkitype.

Troubleshooting

The most common fix for everything: open a fresh terminal (PATH caching). The Windows PowerShell installer refreshes PATH during the run, so this is usually only needed after the script exits. Per-tool troubleshooting now lives in the doc for its layer — 01, 02, 03, 04. Still stuck? bash install.sh status tells you exactly what's missing, and every install function prints its manual fallback commands when it fails.

Contributing

Issues and PRs welcome at github.com/xingh/tekt.md.

Credits

Tekt exists because of the people who build the tools it composes. Every entry in tekt.catalog.yaml carries an upstream: field naming its creators — that field is load-bearing. Particular thanks to the maintainers of Git, Homebrew, Go, Python/pyenv, nvm/Node, VS Code, Docker, rclone, the AWS/S3 tool authors, Tailscale, ngrok, Ollama, Anthropic (Claude Code & the MCP reference servers), the OpenClaw community, Sipeed (PicoClaw), Nous Research (Hermes), zeroclaw-labs, HKUDS (Nanobot), qwibitai (NanoClaw), samanhappy (MCPHub), Danny Avila (LibreChat), and n8n GmbH.

Maintained by Anant Corporation