OpenSpec
Also known as: OpenSpec by Fission AI, Fission AI OpenSpec, @fission-ai/openspec, Fission-AI/OpenSpec, OpenSpec Desktop, openspec.dev
OpenSpec is a free, open source spec driven development framework from Fission AI that puts a lightweight layer of written requirements between developers and their AI coding agents. It installs slash commands into more than thirty assistants, including Claude Code, Codex, Cursor and GitHub Copilot, so each change gets a proposal, delta specs, a design and tasks that the developer reviews before the agent builds, verifies and archives the work into living specs kept in the repository.
OpenSpec is a free, open source spec driven development framework for AI coding agents, made by Fission AI, the company of founder Tabish Bidiwale, a former team lead at the quantum software company Q-CTRL, and backed by Y Combinator. Released under the MIT license as a command line tool, it has passed 75,000 GitHub stars, more than 265,000 developers use it each month, and a new spec is created with it about every two seconds. It adds a lightweight layer of written requirements that the developer and the agent agree on before any code is written, aimed at existing codebases as much as new ones.
OpenSpec does not ship a coding agent; it installs slash commands and skills into whichever of more than thirty assistants a team uses, including Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI and OpenCode. Each change gets its own folder with a proposal, delta specs written as requirements with WHEN and THEN scenarios, a design and a task list.
The agent explores the codebase, proposes the change, works through the tasks writing code and running tests, verifies the result against the artifacts with issues marked critical, warning or suggestion, and archives the change so its deltas merge into the living specs. Schemas define which artifacts exist and in what order, the CLI returns JSON so agents ask it for live state and instructions, validation runs in CI, and Stores, in beta, share read only requirements across repositories. The Desktop beta for macOS adds a review interface where developers accept or undo each requirement, track open questions and hand an approved change to Claude Code, Codex, GitHub Copilot or OpenCode to build.
Everything is free, with no paid plan, hosted service or enterprise edition, and the cost of running it is the model usage of the agent it sits inside. Specs live as Markdown in the customer's own repository, the CLI runs on macOS, Linux and Windows, and anonymous telemetry limited to command names and the version can be switched off. Team features for large codebases and multi repository systems are still being built.
It fits engineering teams that already use coding agents on mature codebases and want plans that outlast a single chat, survive a switch between agents and stay reviewable in pull requests, plus platform teams that own requirements other teams build against. It fits less well for anyone wanting an agent that plans and builds without review, teams needing access controls or audit trails beyond git, or developers who would rather not read and refine a spec before the code arrives.
Vendor details
Canonical URL
https://openspec.dev
Category
Coding agent
Subcategory
Open source spec driven development framework that runs inside other companies' coding agents, with a review desktop app in beta
Funding status
Fission AI, backed by Y Combinator; no funding figures used.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
OpenSpec installs commands and skills for Claude Code, Codex, Cursor, GitHub Copilot (including its cloud agent), Gemini CLI, OpenCode, Amazon Q, JetBrains Junie, Zed and more than twenty other assistants, runs in CI through JSON output and exit codes, files feedback as GitHub issues through the gh CLI, and keeps specs in git. The Desktop beta works with Claude Code, Codex, GitHub Copilot and OpenCode.
In practice
A tech lead asks Claude Code to propose an export feature on a large billing service. OpenSpec creates a change folder with a proposal, delta specs, a design and eight tasks, the lead edits two requirements, and only then does the agent apply the tasks.
A developer starts a change in Cursor and finishes it in Codex the next day. The task list shows which items are checked off and the specs carry the agreed scenarios, so the second agent picks up where the first stopped.
