Spec-driven development tools for coding agents: OpenSpec, Spec Kit and four new frameworks compared

As of 12 October 2026, eight spec-driven development tools work with major coding agents, with OpenSpec recommended for developers already running Claude Code, OpenAI Codex or Cursor with an AGENTS.md because it is MIT-licensed, supported in 38 tools and used about half the tokens of GitHub Spec Kit in one author's two test runs. The four established tools are OpenSpec, GitHub Spec Kit, Kiro and BMAD-METHOD, while four newcomers — loki-mode, spec-superflow, Spectra and Agentforce ADLC — trended on GitHub in September at 111 to 1,073 stars. An ETH Zurich and LogicStar study revised on 29 September found developer-written context files raised resolution rates by 2.4% on average, not a significant change, and added up to 19% to inference cost, while LLM-generated files lowered resolution slightly.

Spec-driven development tools for coding agents: OpenSpec, Spec Kit and four new frameworks compared As of 12 October 2026, eight spec-driven development tools work with the major coding agents: four established ones, OpenSpec https://stackness.dev/tools/openspec , GitHub Spec Kit https://stackness.dev/tools/spec-kit , Kiro https://stackness.dev/tools/kiro and BMAD-METHOD https://stackness.dev/tools/bmad-method , and four that trended on GitHub in September, loki-mode https://stackness.dev/tools/loki-mode , spec-superflow https://stackness.dev/tools/spec-superflow , Spectra https://stackness.dev/tools/spectra and Agentforce ADLC https://stackness.dev/tools/agentforce-adlc . For a developer already running Claude Code https://stackness.dev/tools/claude-code , OpenAI Codex https://stackness.dev/tools/codex-cli or Cursor https://stackness.dev/tools/cursor with an AGENTS.md https://stackness.dev/tools/agents-md , the pick is OpenSpec: MIT-licensed, supported in 38 tools, and in one author's two test runs it used about half the tokens of Spec Kit. Each of the four newcomers comes with a catch. What does a spec-driven tool add over a plain AGENTS.md? A spec-driven tool adds a folder per change on top of the rules an AGENTS.md applies to every task. The folder holds a proposal, a spec and a task list, commands move the agent through the phases, a check runs before code and another after it, and the finished change merges into a living spec. An AGENTS.md has no phases, no per-change record and no check that the code matches what was agreed. An instructions file on its own does little for task success. In an ETH Zurich and LogicStar study https://arxiv.org/abs/2602.11988 revised on 29 September, developer-written context files raised resolution rates by 2.4% on average, not a significant change, and added up to 19% to inference cost; LLM-generated files lowered resolution slightly. Most spec tools still write into the instruction file: Spectra keeps its own block in CLAUDE.md and AGENTS.md , and Kiro loads AGENTS.md as steering. The AGENTS.md, CLAUDE.md and skills https://stackness.dev/blog/agents-md-vs-claude-md-vs-skills-which-layer-each-instruction-belongs-in-and-what-to-delete post covers what belongs in that file. How I picked the eight The four new tools are the repositories tagged for spec-driven development that reached my GitHub trend feed between 12 and 15 September, at 111 to 1,073 stars. The four established ones are the tools the comparison pages ranking for this query cover. I left out the plan modes built into agents, which keep no spec after the session. Stars, releases and downloads are from the GitHub, npm and PyPI APIs on 12 October, and download counts include CI runs. Eight spec-driven development tools The eight split into four established planners that write Markdown specs before code OpenSpec, Spec Kit, Kiro, BMAD-METHOD , an issue-to-pull-request agent loki-mode , a planner with an enforced review record spec-superflow , a desktop app Spectra and a Salesforce-only pack Agentforce ADLC . OpenSpec - Agents: 38 tools in its docs, among them Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI https://stackness.dev/tools/gemini-cli , OpenCode https://stackness.dev/tools/opencode and Kiro - Writes: openspec/changes/<change / with proposal.md , delta specs, design.md and tasks.md ; archiving merges the deltas into openspec/specs/ - Gates: your review of the proposal and openspec validate before code; /opsx:verify after, which "does not block archive" - Licence: MIT; anonymous telemetry on outside CI, off with OPENSPEC TELEMETRY=0 - Latest: v1.14.1, 6 October; 71,729 stars; 762,399 npm downloads in the week to 11 October OpenSpec is a Node CLI from Fission AI that writes /opsx: commands and skills into your agent, with "no rigid phase gates". Thoughtworks' Radar called it a "fluid, minimal workflow" "well-suited for existing systems". Its average daily npm downloads https://api.npmjs.org/downloads/range/2026-07-14:2026-10-11/@fission-ai/openspec rose 2.7 times between mid-July and early October, from about 46,000 to 125,000, counting only days npm reported. When to pick it: you work in an existing codebase and want a spec per change without a process around it. GitHub Spec Kit - Agents: 41 integrations, among them Claude Code, Codex CLI, Cursor, GitHub Copilot the default , Gemini CLI, opencode and Kiro CLI - Writes: .specify/memory/constitution.md , then per feature a numbered folder and Git branch with spec.md , plan.md , tasks.md , research.md , data-model.md and contracts/ - Gates: optional clarify , checklist and analyze before code, and implement asks first when checklist items are open; converge after, repeated until it reports "Converged" - Licence: MIT; needs Python 3.11 and uv - Latest: v1.1.3, 9 October; 140,958 stars; 22,916 PyPI downloads last week, with installs from Git, the route the docs recommend, not counted Spec Kit is GitHub's Python CLI, specify . converge arrived in June 2026, so reviews written before then tested a Spec Kit without it. On a feature that Birgitta Böckeler's past team would have sized at three to five points, Böckeler found Spec Kit's steps and Markdown files https://martinfowler.com/articles/exploring-gen-ai/sdd-3-tools.html "overkill for the size of the problem", and Thoughtworks' teams reported "overly verbose markdown outputs". When to pick it: new features that justify a plan and a data model each, or a team on Copilot. Kiro - Agents: its own IDE and CLI only, on Claude, GPT and open-weight models - Writes: .kiro/specs/<spec /requirements.md with acceptance criteria in EARS notation, design.md and tasks.md - Gates: your approval between requirements, design and tasks, skipped in Quick Spec; "Analyze Requirements" and property-based correctness checks in the IDE - Licence: proprietary; free with 50 credits a month, up to $200 a month for 10,000 pricing https://kiro.dev/pricing/ - Latest: IDE 1.2.56, 8 October; CLI 2.29.0, 9 October Kiro is Amazon Web Services' agentic IDE with specs built in. Böckeler watched it turn a small bug into four user stories with 16 acceptance criteria. When to pick it: you would switch editors to get specs and phase approvals built in. BMAD-METHOD - Agents: 48 platforms; Claude Code, Codex, Cursor and GitHub Copilot marked "preferred" - Writes: a brief, a PRD, an architecture document, then epics and stories - Gates: PRD validation before build; code review, adversarial and edge-case reviews and a retrospective at each epic - Licence: MIT, with a trademark notice - Latest: v6.12.2, 11 October; 54,058 stars; 18,518 npm downloads last week; v7 announced for 16 October BMAD-METHOD from BMad Code, LLC gives the agent agile roles, from analyst to QA, each writing its own document before the first story. When to pick it: a team that wants the whole agile sequence, from product brief to reviewed stories. loki-mode - Agents: Claude Code by default; Codex, Cline and Aider experimental; OpenCode not supported by the current engine - Writes: a receipt per run in .loki/runs/<run-id / , signed with Ed25519, and a pull request with a "NOT PROVEN" list - Gates: a delivery contract and checks written from the task before code; tests and a