v1.0 · open source · Claude-first · now for any agent

A virtualengineering team,in one slash.

Born in Claude Code, Synthex now brings the same specialist workflows to any agent harness with Agent Skills support. Turn a messy idea into a peer-reviewed plan and production-ready code — without you writing a for loop.

Get started →View on GitHub
CLAUDE CODE INSTALL/plugin marketplace add bluminal/lumenai
claude-code · ~/your-project
▸ /plugin marketplace add bluminal/lumenaiadded bluminal/lumenai · 1 plugin available▸ /synthex:init∙ scaffolding .synthex/config.yaml∙ creating docs/reqs/ docs/plans/ docs/specs/✓ ready▸ /synthex:next-prioritytech-lead → identifying ready tasks └ 3 tasks parallelisable · spinning up worktreesfrontend-eng · quality-eng · code-reviewer · security-reviewer …implementing 2.3.1 Real-Time Notification System✓ acceptance criteria met · merging▸
The shift

Stop writing code. Start directing it.

Three convictions that the entire system is built around — they cascade into every command, agent, and quality gate.

/ 01

Structure creates quality.

Discovery → planning → execution as distinct phases with review gates between them. Code without a plan is lumber without blueprints.

/ 02

Specialisation beats generality.

One agent asked to "build a login page" produces something that works. Fifteen specialised roles produce something that's production-ready.

/ 03

Parallel execution changes the game.

Git worktrees isolate every task. Three features get implemented at once without conflict. A sprint becomes an afternoon.

Cursor cloud agents

Fan your plan out to the cloud.

On Cursor, next-priority/synthex:next-priority with --in-cloud hands each parallel plan task to its own cloud agent — a separate VM on a separate branch — and merges the work back only after it validates. Add --loop and it keeps going until the plan is done.

One VM and one branch per task

Each Tech Lead worker forks from your pushed base branch, commits on its own branch, and pushes it. Workers run in the background, so the batch truly runs in parallel while you handle [H] gates.

Merged only after validation

Your session fetches each worker's branch after its tests and approvals pass, then fast-forwards it in, falling back to a merge commit. Commit and push before you start: cloud clones can't see uncommitted edits.

Sticky default, safe fallback

Set next_priority.execution_environment: cloud to make it the default; --local overrides per run. Other hosts print a one-line notice and fall back to local worktrees.

cloud agents
▸ /synthex:next-priority --in-cloud --loop [base: origin/main · environment: cloud] ↳ 3 tasks · 3 cloud agents · running in background ↳ feature/12-login-form pushed ↳ feature/13-session-store pushed ↳ feature/14-audit-log pushed  [T] tests pass · git fetch · git merge --ff-only FETCH_HEAD✓ 3 merged · next batch

Keep team defaults in .synthex/config.yaml and your own in .synthex/config.local.yaml — see how config layers resolve.

Read how cloud agents work →
Lifecycle

Idea to production, traced end-to-end.

Every task maps to a requirement, every change maps to a task, every architectural decision is recorded. Five phases, one continuous loop.

  1. PHASE · 01

    Discover

    Brain-dump → conversation → PRD that captures the why.

  2. PHASE · 02

    Plan

    Milestones, phases, tasks. Peer-reviewed before a line of code.

  3. PHASE · 03

    Build

    Parallel worktrees, specialist agents, mandatory quality gates.

  4. PHASE · 04

    Ship

    Performance audited, observable, and operationally ready.

  5. PHASE · 05

    Learn

    DORA, retros, and follow-through loops back into discovery.

Native looping

Iterate any command until it's done.

Add --loop to next-priority/synthex:next-priority, review-code/synthex:review-code, write-implementation-plan/synthex:write-implementation-plan, or refine-requirements/synthex:refine-requirements — and the command keeps re-running until it emits a configured completion promise. State lives on disk, so the loop survives session restarts and crashes.

Drive a plan to done overnight

Wrap next-priority/synthex:next-priority in a loop and let it pick up unblocked work as you finish [H] reviews in a parallel thread. Per-session state means two sessions in the same project don't collide.

Resume after closing your laptop

State lives at .synthex/loops/<loop-id>.json. Re-open Claude Code, run loop/synthex:loop with --resume, and the loop picks up at the iteration counter where it stopped.

Or loop an arbitrary prompt

The generic loop/synthex:loop primitive iterates any prompt — literal text or a file — until a configured <promise> tag fires. Same iteration framework, same per-session safety, no command body required.

native looping in 4 lines
▸ /synthex:next-priority --loop --completion-promise "PLAN_DONE" --max-iterations 30 [loop next-priority-3f2a iteration 1/30] ↳ 3 tasks · 3 worktrees · parallel✓ merged · 17 tasks remain · re-running  [loop next-priority-3f2a iteration 8/30]✓ every task done · <promise>PLAN_DONE</promise>
Read the looping guide →
The org chart

A specialised team. Three layers. One coherent system.

Modelled on the roles a real software startup would hire — clear scope, structured delegation, honest peer review.

Tier I

Orchestration

Coordinates execution
01 / TL

Tech Lead

Full-stack orchestrator. Breaks tasks down, delegates to specialists, rolls up the result.

02 / FE

Lead Frontend Engineer

Coordinates frontend delivery, enforces UX quality, accessibility, and design-system fidelity.

03 / PM

Product Manager

Interactive requirements gathering. Transforms PRDs into milestone-based plans.

