+BMAD+ · the multi-agent framework for AI coding assistants, built on the BMAD Method
Your AI.A whole team.
BMAD+ gives the AI assistant you already use a team of specialist agents. They ask what they need before they build, split the work under a budget you set, and report what was actually checked.
Works in Claude Code, Gemini CLI, Antigravity, Cursor, Codex CLI, OpenCode, and Aider. One command installs it.
The one-minute overview
BMAD+ in brief
Duration 1:15 · Voice-over: English
What BMAD+ installs, how a pack asks before it acts, how a swarm is counted honestly, and the ten packs.
Transcript
BMAD+ installs roles, workflows and shared context into the AI coding tool you already use. It starts by asking: a pack declares the questions it needs answered before it acts. You answer a few at a time, and what you do not know is recorded as unknown, with its risk. Then Nexus can run a swarm of sub-agents, within a budget you set. Each one reads a brief and writes an answer, and only valid answers are counted. The tally on screen is demo data, and it says plainly that simulated units are not people. For planned work, Nexus requires independent checks before it accepts a result. Core's four roles, nine optional packs, and adapters for seven AI coding tools. Install it with one npx command from your project folder.
BMAD+ is a set of instructions, skills and a small command line that your AI assistant follows. Each move below is something you can run; the panels quote real output from BMAD+’s own specs, fixtures and acceptance run, shortened where marked, and say plainly when a part is simulated.
“bmad-plus” stands for “npx bmad-plus@0.23.1” in the transcripts.
Move 01
Ask the right questions first
A pack can declare the questions it needs answered before it acts. bmad-plus intake asks a few at a time, marks the must-haves, proposes defaults and records what you do not know together with the risk of not knowing. Nothing is invented: the brief it renders becomes the contract for what comes next.
Must-haves first, six questions per message by default
Answered, defaulted or unknown, never silently guessed
The brief says which answers are yours and which are proposals
Simulatedillustration, not captured outputIn your assistant
You
Atelier, design the invoice list of my app.
Atelier
Before I design anything, three questions: which kind of surface is this, who uses it and in what situation, and what is the one job this page must get done?
You
A dashboard. Accountants, at a desk, many times a day. Finding an unpaid invoice in under ten seconds.
Real outputquoted, shortenedIn the terminal
$ bmad-plus intake plan --intake .agents/skills/pack-design/intake.yaml --answers _bmad-output/design/intake.jsondesign-read (design): not ready — 0 answered, 0 defaulted, 0 unknown, 13 not asked yet.Must-haves still open: surface_kind, audience, jobAsk next (at most 6 per message): - surface_kind [choice] Which surface is this? […] (must-have; choices: landing, dashboard, table, […]) - audience [text] Who uses it, and in what situation […] (must-have) - job [text] What is the one job this page must get done […] (must-have) […]$ bmad-plus intake record --intake […] --answers […] --set surface_kind=dashboard --set "audience=[…]" --set "job=[…]"Recorded surface_kind, audience, job in […].$ bmad-plus intake render --intake […] --answers […] --out _bmad-output/design/brief.mdBrief written to _bmad-output/design/brief.md ([…] bytes); the intake is ready.
Move 02
Turn the answers into a plan
Nexus splits approved work into tasks with dependencies, disjoint write scopes and a check for each one, and keeps the run on disk so it survives a closed session. You decide where a checkpoint stops the run for approval, where it only notifies, and where it continues.
Tasks never share a write scope
Checks are literal commands, run without a shell
The plan is a file you can read and review
Real outputquoted, shortenedA Nexus plan, from the specification
$ bmad-plus nexus create --plan nexus-plan.json --directory PROJECT --json$ bmad-plus nexus verify example result --directory PROJECT --json$ bmad-plus nexus accept example result --directory PROJECT --json
Move 03
Let a team work in parallel, under a budget
Agents take roles. A swarm deals a panel, a jury or a fan-out of sub-agents, each with one written brief. The run is capped by a budget (26 sub-agents by default) and goes past it only with your explicit approval. Parallel execution itself depends on your AI tool.
Same profile, same count and same seed give the same cards
Waves of 10 by default
Paid media generation always asks
Simulatedillustration, not captured outputThe budget
This run needs 30 sub-agents · Budget 26 · Waves of 10
Real outputquoted, shortenedA panel of 30 against a budget of 26
$ bmad-plus swarm init --definition swarm/buyer-panel.json --profile swarm/profile.json --context pitch=swarm/pitch.md --n 30 --seed 7<run>: needs approval. The run needs 30 sub-agents and the budget is 26: ask the user, then re-run init with --approved --seed 7.(exit 2, no brief written)$ bmad-plus swarm init […] --n 30 --seed 7 --approvedbuyer-panel-20261018-oqu7: panel of 30 units, budget 26, approval user.Seed 7 (same seed, same cards).$ bmad-plus swarm briefs buyer-panel-20261018-oqu7buyer-panel-20261018-oqu7: 30 of 30 units need a sub-agent, in waves of 10.Wave 1: B001 […] […]Each sub-agent: Read the brief and follow it exactly; it is your whole instruction.
