zing-ai 24.3.5

Zing AI development pipeline installer

Zing!

Zing plans, builds, tests, and ships clean code with AI.
Zing keeps you in charge at each step.
This is the Zen of Zing.

---

The Problem

AI is good at a lot of things. Reducing entropy isn't one of them.

Every AI-generated line of code is a small bet. Sometimes it's exactly right. Often it's close enough. But over hundreds and thousands of those bets, the misses accumulate. Variable names drift. Abstractions get duplicated. Edge cases get papered over. The codebase collapses because it decays.

AI coding assistants are great at writing functions. But shipping software isn't just writing functions. It's understanding what to build, figuring out where it fits in the codebase, breaking work into steps that make sense, building it incrementally, reviewing it properly, and tracking it in your project management tool. Today, you do all of that orchestration yourself, or you don't, and entropy wins.

The Landscape

There are a lot of really good projects tackling AI-assisted development right now, such as GSD, BMAD, Taskmaster AI, Spec Kit, and OpenSpec.

These are all worth checking out. But there's a fundamental philosophical difference between Zing and every other option: they all stop at "plan and execute." They help you structure work for an AI agent, then trust the agent to deliver. The human is involved at the beginning (writing the spec) and at the end (reading the PR), but the middle is a black box.

Why Zing is Different

• It doesn't trust the middle.

  Every stage boundary is a decision point where a human stays in control — not just before the work starts, but during planning, between plan and build, and after the code is written. Decision points are asynchronous and batched so they don't slow you down.

• It's an entropy reduction engine.

  You make the big decisions. The AI handles implementation within the boundaries you set. This isn't vibe coding. It's structured coding with AI as the executor, you as the architect.

• A real-time review dashboard.

  Parallel agents stream findings to a live dashboard in your browser via SSE. You triage findings, pick fix approaches, and answer planning questions in minutes instead of one at a time in the terminal. The AI blocks until you submit, then picks up exactly where it left off.

• The review loop.

  Six specialized review agents catch inconsistencies, flag drift from the plan, and surface entropy before it gets committed. Every review pass actively removes disorder from the system.

• It's a pipeline, not a prompt.

  Each stage feeds into the next. Four parallel audit passes stress-test the plan before code is written. The build follows the audited plan exactly. The code review knows what was intended because it has the spec.

• Parallelism is a first-class concept.

  Codebase exploration, plan evaluation, and code review all fan out across multiple specialized agents simultaneously. Better results because each agent has a focused lens.

• Context isolation per build step.

  Each build step runs in a fresh subagent with only the information it needs, producing one atomic commit per step with verified git hygiene.

• Chain of thought, not chain of hope.

  Each stage narrows the problem space before the next one starts. No deviations from the plan, no bonus features, no drive-by refactoring. Every step has acceptance criteria. Every completion gets a commit.

How they compare

	Zing	GSD	BMAD	Taskmaster AI	Spec Kit	OpenSpec
Capture requirements	✅	☑️	☑️	Via PRD	☑️	☑️
Plan	✅	☑️	☑️	☑️	☑️	☑️
Multi-pass plan audit	✅	—	Partial	—	—	—
Context isolation	✅	☑️	—	☑️	—	—
Multi-agent code review	✅	—	Partial	—	—	—
PR creation & response	✅	—	—	—	—	—
Live review dashboard	✅	—	—	—	—	—
Multi-model support	—	☑️	—	—	—	—
Task dependency graphs	—	—	—	☑️	—	—
Agent-agnostic	—	Partial	—	☑️	☑️	☑️
Runtimes supported	Claude Code, OpenCode	Claude Code, OpenCode, Gemini CLI, Codex	Claude Code, Cursor	13+ IDEs	Agent-agnostic	20+ tools
Issue tracker integration	Linear	—	—	—	GitHub	—

What Zing Does

Zing is a pipeline of specialized AI tools backed by a live review dashboard:

1. Capture (/zing)

Zing starts with a conversation or a Linear ticket URL. It listens, asks the right questions, and saves a structured spec to your .zing/ directory. No templates to fill out. Just talk about what you want to build.

2. Plan (/zing-plan)

Zing explores your codebase in parallel. Multiple agents fan out simultaneously to map relevant files, understand existing patterns, and identify integration points. It asks you targeted questions, then produces a concrete action plan with acceptance criteria for every step.

3. Audit the Plan (/zing-plan-audit)

Before a single line of code is written, parallel evaluation passes stress-test your plan:

• Design Fundamentals: Is this the right approach? Is it overengineered?
• Robustness & Safety: Will it break things? Is it testable?
• Executable Spec: Is every step specific enough to actually build?
• Code Quality: Does it follow the idioms of the codebase?

Each dimension gets a rating, and weak spots come with concrete improvement options.

4. Track (/zing-plan-linear)

Zing creates a Linear project with tickets for each phase, sequential dependencies between them, and the full plan attached as a document. Your project management stays in sync without you touching it.

