Spec-driven development for AI coding assistants.

Follow @0xTab on X for updates · Join the OpenSpec Discord for help and questions.

_{🧪 New: OPSX Workflow — schema-driven, hackable, fluid. Iterate on workflows without code changes.}

OpenSpec aligns humans and AI coding assistants with spec-driven development so you agree on what to build before any code is written. No API keys required.

AI coding assistants are powerful but unpredictable when requirements live in chat history. OpenSpec adds a lightweight specification workflow that locks intent before implementation, giving you deterministic, reviewable outputs.

Key outcomes:

• Human and AI stakeholders agree on specs before work begins.
• Structured change folders (proposals, tasks, and spec updates) keep scope explicit and auditable.
• Shared visibility into what's proposed, active, or archived.
• Works with the AI tools you already use: custom slash commands where supported, context rules everywhere else.

• Lightweight: simple workflow, no API keys, minimal setup.
• Brownfield-first: works great beyond 0→1. OpenSpec separates the source of truth from proposals: openspec/specs/ (current truth) and openspec/changes/ (proposed updates). This keeps diffs explicit and manageable across features.
• Change tracking: proposals, tasks, and spec deltas live together; archiving merges the approved updates back into specs.
• Compared to spec-kit & Kiro: those shine for brand-new features (0→1). OpenSpec also excels when modifying existing behavior (1→n), especially when updates span multiple specs.

See the full comparison in How OpenSpec Compares.

┌────────────────────┐
│ Draft Change       │
│ Proposal           │
└────────┬───────────┘
         │ share intent with your AI
         ▼
┌────────────────────┐
│ Review & Align     │
│ (edit specs/tasks) │◀──── feedback loop ──────┐
└────────┬───────────┘                          │
         │ approved plan                        │
         ▼                                      │
┌────────────────────┐                          │
│ Implement Tasks    │──────────────────────────┘
│ (AI writes code)   │
└────────┬───────────┘
         │ ship the change
         ▼
┌────────────────────┐
│ Archive & Update   │
│ Specs (source)     │
└────────────────────┘

1. Draft a change proposal that captures the spec updates you want.
2. Review the proposal with your AI assistant until everyone agrees.
3. Implement tasks that reference the agreed specs.
4. Archive the change to merge the approved updates back into the source-of-truth specs.

• Node.js >= 20.19.0 - Check your version with node --version

Option A: Using npm

npm install -g @fission-ai/openspec@latest

Verify installation:

openspec --version

Option B: Using Nix (NixOS and Nix package manager)

Run OpenSpec directly without installation:

nix run github:Fission-AI/OpenSpec -- init

Or install to your profile:

nix profile install github:Fission-AI/OpenSpec

Or add to your development environment in flake.nix:

{
  inputs = {
    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
    openspec.url = "github:Fission-AI/OpenSpec";
  };

  outputs = { nixpkgs, openspec, ... }: {
    devShells.x86_64-linux.default = nixpkgs.legacyPackages.x86_64-linux.mkShell {
      buildInputs = [ openspec.packages.x86_64-linux.default ];
    };
  };
}

Verify installation:

openspec --version

Navigate to your project directory:

cd my-project

Run the initialization:

openspec init

What happens during initialization:

• You'll be prompted to pick any natively supported AI tools (Claude Code, CodeBuddy, Cursor, OpenCode, Qoder,etc.); other assistants always rely on the shared AGENTS.md stub
• OpenSpec automatically configures slash commands for the tools you choose and always writes a managed AGENTS.md hand-off at the project root
• A new openspec/ directory structure is created in your project

After setup:

• Primary AI tools can trigger /openspec workflows without additional configuration
• Run openspec list to verify the setup and view any active changes
• If your coding assistant doesn't surface the new slash commands right away, restart it. Slash commands are loaded at startup, so a fresh launch ensures they appear

After openspec init completes, you'll receive a suggested prompt to help populate your project context:

Populate your project context:
"Please read openspec/project.md and help me fill it out with details about my project, tech stack, and conventions"

Use openspec/project.md to define project-level conventions, standards, architectural patterns, and other guidelines that should be followed across all changes.

Here's a real example showing the complete OpenSpec workflow. This works with any AI tool. Those with native slash commands will recognize the shortcuts automatically.

Start by asking your AI to create a change proposal:

You: Create an OpenSpec change proposal for adding profile search filters by role and team
     (Shortcut for tools with slash commands: /openspec:proposal Add profile search filters)

AI:  I'll create an OpenSpec change proposal for profile filters.
     *Scaffolds openspec/changes/add-profile-filters/ with proposal.md, tasks.md, spec deltas.*

Check that the change was created correctly and review the proposal:

$ openspec list                             # Confirm the change folder exists
$ openspec validate add-profile-filters     # Validate spec formatting
$ openspec show add-profile-filters         # Review proposal, tasks, and spec delta

Iterate on the specifications until they match your needs:

You: Can you add acceptance criteria for the role and team filters?

