Wick

WickBuild internal tools and AI agents in Go.

Scaffold a tool, job, or connector with one prompt — Claude writes real Go files in your repo. Ship the same binary as a Slack / Telegram / web AI agent with command-approval, multi-session pool, and zero copy-pasting.

Paste this into any AI agent to get started:

Check if Go is installed by running: go version
If Go is not installed, help me install it first for my OS, then come back to this.

Once Go is ready:
1. Install wick: go install github.com/yogasw/wick@v0.10.2
2. Ask me: what do you want to name your project?
3. Scaffold it: wick init <name>
4. Run: wick dev — then show me what was created.

💬

AI Agents in Slack, Telegram, and the Web

Spawn Claude / Codex / Gemini as long-lived subprocesses — same agent reachable from Slack threads, Telegram chats, and the web UI at the same time.

Multi-session pool with idle-kill + --resume revive · multi-instance providers (two PATs, side-by-side) · workspaces on disk · command gate with 4-mode interactive approval · AskUser MCP tool · everything persisted under ~/.<app>/agents/.

🤖

AI Is the Primary User

Wick is designed for AI agents, not humans. Every convention, file name, and pattern is optimized so Claude knows exactly what to create — no exploration, no guessing.

🗂️

Git Is the Control Plane

No drag-and-drop UI to version. Every tool and job AI creates is real code in real files. `git diff` to review, `git revert` to undo. You own everything.

🧰

Tools, Jobs, & Connectors

Say "add a Slack notifier job" or "add a GitHub connector for our LLM agent". Claude creates the file, registers it, wires the config — for humans, schedulers, and LLMs alike.

🔌

LLM-Ready via MCP

Expose any connector to Claude, Cursor, and other MCP clients. Built-in OAuth 2.1 + Personal Access Tokens, per-call audit log, no protocol code on your side.

🛡️

Command Gate

Every Bash command an agent runs goes through a sidecar binary you ship with the installer. Whitelist via glob, escalate to interactive approval (Approve once / This session / Always / Block), audited per-stage to JSONL.

No env vars. Sibling-of-exe → embedded fallback → PATH. Just works.

👀

See Everything That Was Built

Git history IS your tool inventory. Who built what, when, and why — no separate dashboard or admin panel to maintain.

🔐

Secure by Default

First boot generates a random admin passphrase and forces a setup flow before anything else runs. SSO, per-tool visibility, tag-based access — all editable from /admin without a redeploy.

⚙️

Live Config, No Redeploy

Declare a typed Config struct. Fields become admin-editable rows. Secrets, URLs, toggles — updated live without touching code.

Why teams pick wick for AI agents

Most "AI agent" platforms lock you into their runtime, expose chat-only, and hide the moving parts. Wick does the opposite:

You bring	Wick gives you
Your Claude / Codex / Gemini install (with your MCP servers, skills, memory)	A pool that spawns them as subprocesses, idle-kills, resumes by `cli_session_id`
Your Slack workspace, Telegram bot, or just a browser	Three transports → one session per thread / chat / conversation, all live at once
Your PAT (or two — `claude/work` + `claude/personal`)	Multi-instance provider config, per-instance env / args, status-cached `--version` probes
Your folder of repos (or any path on disk)	Workspaces — managed or custom path, multi-session sharing, no git worktree
Your trust threshold	Command Gate sidecar: whitelist + 4-mode interactive approval, fail-safe block on infra failure

You bring

Wick gives you

Your Claude / Codex / Gemini install (with your MCP servers, skills, memory)

A pool that spawns them as subprocesses, idle-kills, resumes by cli_session_id

Your Slack workspace, Telegram bot, or just a browser

Three transports → one session per thread / chat / conversation, all live at once

Your PAT (or two — claude/work + claude/personal)

Multi-instance provider config, per-instance env / args, status-cached --version probes

Your folder of repos (or any path on disk)

Workspaces — managed or custom path, multi-session sharing, no git worktree

Your trust threshold

Command Gate sidecar: whitelist + 4-mode interactive approval, fail-safe block on infra failure

Read AI Agents for the headline tour, or jump to the deep-dives:

Workspaces — folders on disk, managed vs custom path, the built-in default

Providers — multi-instance config, binary resolution chain, status cache

Channels — Slack (Socket + HTTP), Telegram (long-poll), Web (SSE), AskUser MCP

Pool & Sessions — slot allocation, message buffer, resume flow

Command Gate — <app>-gate sidecar, shared spec/socket/audit, daily tail log