5. Build (/zing-build)

Zing executes the plan step by step. After each step, it commits, updates the progress checklist, and moves on. No scope creep, no unsolicited refactoring, no features that weren't in the plan. Just disciplined, incremental delivery against the spec.

6. Review (/zing-build-audit)

Four parallel review agents examine your branch's changes like senior developers:

• Correctness: Logic errors, edge cases, error handling
• Security & Reliability: Vulnerabilities, production readiness
• Quality & Style: Naming, readability, idiomatic code
• Coverage & Performance: Test gaps, bottleneck risks

Findings stream into a live review dashboard in your browser. You can watch agents work in real time — spinners show which agents are running, notification dots flag new findings on each tab, and build logs are available for diagnostics. When agents finish, you triage each finding directly in the browser: accept, drop, downgrade, or discuss. Choose from suggested fix approaches or write your own. Text responses and selections auto-save as you go. When you're ready, hit Submit and Zing picks up where it left off — applying fixes, writing a review report, and moving to the next stage.

7. Ship

When the review is clean, Zing offers to open your pull request. Draft by default, because you're still in control.

8. PR Respond (/zing-pr-respond)

Given a PR link or number, Zing checks out the branch, fetches all unresolved review comments, and walks you through each one. It analyzes whether the comment needs a code fix, a reply, or is already addressed. When there's one clear fix it proposes the change; when there are multiple approaches it asks you to pick. After all comments are handled, it commits and pushes the changes, posts replies on GitHub, and resolves the threads. Then it polls CI with increasing backoff until all checks complete. If any checks fail, Zing investigates the logs, fixes the failures, pushes again, and loops back to check for any new unresolved comments — repeating the full cycle until CI is green. Finally, it re-requests reviews from stale reviewers so the PR is ready for another look.

9. PR Review (/zing-pr-audit)

Once a PR is open, Zing can review it the way a senior developer would — on GitHub, with line-level comments. It checks out the PR branch, reads every changed file in full, fans out four parallel review agents (the same ones from the build review), and walks you through each finding before submitting. The final review is posted via the GitHub API with inline comments on the exact lines that matter, severity ratings, and code suggestions where the fix is obvious. The review action (approve, comment, or request changes) is your call. A local markdown report is also saved so you can feed findings straight back into /zing-plan if fixes are needed.

10. Code Audit (/zing-custom-audit)

Point Zing at any area of your codebase — files, directories, or just a description like "the authentication module" — and it performs a focused audit. Zing resolves your description to concrete files, confirms the scope with you, then fans out six parallel review agents to analyze the code as it stands today. Each finding is walked through one by one so you can validate, discuss, or dismiss it. Confirmed findings are written to a markdown report you can feed into /zing-plan to start fixing them.

---

Installation

Prerequisites

• Python >= 3.12
• uv

Recommended MCP Servers

Zing works best when your AI coding assistant has the following MCP servers installed and configured:

• Serena — Semantic code tools via LSP for token-efficient code exploration and precise symbol-level editing
• AI Distiller — Compact code structure extraction and specialized analysis (security audits, bug hunting, refactoring)
• CodeGraphContext — Code graph analysis for understanding call chains, detecting dead code, and architectural queries
• Context7 — Up-to-date library documentation and code examples, so your assistant doesn't rely on stale training data

It also benefits from having the GitHub CLI (gh) installed for creating pull requests, managing issues, and interacting with GitHub directly from the command line.

These servers give Zing's agents deeper insight into your codebase during planning, building, and reviewing.

Start the dashboard

Before using any review command, start the Zing server in a separate terminal:

zing-ai mcp

This launches the MCP server and review dashboard on http://127.0.0.1:9876. The dashboard is available at http://127.0.0.1:9876/dashboard. Leave this running while you work — review commands won't function without it.

The MCP connection is automatically registered with Claude Code during installation (via mcp-remote), so Claude Code will connect to the running server when review tools are invoked.

Permissions

The review commands use the MCP server and curl to coordinate agent workflows. To avoid repeated permission prompts, add the following to your global Claude Code settings (~/.claude/settings.json) under permissions.allow:

"Bash(curl:*)",
"mcp__zing-ai__*"

Install from PyPI

uv tool install zing-ai

Install from GitHub

Bleeding edge (latest features, may have rough edges):

uv tool install --force git+https://github.com/Farmer-Pete/Zing

Set up commands for your AI coding assistant

Interactive mode (asks which runtime to install for):

zing-ai install

Claude Code:

zing-ai install --claude

OpenCode:

zing-ai install --opencode

Both:

zing-ai install --all

---

Updating

From PyPI:

uv tool upgrade zing-ai
zing-ai install --claude   # or --opencode or --all

From GitHub (reinstall to pull latest):

uv tool install --force git+https://github.com/Farmer-Pete/Zing
zing-ai install --claude   # or --opencode or --all

If you've customized any commands, they'll be backed up to a zing-patches/ directory before being overwritten. To see your backed-up customizations:

zing-ai reapply-patches --claude   # or --opencode

---

Usage

After installation, the zing commands are available as slash commands in your AI coding assistant:

• /zing — Start a new zing (capture what you want to build)
• /zing:plan — Break it down into an actionable plan
• /zing:plan-audit — Audit the plan for soundness
• /zing:build — Execute the plan step by step
• /zing:build-audit — Review the code changes
• /zing:custom-audit — Audit existing code for issues
• /zing:pr-respond — Address unresolved PR review comments
• /zing:pr-audit — Review a pull request on GitHub
• /zing:plan-linear — Create Linear tickets from the plan

In OpenCode, use flat naming: /zing-plan, /zing-build, etc.