A platform team keeps authentication requirements in a shared store that product teams reference read only, and CI runs validation on every pull request so a change without spec deltas fails.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
8.0 / 14 capabilities · 57%
| Integrations & Tool Calling | Partial |
|---|---|
|
The init wizard installs slash commands and skills for more than thirty coding assistants, including Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Amazon Q, JetBrains Junie and Zed, and can deliver instructions to GitHub Copilot's cloud agent. Connections to issue trackers, chat tools or cloud services come from the hosted agents, since the framework adds no service connectors of its own. Sourceopenspec.devread 2026-11-18 |
|
| Workflow Orchestration | Full |
|
Each change moves through a defined lifecycle of explore, propose, apply and archive, and schemas set the artifacts and their dependencies, so the fast forward command builds proposal, specs, design and tasks in dependency order before the agent works through the task list. The status and instructions commands tell the agent which artifact is ready next, and the onboarding run chains eleven phases from finding an improvement to archiving it. Sourcegithub.com/fission-ai/openspec/blob/main/docs/commands.mdread 2026-11-18 |
|
| Knowledge Grounding & RAG | Partial |
|
Specs live in the repository as living requirements, with scenarios written as WHEN and THEN steps, so every agent session reads the same statement of what the system should do. Stores, in beta, keep shared requirements in a separate repository that other teams reference read only, and the context command assembles them for the agent. There is no retrieval index or embedding search, since agents read the Markdown files directly. Sourcegithub.com/fission-ai/openspecread 2026-11-18 |
|
| Human Oversight & Guardrails | Partial |
|
The propose step stops before any code is written so the developer reviews the plan, the update command confirms every edit one artifact at a time, the continue command lets the developer set the pace, and bulk archiving asks before overwriting specs. The Desktop beta adds accept and undo on each requirement and flags open questions. There are no allow lists or blocks over what the agent may run while it applies a change. Sourceopenspec.dev/blog/openspec-desktopread 2026-11-18 |
|
| Security, Identity & Governance | Partial |
|
The CLI runs locally with no account, API key or hosted service, and specs never leave the repository, while usage telemetry collects command names and the version only, switches off in CI and can be disabled with a setting or an environment variable. Stores let a platform team own requirements that other teams read only. There is no single sign on, role based access or audit trail of its own beyond git. Sourcegithub.com/fission-ai/openspec/blob/main/docs/cli.mdread 2026-11-18 |
|
| Observability & Auditability | Partial |
|
The list, show, status and validate commands report every change's artifacts and progress, most of them as JSON, and the view command opens a terminal dashboard of changes and specs. The dated archive keeps each change's proposal, design and tasks as a record of why the code changed. There are no traces or logs of agent runs, so what an agent actually did lives in its own transcript and in git history. Sourcegithub.com/fission-ai/openspec/blob/main/docs/cli.mdread 2026-11-18 |
|
| Memory & State Persistence | Partial |
|
Change folders, checked off items in the task list and the dated archive of completed changes persist plans and progress across sessions, so work can pause and resume or move from one coding agent to another without losing what was decided. That state lives in Markdown files in the repository rather than in an agent memory, and preferences or lessons learned across tasks stay with each hosted agent. Sourceopenspec.dev/faqread 2026-11-18 |
|
| Deployment & Data Residency | Full |
|
The CLI installs through npm, pnpm, yarn, bun, Homebrew, Nix or a standalone binary that needs no Node.js, and runs on macOS, Linux and Windows. Specs and changes are plain Markdown inside the customer's own repository, so nothing is hosted by the vendor. The Desktop beta runs on macOS with Windows support coming, and it needs the CLI set up in the project first. Sourcegithub.com/fission-ai/openspecread 2026-11-18 |
|
| Prebuilt Agents / Templates / Packs | Partial |
|
The default spec driven schema ships ready to use with templates for the proposal, specs, design and tasks, the expanded profile adds further workflow commands, and community schema bundles from separate repositories package opinionated workflows. There are no prebuilt agents, since the framework runs inside whichever coding agent the developer already uses, and the onboarding command walks a new user through one full change. Sourcegithub.com/fission-ai/openspecread 2026-11-18 |
