A wizard configures @specrouter-dev — an orchestrator agent that drives the whole HU → SPEC → CODE → QA flow for OpenCode and Claude Code. Plain markdown prompts and skills. You write them, you ship the ZIP, you keep the code.
Every phase transition, every artifact, every gate — visible in a local kanban. No cloud. No account. Just your docs/.
# SpecRouter.dev (Orchestrator) You drive a 4-phase SDD flow. Between phases you stop and wait for human approval. ## What you load Each phase is a sub-agent — invoke it as the orchestrator signals: - @phase-hu → docs/user-stories/US-<id>-<slug>.md - @phase-spec → docs/specs/SPEC-<id>-<slug>.md - @phase-code → implementation + tests + commit - @phase-qa → docs/qa/QA-<id>-<slug>.md (verdict) ## Approval modes (picked in the wizard) strict · assisted · autonomous · custom ## Opt-in skills you can also load @vault — token usage report @backlog — capture a product idea
Generates ready-to-use ZIPs for
@specrouter-dev.opencode/agents/.opencode/skills//specrouter-dev.claude/agents/.claude/skills/Three steps. No GitHub login. No account. No telemetry.
OpenCode or Claude Code. The single-agent orchestrator (recommended) or four separate agents. Strict, assisted, autonomous, or custom approval between phases. The wizard remembers your choices — re-run anytime to iterate.
Describe each agent in one sentence. The LLM drafts the system prompt — you edit, accept, or rewrite. Same for skills. Or skip AI and write them yourself. You stay in control of every word in the ZIP.
One ZIP with the orchestrator, phase sub-agents, opt-in skills, and project files. Drop it in your repo, add your API key, run @bootstrap. Five minutes from zero to a working SDD team.
Pick the mode in the wizard. You can change it any time by regenerating the ZIP — your custom prompts and skills stay.
The wizard emits a single agent — @specrouter-dev — that owns the full HU → SPEC → CODE → QA flow. Phase sub-agents are loaded on demand. You talk to one handle, it sequences everything for you, and pauses for approval between phases.
The wizard emits four separate agents — @product-owner, @sdd, @coder, @qa. You invoke each one yourself in any order. Every phase is its own hard gate: nothing moves forward until you say so.
Both modes ship with the same approval flow knob — strict, assisted (default), autonomous, or custom. Both ship with the same opt-in skills (@vault, @backlog). The only thing that changes is how many agents you see in your Tab picker.
Most agent tools ship a curated preset and ask you to live with it. SpecRouter ships a generator and a blank canvas.
SpecRouter ships a generator, not a curated preset. Every agent prompt starts blank — or from a starter you can rename, rewrite, or delete. No persona baked in.
The ZIP is plain markdown. Edit it, fork it, version it in your repo. SpecRouter does not phone home, inject telemetry, or wrap your agents in a proxy. MIT-licensed code, all of it.
Pick strict gates or relaxed ones. Bundle the optional local dashboard. Add or remove phase sub-agents as your project grows. SpecRouter fits your team, not the other way around.
Three audiences. One generator. Zero opinions baked into the output.
You're tired of copy-pasting prompt fragments between projects. Generate a clean SDD flow once, version it in your repo, and re-import it next time.
Every teammate ships a slightly different AI flow. SpecRouter gives you a baseline the team agrees on, with room for each project to diverge.
Auditable. Forkable. No proxy. SpecRouter generates artifacts you can review line by line and ship to your compliance team without translation.
Every handle below is a file in your ZIP. Edit them. Commit them. Rename them. They are plain markdown.
@bootstrapInitialize project context, then hand off to @specrouter-dev.
@specrouter-devThe orchestrator. Drives HU → SPEC → CODE → QA, pauses between phases for human approval.
@phase-huPhase sub-agent — produces a single user story. Loaded by the orchestrator; you can also invoke directly.
@phase-specPhase sub-agent — refines the HU into a technical spec with ordered tasks.
@phase-codePhase sub-agent — implements the spec with TDD, writes the commit.
@phase-qaPhase sub-agent — verifies the implementation against the original HU. Closes the flow.
@product-ownerProduces the user story (HU) with acceptance criteria.
@sddRefines the HU into a technical spec. Independent approval gate.
@coderImplements the spec with TDD. Independent approval gate.
@qaVerifies the implementation. Independent approval gate. Closes the flow.
@vaultReports token usage across the SDD flow's LLM calls (per-phase breakdown + grand total).