Move 04
Check what came back
Each answer file is validated and only valid answers are counted. Nexus runs the plan’s checks itself. A code review gate answers incomplete, findings or clean, and says that clean does not prove the code correct. The tally opens with the honesty disclosure.
A rejected answer writes nothing
Skipped units are named, not hidden
Simulated units are not people
Real outputquoted, shortenedThe honest result of a run
$ bmad-plus swarm check buyer-panel-20261018-oqu7buyer-panel-20261018-oqu7: 1 of 30 valid answers. B002 missing […]$ bmad-plus swarm save buyer-panel-20261018-oqu7 B002 --input swarm/reply-bad.txtSwarm: […] buys must be true or false(nothing written)$ bmad-plus swarm tally buyer-panel-20261018-oqu7[…]/tally.md1 of 30 units counted; 29 skipped.Disclosure: Simulated units are not people. Read rates as an upper bound and confirm with real ones.
Review gate verdicts
incompleteexit 2
findingsexit 1
cleanexit 0
Move 05
Keep what was learned
Decisions, lessons and patterns live in plain files in your project, so the next session starts from them. Recall is a ranked search over those files and the journal; promotions are only ever proposed. Uninstalling BMAD+ keeps your memory.
No server, no vector service required
A global brain is linked, never overwritten
Dates and sources stay with each entry
Real format, sample dataProject memory
$ bmad-plus mem write --agent forge --task "Switch sessions to signed cookies" --outcome success --ci passmem write: event […] appended ([…]) agent=forge cli=- model=- outcome=success$ bmad-plus mem recall "session cookies"mem recall: 2 match(es) for "session cookies" (scope: project) 1. [0.812] note decisions.md § […] 2. […]
Real outputquoted, shortenedOwnership, not takeover
$ bmad-plus uninstall --yesProject memory, settings, and unknown or modified files have been preserved.
Plain files in your repository
.agents/memory/decisions.md
.agents/memory/lessons.md
.agents/memory/patterns.md
.agents/memory/context.md
.agents/memory/sessions/
02
Packs
The right expertise,when you need it.
Core is always installed. Pick the other packs during installation, or later with the same command. Each pack adds agents, skills or supporting tools; select one to see what it does, a prompt to try and the command that installs it.
Core
Always installed
The four agents that carry a project from idea to delivery, and the skills behind autopilot, parallel work, swarms, intake, human acceptance recipes and update checks.
Atlas, help me define a small invoicing app for freelancers. Ask about the users first.
Install
npx bmad-plus@0.23.1 install
Needs
Nothing beyond Node.js.
OSINT
Governed research on public sources. Before any item about a named person, Shadow requires a recorded purpose and legal basis. Dossiers carry confidence grades A to D and an AI-generated disclaimer.
Governed use: named-person investigations stop unless you state a lawful basis and purpose.
One orchestrator that routes a compliance question to the right specialist, maps controls across frameworks and writes a consolidated report. Every report is marked AI-generated and not legal advice. Evidence tooling included: assurance cases and an AI processing register.
Search optimization and GEO: being found in AI-generated answers as well as in search results. Scout crawls and inspects the site technically and for performance, Judge assesses content quality, structured data and AI-search readiness, Chief scores, plans and reports. A 6-phase audit with a PageSpeed loop. Without Python it runs in prompts-only mode and marks unverified categories instead of scoring them.
Optional: Python 3.11 or newer, set up in an isolated environment with --provision-python. GOOGLE_API_KEY is optional.
Memory
Persistent memory in plain files: decisions, lessons, patterns, context and session handoffs. It detects and links an existing brain and never overwrites memory files.
Works alongside the other packs; where roles overlap, the pack’s own rules take over.
Dev Studio
A full software lifecycle in 38 workflows, from analysis and planning to architecture, implementation and review, spread over six personas that sit beside Core without replacing it.
Atelier interviews you with a design read, then builds, audits or redesigns an interface under checkable rules: semantic tokens, WCAG AA, every state covered and a pre-flight verdict quoted at the end.
Atelier, run the design read for the invoice list of my app, build the screen, and end with the pre-flight verdict.
Install
npx bmad-plus@0.23.1 install --packs design
Needs
The design-skill package, installed with the pack.