|
| Triggers & Channel Coverage | Not documented |
|
Every step begins with a person typing a slash command or a script calling the CLI, and nothing wakes on a schedule, a webhook or a repository event. Continuous integration can run validation over every change, but that check runs inside a pipeline the team already owns, and the framework itself never starts an agent or a change on its own. Desktop chats also start only when the developer opens them. Sourcegithub.com/fission-ai/openspec/blob/main/docs/cli.mdread 2026-11-18 |
|
| Model Flexibility & Routing | Full |
|
The framework works with any of more than thirty coding assistants, so the choice of model follows the agent: Claude models through Claude Code, OpenAI models through Codex, Gemini through Gemini CLI, or any provider through OpenCode, and specs carry over when a team switches agents. The Desktop beta runs Claude Code, Codex, GitHub Copilot and OpenCode with the model chosen for each chat. Sourceopenspec.dev/faqread 2026-11-18 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
The CLI creates, validates and archives changes and returns JSON from list, show, validate, status, instructions, templates, schemas and context, with clear exit codes for scripts and CI, and the instructions command exists so agents fetch live guidance rather than static prompts. Custom schemas are created, forked and validated from the CLI to define new artifacts and workflows. There is no MCP server or programmatic library. Sourcegithub.com/fission-ai/openspec/blob/main/docs/cli.mdread 2026-11-18 |
|
| Testing, Debugging & Optimization | Partial |
|
The verify command checks an implementation against its artifacts for completeness, correctness and coherence, searches the codebase for evidence and rates issues as critical, warning or suggestion, while validate checks specs and changes in strict mode and in CI. The verify step sits in the expanded profile and does not block archiving, and there is no scoring or comparison of agents across tasks. Sourcegithub.com/fission-ai/openspec/blob/main/docs/commands.mdread 2026-11-18 |
|
| Browser / Computer-use | Not documented |
|
Nothing in the framework drives a browser or a desktop. Agents implement tasks in the repository through their own tools, and any browser testing comes from the coding agent or the team's test suite rather than from the spec layer, which only plans, tracks and checks changes against written requirements and scenarios. Verification reads code and tests, not rendered pages. Sourceopenspec.devread 2026-11-18 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free and open source under the MIT license, with no paid tier.
No charge for the framework; the cost is the model usage of the coding agent it runs inside, billed by that agent's provider.
What is public
OpenSpec is free and open source under the MIT license, with no paid tier, hosted service or enterprise edition.
Billing mechanics
Nothing is billed by Fission AI; the coding agent that runs the commands bills its own model usage.
Cost watchouts
Each change spends agent tokens on exploring, proposing, verifying and archiving as well as on writing the code.
Variable cost rationale
The framework is free, so spend follows how many changes a team runs and how many agent turns each planning and verification step takes.
Additional watchouts
Stores, the shared requirements feature for teams, is in beta and may change, and usage telemetry stays on until it is switched off.
Sales call required
No, self serve available
Free / trial
Free MIT licensed CLI, plus a free Desktop beta for macOS
Lowest paid plan
None. There is no paid plan or enterprise edition.
Commercial notes
The CLI installs from npm, Homebrew, Nix or a standalone binary, and the Desktop beta for macOS is a free download that needs the CLI set up first.
Missing data
Fission AI states no price or plan for the team features it is building.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- 10Web — Agentic website platform whose AI agents build, host and manage…
- AgentUI — Managed AI app builder for operations teams that generates and hosts…
- Aider — Open source, model agnostic terminal coding agent that edits your…
- Amp — Amp is a coding agent from Amp Frontier Corporation that runs each…
- Anthropic Claude Code — Claude Code is Anthropic's agentic coding tool
Alternatives to OpenSpec
Other products that do the work OpenSpec does, and what each one is.
- ThinkReviewAI code review
- CodeGPTCoding agent for VS Code, JetBrains IDEs and Cursor
- Zed
- Blackbox AIFrontier inference platform
- EllipsisManaged agent platform for coding agents defined as YAML in your repo
- OpenCode