Tier II

Specialists

Domain expertise
04

Architect

System design, ADRs, plan-review feasibility, technology trade-offs.

05

Code Reviewer

Craftsmanship review — correctness, maintainability, reuse, conventions.

06

Security Reviewer

Mandatory gate. Secrets, injection, auth, supply chain, OWASP.

07

Quality Engineer

Test strategy, gap analysis, behavior-driven test authoring.

08

Terraform Reviewer

Plan review with cost, blast-radius and security verdict for AWS.

09

Design System

Token ownership, component governance, compliance audit.

10

Performance Engineer

Quantified, full-stack performance analysis with budgets.

11

SRE

SLOs, observability, runbooks, deployment risk, postmortems.

12

Technical Writer

API docs, guides, migrations, changelogs verifiable against code.

Tier III

Research & Analysis

Continuous improvement
13

UX Researcher

Personas, journey maps, opportunity-solution trees, heuristic eval.

14

Metrics Analyst

DORA + HEART + AARRR + OKRs. Trend, benchmark, learn.

15

Retrospective Facilitator

Planned-vs-actual, blameless retros, tracked improvement loops.

Commands

Delivery workflows, one lifecycle.

Each command orchestrates a subset of agents. Mix them, chain them, or let next-priority run the whole loop.

  • init/synthex:initsetupScaffold project structure and Synthex configuration.
  • configure-multi-model/synthex:configure-multi-modelsetupWizard that detects installed model CLIs, checks auth, and enables multi-model review.
  • configure-teams/synthex:configure-teamssetupWizard for standing review pools — enable routing and pick routing and matching modes.
  • dismiss-upgrade-nudge/synthex:dismiss-upgrade-nudgesetupSilence the session-start upgrade nudge for this project.
  • star/synthex:starsetupOffer to star the LumenAI marketplace repository on GitHub.
  • refine-requirements/synthex:refine-requirementsdiscoverMulti-agent PRD review loop — Product Manager, Tech Lead, and Lead Frontend Engineer tighten clarity, completeness, and communicability.
  • write-implementation-plan/synthex:write-implementation-plandiscoverPeer-reviewed plan from a PRD — Architect + Design System agent + Tech Lead.
  • next-priority/synthex:next-prioritybuildPick top tasks and fan them out to parallel Tech Leads — in local worktrees, or as Cursor cloud agents with --in-cloud.
  • loop/synthex:loopbuildNative looping primitive — iterate any prompt until a completion promise fires or max-iterations hits.
  • list-loops/synthex:list-loopsbuildEnumerate running and recent loops in the project. Read-only.
  • cancel-loop/synthex:cancel-loopbuildCancel a loop by id, or every running loop with --all. Idempotent.
  • review-code/synthex:review-codebuildMulti-perspective review — Code + Security + (optional) Performance, prompt-file, and multi-model reviewers.
  • start-review-team/synthex:start-review-teambuildStart a standing review pool — warm reviewers that take reviews without per-review spawn latency.
  • stop-review-team/synthex:stop-review-teambuildDrain in-flight reviews and shut down one or all standing review pools.
  • list-teams/synthex:list-teamsbuildList running standing review pools and per-invocation teams, with composition and load.
  • write-adr/synthex:write-adrbuildInteractive ADR authoring with the Architect.
  • write-rfc/synthex:write-rfcbuildRFC drafting with Architect & PM, parallel review pass.
  • test-coverage-analysis/synthex:test-coverage-analysisbuildRisk-prioritised gap analysis. Optional fix mode writes tests.
  • design-system-audit/synthex:design-system-auditbuildAudit code for token compliance and component governance.
  • performance-audit/synthex:performance-auditshipQuantified analysis across frontend, backend, db, infra.
  • reliability-review/synthex:reliability-reviewoperateSLO coverage, observability gaps, runbook & deploy-risk review.
  • retrospective/synthex:retrospectivelearnPlanned-vs-actual + structured retro with tracked follow-through.
Full command reference →
"I'm still in control when I need to be. I'm still monitoring outputs, validating decisions. Claude provides the autopilot — but I'm the pilot. I decide when to take the stick."
— AJ BROWN · ON AGENTIC SOFTWARE DELIVERY
Quickstart

Claude Code first. Any skills-capable agent next.

Synthex is markdown all the way down. Claude Code gets the native plugin experience; Codex, Gemini CLI, OpenCode, and other Agent Skills hosts use the same workflows.

  1. Start with your agent

    Claude Code and Codex install Synthex as a plugin. Other compatible harnesses load its shared Agent Skills bundle.

    /plugin marketplace add bluminal/lumenai

  2. Install the workflow bundle

    Use the compatibility guide for the exact command for your harness, then scaffold .synthex/config.yaml and the standard docs/ tree.

    Choose your agent →

  3. Describe what you want, then ship it

    Write a PRD with the Product Manager, generate a peer-reviewed plan, then let parallel Tech Lead instances execute.

    /synthex:next-priority
a typical session
▸ /synthex:write-implementation-planproduct-manager · architect · design-system-agent · tech-lead ↳ peer review · 3 cycles · approved✓ docs/plans/main.md written ▸ /synthex:next-priority ↳ 3 tasks · 3 worktrees · parallel code-review · security-review · tests✓ merged · plan advanced ▸ /synthex:retrospective ↳ planned-vs-actual · 18/20 tasks · 2 deferred✓ next cycle informed
Read the docs →