🚀 Vizier

Turn app ideas into structured, agent-ready build plans in seconds.

Vizier is a VS Code extension that takes a loose app idea and transforms it into a complete, structured blueprint — PRD, tech stack, data model, API contract, build tasks, and a decision register — then exports it to your workspace so you (or your AI coding agent) can build from it.

✨ Trusted by developers building with Claude, GPT-4, Ollama, and more.

Features

• Smart Classification — Automatically categorizes your idea (SaaS, Mobile, CLI tool, Browser extension, Game, Internal tool) using AI, with keyword-based fallback.
• Repo-Aware Planning — Scans your workspace (package files, README, existing agent rules) so plans respect your current stack instead of guessing greenfield. No source code is read or sent.
• Category-Aware Questionnaire — Asks different, relevant questions based on your app type, plus optional expert-perspective lenses.
• Complete Blueprint Pipeline — 6-stage generation: PRD → Architecture → Data Model → API Contract → Tasks → Decisions.
• API Contract Design — Concrete REST endpoints with method, path, auth, and request/response shapes.
• Estimates & Story Points — Every task gets an effort size, engineering hours, and Fibonacci story points.
• Dependency-Aware Task Graph — Build tasks ordered by dependencies with cycle detection, plus auto-generated testing tasks.
• Structured Validation — All LLM output is schema-validated with zod and automatically retried with feedback on failure.
• Context Packs — Per-task scoped context so agents get just what each task needs.
• Agent-Specific Export — Generates .cursorrules, CLAUDE.md, or AGENTS.md automatically based on what's installed in your workspace.
• Plan Progress Monitoring (private, local) — Track how much of the exported plan has been executed — done / in-progress / blocked, with git/test-backed verification and a live status-bar chip. Runs entirely on your machine.
• Multi-Model Support — Anthropic (Claude), any OpenAI-compatible API, omniroute (auto-switching gateway), or a local Ollama server (fully on-prem, no API key).
• Provider Resilience & Cost Control — Optional fallback provider (automatic failover), local response caching, per-stage model selection, a monthly token budget, and stable mode.
• Real Tracker Integrations — Push plan tasks to Jira, Linear, or GitHub Issues (or a generic webhook) as real issues, with dry-run preview.
• Privacy-First by Design — Planning sends only your idea + a workspace summary. Monitoring is 100% local. API keys live in VS Code Secret Storage.

Requirements

• VS Code 1.85.0 or higher
• A model provider. Pick one in settings (vizier.provider):
  • Anthropic (Claude) — get an API key, or
  • OpenAI-compatible API (OpenAI, OpenRouter, etc.), or
  • omniroute (an OpenAI-compatible gateway), or
  • Ollama — local, no API key required. Install Ollama and pull a model, e.g. ollama pull llama3.2.

Installation

1. Open VS Code
2. Open the Extensions view (Ctrl+Shift+X)
3. Search for "Vizier"
4. Click Install
5. Open Settings, set vizier.provider and the matching API key (or pick ollama for fully local use)

Quick Start

1. Open the Vizier sidebar (rocket icon in the activity bar)
2. Describe your app idea in 1–3 sentences, click Plan This App
3. Answer a few short questions (or skip them)
4. Wait ~15–30 seconds for your blueprint
5. Review the blueprint, tasks, and decisions
6. Export to Files — writes everything to a plan/ folder in your workspace
7. As you build, run Vizier: Check Plan Progress to see live execution status — fully locally

Tip: You can also start from the command palette with Vizier: Plan New App.

Configuration

Output

After generation, Vizier writes these files to your workspace:

plan/
  plan.json           - Machine-readable plan (source of truth for progress monitoring)
  overview.md         - Product requirements document
  architecture.md     - Tech stack with rationale
  schema.md           - Data model with entities
  api.md              - API contract (endpoints, auth, request/response)
  tasks.md            - Build tasks in dependency order (with estimates)
  decisions.md        - Architectural decision register
  perspectives.md     - Expert-perspective sections (if any selected)
  context/            - Per-task context packs
    TASK-001.md
    TASK-002.md
    ...
  status.md           - Plan progress report (written by "Check Plan Progress")
.cursorrules         - Cursor rules (if Cursor detected)
CLAUDE.md            - Claude Code rules (if Claude Code detected)
AGENTS.md            - Generic agent rules (fallback)

Every generated document includes a disclaimer footer. plan/status.md additionally states that monitoring is local and private.

Plan Progress Monitoring

The Vizier: Check Plan Progress command (status-bar chip, command palette, or the button in the sidebar) compares the exported plan against your workspace to report:

• Status per task — done / in-progress / blocked / not started, computed from file existence and task references
• Verification — tasks are marked verified only when backed by real evidence: git commits referencing the task, or tests that actually passed (parsed from Jest/JSON and JUnit/XML reports)
• Coverage & trend — project coverage and a progress history for burndown-style trends

All of this runs entirely on your machine. No source code, file contents, or repository data is ever transmitted.

Privacy & Disclaimers

• Planning: only your idea, questionnaire answers, and a summary of your workspace (package files, README, existing rules) are sent to your selected LLM provider. No source code is sent.
• Monitoring: runs 100% locally; it inspects file existence and task references only. No data leaves your machine.
• Keys: stored in VS Code Secret Storage, not plaintext.
• AI is not a substitute for judgment. Plans may be inaccurate, incomplete, or unsafe. Review, validate, and test everything. See DISCLAIMERS.md.

FAQ

Q: Which AI models are supported? A: Anthropic Claude, any OpenAI-compatible API, omniroute (OpenAI-compatible, can auto-switch models), or a local Ollama server (fully on-prem, no API key). Choose via vizier.provider.

Q: Is my code sent to the AI? A: During planning, no source code is sent — only your idea, answers, and a repo summary. Plan-progress monitoring is fully local and never transmits any code. With the Ollama provider, generation is also fully local.

Q: What does "omniroute" do? A: It's treated as an OpenAI-compatible endpoint where vizier.omnirouteModel: "auto" lets the gateway switch models (automode) while Vizier retains all planning context. Set vizier.omnirouteBaseUrl to your gateway.

Q: How is my API key stored? A: API keys are stored in VS Code Secret Storage. They are never written to plaintext settings.

Q: Does the status-bar chip upload anything? A: No. Plan-progress monitoring is local-only. If you enable the optional AI narrative, it is generated from plan metadata only (task ids, titles, status counts) — never from your code.

License

MIT — see LICENSE.md.