Spec-Driven Development Tools in 2026: An Honest Field Guide

The field

GitHub Spec Kit (free, MIT-licensed: a permissive open-source license that lets you use and modify the code freely). The baseline, at roughly 130k+ GitHub stars: slash commands that walk any of 30+ coding agents from constitution to spec to plan to tasks to implementation, producing markdown artifacts. Its convergence check, /speckit.converge, compares codebase to spec and appends remaining work. The check is a prompt template the agent executes: judgment, not a gate. Our pairwise comparison goes deeper.

Kiro (AWS; free tier, then paid monthly tiers billed in usage credits). A standalone agentic IDE with specs built in: requirements in EARS notation (a constrained sentence format for writing testable requirements), design, and tasks per feature. Credit where due: Kiro documents real machine-executed verification, including property-based testing (auto-generated tests that check a rule holds across many random inputs, not one example) and automated reasoning, both driven by the requirements, and its own docs are honest that this is evidence of correctness rather than proof. In this field, that makes it the other tool taking deterministic checking seriously. It asks you to adopt its IDE, and it is proprietary.

OpenSpec (free, MIT). Change-proposal SDD for multiple coding agents: each change is a proposal folder with spec deltas, applied and archived. Popular (roughly 65k+ GitHub stars) and brownfield-friendly (built for changing an existing codebase, not just starting a fresh one). Verification is an agent-run step; the repo ships structural schemas, but no standalone machine gate is documented in the current README.

BMAD-METHOD (free, MIT). The methodology-heavy option: an installable pack of agent workflows running a scale-adaptive loop from clarify to plan to build to learn, with role agents. Ambitious and actively developed (roughly 50k+ stars). Validation is agent-assisted review; no deterministic gate documented.

Cursor Plan Mode (part of Cursor, paid). Not a full SDD system: the agent researches, asks clarifying questions, and writes a markdown plan you review before it builds. A planning step, with no documented verification that the implementation honored the plan.

SpecWeaver (ours; $37 one-time). The narrow tool in the list, on purpose: a CLI (a command-line tool: install it, run it, no accounts or API keys). It does not structure your workflow. It makes the spec checkable: acceptance criteria that carry shell commands whose exit code (the numeric pass or fail result a command returns: 0 passes, anything else fails) is the verdict, drift detection against the files and symbols the spec declares, and a gate command that passes or blocks with one exit code, annotating GitHub pull requests. Works with any agent that accepts structured specs.

The comparison that matters

Read the last column twice. Three of six tools resolve "is it done?" with model judgment, which inherits the model's generosity, and Cursor Plan Mode never reaches the question. Two make a machine answer: Kiro inside its IDE, SpecWeaver as a $37 CLI you drop into whatever you already use. That framing is not ours alone: Augment Code's 2026 roundup ranks the field on verification architecture and measured spec drift as a core criterion.

How to choose

• Want free structure for any agent? Spec Kit, and OpenSpec for change-heavy existing codebases.
• Want an IDE that owns the whole loop, verification included? Kiro, if credit pricing and a proprietary editor fit you.
• Want a methodology to think in? BMAD.
• Want a deterministic gate without changing anything else? That is SpecWeaver, and it composes with the others: structure upstream with Spec Kit, gate downstream with us.