Works alongside the other packs; where roles overlap, the pack’s own rules take over.
Atelier and the ask-first story
The Design pack is the one that ships an intake today: a design read that comes before any pixel. The same mechanism is open to any pack, and to your own.
Two mechanisms that stop an assistant from guessing and from overspending. Both are plain files plus a command that checks them, so you can read what happened afterwards.
Intake: the questions come before the work
A pack that declares an intake does not start until its must-have questions are answered. The host assistant asks them in your language, a few at a time; the CLI validates and records each answer; the rendered brief is what the next step is held to.
1
Plan
bmad-plus intake plan lists what is still open, must-haves first.
2
Ask
Your assistant asks at most six questions in one message (the default; each intake can set its own limit).
3
Record
Each answer is stored as answered, defaulted or unknown. An unknown copies its risk.
4
Render
The brief marks every answer with its status. It never overwrites a brief it did not write.
answered
You said it.
defaulted
A proposed default stood. The brief says it is a proposal, not your fact.
unknown
You did not know. The risk of not knowing is written next to it.
A swarm deals units, writes one brief per sub-agent, validates every answer file and counts only the valid ones. The runtime never calls a model, opens no network connection and starts no process: your assistant launches the sub-agents, the CLI keeps the books.
Panel
Units dealt from a profile by exact quotas and a seed, such as simulated buyers or readers.
Jury
Units listed in the definition, each with its own lens on the same material.
Fan-out
Units supplied at init, each with a disjoint write scope; exportable as a durable Nexus plan.
Simulatedillustration, not captured outputThe budget
This run needs 30 sub-agents · Budget 26 · Waves of 10
The default is 26 sub-agents per run in waves of 10. A run that needs more is refused until a person approves it.
Limits. 1 to 1000 units per run; briefs and answers stay in .bmad-plus/swarm/ in your project.
Simulated units are not people. A rate from a panel is an upper bound to confirm with real ones, its quotes are not testimonials, and every tally opens with this sentence.
“Simulated units are not people. Read rates as an upper bound and confirm with real ones.”
One generated spine, AGENTS.md, plus the instruction file each tool expects. The files are generated from the registry and never hand-edited; Instructions it did not write are preserved, or copied to a .bak file before a non-interactive install (--yes) replaces them.
Claude Code
Writes
CLAUDE.md
Detected from
.claude/
Gemini CLI
Writes
GEMINI.md
Detected from
.gemini/
Antigravity
Writes
GEMINI.md
Detected from
.gemini/antigravity/
Shares the file with Gemini CLI
Cursor
Writes
.cursor/rules/bmad-plus.mdc
Detected from
.cursor/
Codex CLI
Writes
.codex/AGENTS.md
Detected from
.codex/
OpenCode
Writes
.opencode/AGENTS.md
Detected from
.opencode/
Aider
Writes
CONVENTIONS.md
Detected from
.aider.conf.yml
Aider also needs read: CONVENTIONS.md in .aider.conf.yml, or /read CONVENTIONS.md in the session.
Host-managed: your tool supplies the model, the permissions and any parallel execution. BMAD+ is not an agent runtime; the only program it launches is one your Nexus plan declares.
The commands bmad-help, autopilot and parallel are instructions you give your assistant, not shell commands.
Models are not tied to a vendor: Claude, GPT, Gemini and local models are supported by contract.
10 installer languages. The installer and uninstaller screens are available in English, French, Spanish, German, Brazilian Portuguese, Russian, Chinese, Hebrew, Japanese and Italian: pass --lang. Every other command prints English.
07
Security and data handling
Memory you own.Trust you can check.
Everything BMAD+ keeps is a file in your repository. Everything it could send off your machine is declared, and a build gate fails when code does something undeclared.
Memory in plain files
Decisions, lessons, patterns and context live in .agents/memory/. A project brain is found in this order: BMAD_PLUS_BRAIN, a _brain/ in the project or above it, ~/.bmad-plus/brain, then a non-empty ~/.claude/memory. The link is recorded, nothing is overwritten.
decisions.md
lessons.md
patterns.md
context.md
sessions/
Ownership, not takeover
The installer records a SHA-256 for every file it writes. Uninstall removes only files that still match, plus the manifest. Your config, your memory, your outputs and anything you edited stay.
Customised files are preserved on update and reported as conflicts
Every update leaves a receipt and backups, and can be restored
doctor reports the inventory: complete, incomplete, legacy or invalid
Updates that ask first
The default policy only notifies. Automatic updates need an explicit semver range, a fresh registry answer and an exact expected version. Offline or stale means unknown, never “up to date”.