signed seal after; an unverified run opens a draft pull request - Licence: Business Source License 1.1, free for non-commercial and internal use, Apache 2.0 from 2030 - Latest: v11.3.9, 9 October; 1,093 stars; 5,133 npm downloads last week loki-mode from Autonomi turns an issue, spec or one-line task into a pull request. Three claims that travelled with it in September do not hold today: its README says it does not deploy, the current engine drops OpenCode and leaves four providers, and its own table marks one of its eight gates advisory, gates that belong to the older engine. One account made 3,796 commits in 30 days, about half of them naming Claude in the message, and 315 npm versions since mid-July inflate its downloads. When to pick it: an issue-to-pull-request loop with a receipt you can verify in CI, for internal use. spec-superflow - Agents: 19 platforms in its docs 17 in its repository description ; session hooks only on Claude Code, Cursor, Copilot CLI, Gemini CLI and CodeBuddy - Writes: changes/<name /proposal.md and tasks.md , plus delta specs and design.md when needed - Gates: one approval of the proposal and tasks before code; one review of the whole Git range and a recorded verification command after, ending verified or accepted-risk - Licence: MIT; README in Chinese, English in docs/README en.md - Latest: v2.0.1, 23 September; 851 stars; 111 npm downloads last week; no commits since 24 September spec-superflow combines OpenSpec's planning folders with the review and debugging discipline of Superpowers https://stackness.dev/tools/superpowers , without depending on either at runtime. When to pick it: OpenSpec's folders plus an enforced review record, if a quiet repository does not worry you. Spectra - Agents: Claude Code, Codex, Cursor, GitHub Copilot, Antigravity and Junie https://stackness.dev/tools/junie in the README; the website still says 23 - Writes: a proposal, specs, design and tasks per change under docs/spectra/ , and a Spectra block in CLAUDE.md and AGENTS.md - Gates: /spectra-discuss and /spectra-analyze before code; /spectra-verify , a review, a security audit and a drift check after - Licence: free, closed source; only the npm launcher is MIT - Latest: v3.0.0, 11 September; 747 stars; 1,078 downloads of the v3.0.0 app Spectra is a desktop app for macOS and Windows with a CLI, from Chien-Long Kao at 5xCampus. It started as an OpenSpec GUI, and its GitHub repository holds releases, not source. When to pick it: a desktop view over spec changes, if a closed binary is acceptable. Agentforce ADLC - Agents: Claude Code, Codex or Cursor - Writes: .agent files in Salesforce's Agent Script, Flow XML, Apex classes and tests - Gates: a safety gate before authoring; tests, adversarial probes and an OWASP LLM Top 10 grade; safety gate, validate, publish and activate before deploy - Licence: CC BY-NC 4.0, non-commercial - Latest: no releases; last commit 12 September; 114 stars Agentforce ADLC, the Agent Development Life Cycle from Salesforce AI Research, builds Salesforce Agentforce agents, not general application code, and needs a Salesforce org with Agentforce enabled. When to pick it: research or evaluation work on Agentforce agents; the licence rules out commercial use. Only Agentforce ADLC gates the deploy Seven of the eight spec-driven tools stop at a pull request or a merge, and only Agentforce ADLC checks before deploying, because it is the only one that deploys. Most of the gates run before code. After code, OpenSpec's verify is advisory, while spec-superflow and loki-mode record a verification result before a change counts as done. | Tool | Before code | After code | Before deploy | |---|---|---|---| | OpenSpec | Your review, validate | verify , advisory | None | | Spec Kit | clarify , checklist , analyze optional | converge | None | | Kiro | Phase approvals, Analyze Requirements | Property-based tests | None | | BMAD-METHOD | PRD validation | Reviews and a retrospective per epic | None | | loki-mode | Contract from the task | Tests, signed seal | None, it opens a pull request | | spec-superflow | One approval | Git-range review, recorded verification | None | | Spectra | discuss , analyze | verify , review, audit, drift | None | | Agentforce ADLC | Safety gate | Tests, adversarial probes | Safety gate, validate, publish | What does spec-first cost in tokens and time? One author has published measured token costs, and OpenSpec used about half of Spec Kit's. Jamie Telin ran one task twice https://medium.com/it-chronicles/is-your-safe-choice-burning-your-budget-1cfddf8782e4 per framework on GPT-5.2 and GPT-5.2-Codex in March 2026: 57,740 against 120,947 tokens in the first run, and 91,729 against 181,040 in the second, where planning alone took 38,117 against 96,298. The test had no baseline without a spec and no timings. Wall-clock figures are weaker. loki-mode's maker measured one 16-minute run https://github.com/asklokesh/loki-mode/blob/main/benchmarks/results/gate-profile.json in which eight of eleven gates took 2 seconds combined. A vendor benchmark from Uvik https://uvik.net/blog/spec-driven-development-benchmark/ gives a run date of 15 October 2026, after its own last edit, and its raw data is not reachable, so I left its numbers out. As of 12 October, no independent benchmark compares these tools against no spec with public data. Spec Kit files sit beside OpenCode and Copilot most often Committed Spec Kit command files show up most often in OpenCode and GitHub Copilot folders. A GitHub code search on 12 October found Spec Kit's specify command file 795 times in OpenCode folders, 583 times as a Copilot agent file, 460 in Cursor folders, 280 in Codex folders and 226 in Claude Code command folders. specify init suggests ignoring the agent folder in Git, so these count public files, not users. OpenSpec follows a similar order, led by OpenCode 909 and Codex 726 . On Stackness, as of 12 October 2026, 8 real profiles list Claude Code, 7 list Cursor and 2 list Codex, and none lists Spec Kit, Kiro or AGENTS.md data sources https://stackness.dev/about/data-sources . The numbers are small. The agents sit in the AI coding tools developers list https://stackness.dev/categories/ai-tools , and the habit the tools formalise is the documenting architecture and intent https://stackness.dev/moves/document-system-architecture-and-intent-explicitly-as-agents-write-more-code move. The eight side by side The eight differ most in licence and agent support. Three are MIT tools that support 38 or more agents, spec-superflow is MIT with 19, and the other four are tied to one editor, one main agent, one desktop app or one platform. | Tool | Type | Agents | Licence | Latest | Installs last week | |---|---|---|---|---|---| | OpenSpec | CLI and skills | 38 | MIT | 1.14.1, 6 Oct | 762,399 npm | | Spec Kit | CLI and skills | 41 | MIT | 1.1.3, 9 Oct | 22,916 PyPI | | Kiro | IDE and CLI | Its own | Proprietary | IDE 1.2.56, 8 Oct | Not published | | BMAD-METHOD | Method and skills | 48 | MIT | 6.12.2, 11 Oct | 18,518 npm | | loki-mode | Agent CLI | Claude Code, 3 experimental | BUSL 1.1 | 11.3.9, 9 Oct | 5,133 npm | | spec-superflow | Skills and CLI | 19 | MIT | 2.0.1, 23 Sep | 111 npm | | Spectra | Desktop app and CLI | 6 | Closed, free | 3.0.0, 11 Sep | 19 npm launcher | | Agentforce ADLC | Claude Code plugin | 3 | CC BY-NC | No release | Not on npm | Which one for which setup - Existing codebase, any agent: OpenSpec. - New features with a plan each, or a team on Copilot: Spec Kit. - Specs and phase approvals built into the editor: Kiro. - The whole agile sequence, brief to stories: BMAD-METHOD. - Issue to pull request with a signed receipt, internal use: loki-mode. - OpenSpec's folders plus an enforced review record: spec-superflow. - A desktop view over spec changes: Spectra. - Salesforce Agentforce agents, non-commercial: Agentforce ADLC. - Small changes with a good AGENTS.md: none of them. Tools in this post Use any of these tools? Put them on a Stackness profile, say how you use each one and see who pairs them the same way. It takes a couple of minutes. Show my stack https://stackness.dev/register