ticky / devouswitz

/ticky/named agents over MCP

>AI APIs, provider CLIs and local models >relay tasks, run parallel reviews, combine results >one Python package, no runtime dependencies

your MCP harness
  ├─ ask_wren  → Codex
  ├─ ask_rook  → Claude
  ├─ ask_sage  → local API
  └─ ticky_team → coordinate

/setup/macOS · Linux · Windows

1. Prerequisites

Python 3.11+ and Git. Connect an API endpoint or install a provider CLI:

Codex Claude Code Gemini CLI Grok Ollama

2. Get ticky

git clone https://github.com/devouswitz/ticky.git
cd ticky
./ticky setup

Windows PowerShell

git clone https://github.com/devouswitz/ticky.git
cd ticky
py -3 ticky setup

Choose a connection in setup

Connection	Setup
`api`	Protocol, endpoint, model ID and optional key. Supports OpenAI Chat, Responses, Anthropic Messages, Gemini and compatible local or hosted endpoints.
Provider CLI	Use an existing login or a separate account.
`custom` API	Map a JSON request and response for another service.
`command`	Connect an executable or SDK bridge. Prompt on stdin, answer on stdout.

Direct APIs return text. Custom commands control their own tools and require full access. Provider billing and limits apply.

3. Open a session

./ticky ui

py -3 ticky ui

Connect a harness

./ticky install codex

py -3 ticky install codex

Restart the selected harness after installing. For another MCP client, run ./ticky mcp-json.

/session/commands

Command	Action
`@name task`	Send a task to one agent.
`/team name,name task`	Pass a task and earlier results through agents from any configured provider.
`/paste name`	Compose a multiline task. Finish with `/send` or discard with `/cancel`.
`/use name`	Pin an agent for plain text tasks.
`/agents`	Show models, effort and access.
`/model name model effort`	Change an agent's model and thinking effort.
`/new`	Clear conversation context.
`/doctor`	Check MCP dispatch without spending model credits.

For parallel reviews and a final synthesis: ticky team reviewer,researcher "Review this design" --mode parallel --lead editor. Use your roster's names. Parallel contributors must be read-only.