@backlog <idea>Captures a product idea to docs/backlog.md (timestamped, git-tracked).
@backlog listReads docs/backlog.md and outputs the pending ideas.
The dashboard shows your SDD flow organized in a kanban that updates as your agents work. Every phase transition, every artifact, every gate — visible at a glance. Runs locally. No cloud. No account. Just your docs/.
See your HUs move from pending → approved → in progress → done. Every phase transition is visible.
Watch your flow progress in real time. Know exactly where each feature stands.
Find any HU, spec, or QA report across your entire project. No more grep.
Trace every HU to its SPEC, modified files, and tests. Full lifecycle visibility at a glance.
Reads from docs/ — the markdown you already commit. Zero sync, zero drift.
The dashboard exposes the same data the agents see. To enable, check the “SpecRouter dashboard” box in the wizard's flow-preferences step.
The generator stays free. Pro funds the hosted features so the open-source core keeps moving.
The full generator. Both delivery modes (single-agent & four-agents). All four phases. Both tools. Use it as long as you want.
Generate my ZIPFor teams that want hosted prompt libraries, shared dashboards, and priority AI refinement.
Join the waitlistNo pricing set yet — we will email you when it ships.
SpecRouter is MIT-licensed open source. Read the source on GitHub.
Most agent tools hide their hooks behind a marketing page. We'd rather show our work.
Source on GitHub. License is MIT. No proprietary hooks, no SaaS-only features, no upgrade wall for the things the generator already does.
We do not track what you generate, what you download, or what you ship. The wizard runs in your browser; the LLM customization step uses your key in memory and discards it.
The ZIP is plain markdown. Edit it in your repo. Fork it. Ship it without us. We win by being useful, not by holding your agents hostage.
We never see, store, or proxy your LLM API key. Pay your model provider directly — pay cents, not dollars. No middleman markup.
OpenCode and Claude Code today. Both consume plain markdown for agents and skills, so the generated ZIP drops in without extra adapters. Cursor, Codex, Kiro, and Continue.dev are on the roadmap — vote on the next one by opening a GitHub issue.
No. SpecRouter is a generator that runs in your browser — no login, no signup, no server-side library. You bring your own LLM API key for the prompt-customization step; the key is used in memory and immediately discarded. The generated ZIP is yours.
Most agent-configuration tools ship a curated preset — their authors' opinions baked in — and ask you to live with it. SpecRouter ships a generator and a blank canvas: every agent prompt is yours to define, rename, or delete. No preset persona, no routing rules you didn't pick, no telemetry wrapper around your agents, no SaaS lock-in.
Yes — that's the whole point. The ZIP contains plain markdown files. Edit them in your repo, commit them, ship them. The next time you run SpecRouter you can re-import your edited prompts or start fresh.
Yes. The generator is MIT-licensed open source and will stay free forever. A Pro tier is on the roadmap for hosted team features (shared prompt library, hosted dashboards, priority AI refinement), but the generator, the ZIP output, and the core flow remain free. Read more on the pricing page.
We have not set Pro pricing yet — we will email waitlist subscribers when we do. Pro is opt-in only; existing free users keep everything they have today even when Pro ships.
Yes. Your key is held in your browser's memory only, used for the single customization request you triggered, and never written to disk, sent to our servers, or included in the ZIP. Three independent regression suites verify this — see the security tests in the repo.
The generator runs entirely in your browser and works without network access for the UI. The only network call is the optional AI customization step (POST /api/generate-agent), which requires your key and a reachable model provider.
It depends on the mode you picked. Single-agent (default) ships one orchestrator (@specrouter-dev) plus four phase sub-agents (@phase-hu, @phase-spec, @phase-code, @phase-qa) and the @bootstrap entry-point. Four-agents ships four separate agents (@product-owner, @sdd, @coder, @qa). Both modes can opt into @vault (token usage report) and @backlog (capture product ideas with @backlog <idea>, retrieve with @backlog list).
@specrouter-dev is the orchestrator — it owns ordering, gates, and human approval. The @phase-* sub-agents contain the procedures the orchestrator runs (one per phase). You can also invoke a phase sub-agent directly when you want to skip the orchestrator for a single re-run, but you'll lose the gating and the audit trail of the full flow.
Pick your tool. Pick your flow. Generate. You keep the prompts, you keep the ZIP, you keep the code.
Free during beta · MIT licensed · No signup