---

Architecture

Primary Pipeline

The main flow chains each stage into the next automatically:

flowchart LR
    zing["/zing"] --> new["/zing:new"]
    new --> plan["/zing:plan"]
    plan --> planaudit["/zing:plan-audit"]
    plan -.->|"optional"| planlinear["/zing:plan-linear"]
    planaudit --> build["/zing:build"]
    build --> buildaudit["/zing:build-audit"]
    buildaudit -.->|"fix findings"| zing

All Command Flows

Commands can also be invoked independently. Audit commands optionally feed findings back into planning:

flowchart TD
    zing["/zing"] --> new["/zing:new"]
    new --> plan["/zing:plan"]
    plan --> planaudit["/zing:plan-audit"]
    planaudit --> build["/zing:build"]
    build --> buildaudit["/zing:build-audit"]

    plan --> planlinear["/zing:plan-linear"]
    planlinear -.->|"Start build"| build

    buildaudit -.->|"Fix findings"| plan
    customaudit["/zing:custom-audit"] -.->|"Fix findings"| plan
    praudit["/zing:pr-audit"] -.->|"Fix findings"| plan

    prrespond["/zing:pr-respond"]

    style zing fill:#4a9eff,color:#fff
    style prrespond fill:#6c757d,color:#fff
    style customaudit fill:#6c757d,color:#fff
    style praudit fill:#6c757d,color:#fff
    style planlinear fill:#6c757d,color:#fff

MCP Tool Lifecycle

The Zing MCP server tracks sessions and steps across the entire pipeline via a shared dashboard:

sequenceDiagram
    participant New as /zing:new
    participant MCP as Zing MCP Server
    participant Dashboard as Review Dashboard
    participant Plan as /zing:plan
    participant Audit as Audit Commands

    New->>MCP: session_create(title, steps)
    MCP-->>New: session_id + step IDs
    Note over New: Stores IDs in zing file frontmatter

    Plan->>MCP: session_update(session_id)
    Plan->>MCP: step_start(session_id, step_id)
    loop Subagents
        Plan->>MCP: agent_start(session_id, step_id, name)
        Plan->>MCP: step_log(session_id, step_id, name, message)
        Plan->>MCP: agent_stop(session_id, step_id, name)
    end

    Audit->>MCP: session_update(session_id)
    Audit->>MCP: step_start(session_id, step_id)
    loop Review Agents (×6 parallel)
        Audit->>MCP: agent_start(session_id, step_id, name)
        Audit->>MCP: finding_submit(session_id, step_id, finding)
        Audit->>MCP: agent_stop(session_id, step_id, name)
    end
    Audit->>MCP: review_wait(session_id, step_id)
    MCP-->>Dashboard: Stream findings via SSE
    Dashboard-->>MCP: User triage responses
    MCP-->>Audit: Review complete

Review Agent Architecture

All audit commands (build-audit, custom-audit, pr-audit) fan out six parallel review agents:

flowchart TD
    audit["Audit Command"] --> fan{"Fan out"}
    fan --> a1["Architecture<br/>& Design"]
    fan --> a2["Correctness<br/>& State"]
    fan --> a3["Security"]
    fan --> a4["UI &<br/>Readability"]
    fan --> a5["Performance"]
    fan --> a6["Testing &<br/>Observability"]

    a1 --> dedup["Deduplicate<br/>(by type + title)"]
    a2 --> dedup
    a3 --> dedup
    a4 --> dedup
    a5 --> dedup
    a6 --> dedup

    dedup --> submit["finding_submit() → Dashboard"]
    submit --> wait["review_wait()"]
    wait --> user["User triages in browser"]

    style audit fill:#4a9eff,color:#fff
    style user fill:#28a745,color:#fff

Session Continuity

The zing file frontmatter is the glue that connects all pipeline stages to a single session:

flowchart LR
    subgraph "Zing File (.zing/my-feature.md)"
        fm["session: abc123<br/>steps:<br/>  plan: step-1<br/>  plan-audit: step-2<br/>  build: step-3<br/>  build-audit: step-4"]
    end

    fm -->|"reads session + step IDs"| plan["/zing:plan<br/>step_start(step-1)"]
    fm -->|"reads session + step IDs"| pa["/zing:plan-audit<br/>step_start(step-2)"]
    fm -->|"reads session + step IDs"| build["/zing:build<br/>step_start(step-3)"]
    fm -->|"reads session + step IDs"| ba["/zing:build-audit<br/>step_start(step-4)"]