Piyaz Docs

Install the Piyaz plugin against the hosted app and drive your first project from your agent.

Piyaz is hosted at app.piyaz.ai. The fastest path is to sign in to the hosted app, install the plugin into your coding agent, and connect. This page uses Claude Code; the other platforms have their own guides as well (see the connect guides).

Want to run the server yourself? See Run locally for the Postgres + Bun setup.

Install the plugin

Go to app.piyaz.ai and sign in. This is the account your agent connects to.

Add the marketplace and install

claude plugin marketplace add FrkAk/piyaz
claude plugin install piyaz@piyaz

Connect

In Claude Code, run /mcp, select piyaz, and complete the browser sign-in. OAuth runs once per machine.

Verify

Ask your agent:

/piyaz list my projects

Your agent returns your project list. An empty list is expected on a fresh account.

Now Claude Code has access to the 9 MCP tools, the /piyaz skill, the full workflow skill set, and the dispatchable agents. For Codex, Cursor, or Antigravity, see the connect guides.

Your first project

Load the skill with /piyaz, then say what you want in your own words. Once it is engaged you just keep talking, no prefix is needed again.

/piyaz I want to build a realtime dashboard for server metrics such as cpu/mem/requests

Piyaz brainstorms with you first. It restates the idea, asks one focused question at a time about the gaps (how fresh the data must be, who can see it), pushes back on weak choices, and proposes a default where you have none. You approve the brief. In the same flow it decomposes that brief into a task graph. Your project is now set up with a set of tasks already wired by their dependencies and relations.

MET-1  Add the metrics ingestion endpoint   draft
MET-2  Stream metrics over WebSocket         blocked by MET-1
MET-3  Build the live chart component        blocked by MET-2
MET-4  Add an auth guard to the dashboard    related to MET-3

This is a graph, not a flat checklist. The dependency edges decide which task can start and which task has to wait.

See tasks

Ask Piyaz to navigate the graph for you.

what is the next task?

It points at MET-1, the one task with nothing blocking it, and notes that it is still a draft, so it needs a plan before work starts. You did not have to go through a list of tasks, Piyaz found a starting point itself.

Ask about any task by name.

what is MET-3 about?

It returns the description, the acceptance criteria, and what MET-3 is waiting on. The graph is the source of truth, so the answer is always current.

Work on each task in its own fresh session, so the context window stays focused on one piece of work. Composer does this for you, running every task in a clean agent context.

Refine before you build

A draft is a starting point, not a contract. Refine it using plain language; you can also ask the agent to look things up before it settles the details.

lets refine MET-1, tell me more about rate-limit and also search docs to define better

It explains the rate-limit options, checks the relevant docs, then tightens the description and the acceptance criteria and writes the implementation plan. The more refined a task, the better every agent that picks it up performs.

Hand the loop to composer

On Claude Code, run the whole thing end to end.

/piyaz:composer

On the Pro plan, composer runs inside a Dynamic workflow that is off by default. Enable the Dynamic workflows row in /config first (Claude Code v2.1.154+). See the composer pipeline.

Composer picks the top ready task on the critical path, researches it against your codebase, writes the plan, implements it on a branch, and opens a pull request. It stops there for your review. You merge, and it records what shipped, propagates the result to unblock the next task, and picks that one up. See the composer pipeline.

Go further

Work on several ready tasks at once with parallel dispatch.
See the full feature graph in the web UI's Graph mode, or read the end-to-end cookbook.
The exact tool calls behind all of this are in the MCP reference, if you want to take a better look at them.

Next: Core concepts · MCP tools · Self-host

Last updated July 25, 2026

Quickstart