AI:  I'll update the spec delta with scenarios for role and team filters.
     *Edits openspec/changes/add-profile-filters/specs/profile/spec.md and tasks.md.*

Once specs look good, start implementation:

You: The specs look good. Let's implement this change.
     (Shortcut for tools with slash commands: /openspec:apply add-profile-filters)

AI:  I'll work through the tasks in the add-profile-filters change.
     *Implements tasks from openspec/changes/add-profile-filters/tasks.md*
     *Marks tasks complete: Task 1.1 ✓, Task 1.2 ✓, Task 2.1 ✓...*

After implementation is complete, archive the change:

AI:  All tasks are complete. The implementation is ready.

You: Please archive the change
     (Shortcut for tools with slash commands: /openspec:archive add-profile-filters)

AI:  I'll archive the add-profile-filters change.
    *Runs: openspec archive add-profile-filters --yes*
     ✓ Change archived successfully. Specs updated. Ready for the next feature!

Or run the command yourself in terminal:

$ openspec archive add-profile-filters --yes  # Archive the completed change without prompts

Note: Tools with native slash commands (Claude Code, CodeBuddy, Cursor, Codex, Qoder, RooCode) can use the shortcuts shown. All other tools work with natural language requests to "create an OpenSpec proposal", "apply the OpenSpec change", or "archive the change".

openspec list               # View active change folders
openspec view               # Interactive dashboard of specs and changes
openspec show <change>      # Display change details (proposal, tasks, spec updates)
openspec validate <change>  # Check spec formatting and structure
openspec archive <change> [--yes|-y]   # Move a completed change into archive/ (non-interactive with --yes)

When you ask your AI assistant to "add two-factor authentication", it creates:

openspec/
├── specs/
│   └── auth/
│       └── spec.md           # Current auth spec (if exists)
└── changes/
    └── add-2fa/              # AI creates this entire structure
        ├── proposal.md       # Why and what changes
        ├── tasks.md          # Implementation checklist
        ├── design.md         # Technical decisions (optional)
        └── specs/
            └── auth/
                └── spec.md   # Delta showing additions

# Auth Specification

## Purpose
Authentication and session management.

## Requirements
### Requirement: User Authentication
The system SHALL issue a JWT on successful login.

#### Scenario: Valid credentials
- WHEN a user submits valid credentials
- THEN a JWT is returned

# Delta for Auth

## ADDED Requirements
### Requirement: Two-Factor Authentication
The system MUST require a second factor during login.

#### Scenario: OTP required
- WHEN a user submits valid credentials
- THEN an OTP challenge is required

## 1. Database Setup
- [ ] 1.1 Add OTP secret column to users table
- [ ] 1.2 Create OTP verification logs table

## 2. Backend Implementation  
- [ ] 2.1 Add OTP generation endpoint
- [ ] 2.2 Modify login flow to require OTP
- [ ] 2.3 Add OTP verification endpoint

## 3. Frontend Updates
- [ ] 3.1 Create OTP input component
- [ ] 3.2 Update login flow UI

Important: You don't create these files manually. Your AI assistant generates them based on your requirements and the existing codebase.

Deltas are "patches" that show how specs change:

• ## ADDED Requirements - New capabilities
• ## MODIFIED Requirements - Changed behavior (include complete updated text)
• ## REMOVED Requirements - Deprecated features

Format requirements:

• Use ### Requirement: <name> for headers
• Every requirement needs at least one #### Scenario: block
• Use SHALL/MUST in requirement text

OpenSpec’s two-folder model (openspec/specs/ for the current truth, openspec/changes/ for proposed updates) keeps state and diffs separate. This scales when you modify existing features or touch multiple specs. spec-kit is strong for greenfield/0→1 but provides less structure for cross-spec updates and evolving features.

OpenSpec groups every change for a feature in one folder (openspec/changes/feature-name/), making it easy to track related specs, tasks, and designs together. Kiro spreads updates across multiple spec folders, which can make feature tracking harder.

Without specs, AI coding assistants generate code from vague prompts, often missing requirements or adding unwanted features. OpenSpec brings predictability by agreeing on the desired behavior before any code is written.

1. Initialize OpenSpec – Run openspec init in your repo.
2. Start with new features – Ask your AI to capture upcoming work as change proposals.
3. Grow incrementally – Each change archives into living specs that document your system.
4. Stay flexible – Different teammates can use Claude Code, CodeBuddy, Cursor, or any AGENTS.md-compatible tool while sharing the same specs.

Run openspec update whenever someone switches tools so your agents pick up the latest instructions and slash-command bindings.

1. Upgrade the package

   npm install -g @fission-ai/openspec@latest
   

2. Refresh agent instructions
   • Run openspec update inside each project to regenerate AI guidance and ensure the latest slash commands are active.

You can always go back:

  proposal ──→ specs ──→ design ──→ tasks ──→ implement
     ▲           ▲          ▲                    │
     └───────────┴──────────┴────────────────────┘

• Install dependencies: pnpm install
• Build: pnpm run build
• Test: pnpm test
• Develop CLI locally: pnpm run dev or pnpm run dev:cli
• Conventional commits (one-line): type(scope): subject

MIT