It is not a model or an agent runtime; it launches only the commands your Nexus plan declares.
No background process, daemon or scheduler.
No telemetry and no silent self-update.
No automatic fixes from a review.
No telemetry. 11 flows can leave your machine.
Each one is declared in the registry with who receives what and when, and each starts only on a basis you control. BMAD+ never updates itself silently.
Update check
Opt-out
Asks npm which release is current. Turn it off with update-policy --mode off.
Update apply
Opt-in
Downloads the exact release you confirmed.
Python setup
Opt-in
Installs a pack’s pinned requirements with uv, pipx or pip.
Nexus worker
On invocation
Launches the command or Codex run your plan declares.
Nexus verifier
On invocation
Runs the check command your plan declares, without a shell.
Acceptance page
On invocation
Saves a tester’s verdicts to the address that serves the page.
OSINT providers
With your key
Sends queries to the providers whose keys you set, over HTTPS only.
SEO crawl
On invocation
Fetches the site you audit, with address checks against SSRF.
Google APIs
With your key
PageSpeed and CrUX, with the key in a header, never in a URL.
SEO screenshots
On invocation
A browser pinned to the address you checked, nothing else.
Short presentations with a voice-over. Every video has captions and a text transcript.
AI said “done.” Was it?
Duration 1:29 · Voice-over: English
How 39 agents rebuilt this site in under twelve hours, what the reviewers caught, and the five steps and commands to replay it yourself.
Transcript
AI said “done.” Was it? Thirty-nine agents rebuilt a site in under twelve hours. Reviewers found eleven false claims. They were wrong twice, too. You have probably been burned by “done” before. Here is how to check: five steps, five commands. One. How do you stop invented facts? Write the rules first: from scratch, facts from the code, simulated means labelled. Two. What should it ask before it starts? The intake lists what is known and what is missing. Three. How do agents stay out of each other’s way? One brief per page, one copy of the repo each. BMAD+ deals the briefs and caps each wave. Your assistant launches the agents. Four. Who checks the work? Five readers that wrote none of it, replaying commands on a clean server. And the gate says no until every file is accounted for. A promise that your files are kept. A command that never existed. A count off by one. Five. Fix, then prove: nineteen fix commits, 1,879 tests, zero accessibility violations. A person kept the last call. Nobody has measured yet whether readers find it clearer. You get five commands, three swarm files, the review gate and seven mistakes to avoid. Start with step one. “Done” is a claim. Replay it.
The intake questions, the budget of 26 and the approval step, waves of sub-agents, and a tally that says simulated units are not people.
Transcript
Before a pack acts, it asks. The intake puts a few questions at a time and marks the must-haves. Where you have no answer, it proposes a default, or records the question as unknown with its risk. The answers become a brief, and nothing in it is invented. Next, the swarm. By default, a run may use 26 sub-agents before the assistant must ask you. Ask for more, and the run waits, marked needs approval, until a person says yes. Once approved, sub-agents work in waves of ten by default, each reading one brief and writing one answer. The runtime validates every answer file and counts only the valid ones. This tally is demo data, and it opens with its disclosure: simulated units are not people. Read the rates as an upper bound, and confirm them with real people.
The design read, tokens and states, and a pre-flight verdict taken from the checklist.
Transcript
Atelier, the tenth pack, builds, audits and redesigns interfaces under checkable rules. It starts with a design read: questions about the surface, the audience and the one job the page must do. A dial you leave open is set from the surface, and the brief says it was inferred. Then it builds with semantic tokens, WCAG AA contrast and every state covered. It ends with a pre-flight. Ready or not ready is read from the checklist, not from how the screen looks. The installer copies its three skills, and doctor checks them.
A tour of the ten packs and their agents, with the count of 48 shown as 15 installed, 27 in Shield and 6 in Dev Studio.
Transcript
BMAD+ ships as ten packs. Core is required: Atlas for strategy, Forge for architecture and development, Sentinel for quality, Nexus for coordination. OSINT brings Shadow, for governed public-source research. Maker designs, validates and packages new agents. Shield covers compliance with 27 specialist agents and 11 workflows, from GDPR to ISO 27001. SEO adds three agents, Scout, Chief and Judge, and a six-phase audit. Memory brings Zecher, who keeps notes, decisions and session handoffs. Dev Studio has six roles and 38 workflows, from analysis to review. Backup makes timestamped backups, Animated builds scroll-driven websites from video, and Design brings Atelier. In all, that makes 48 agents.
One install command, a health check with doctor, updates that keep your edits, and an uninstall that leaves memory and configuration in place.
Transcript
Run one command from your project folder, and choose your adapters and packs. The installer writes the agent instructions, the shared project spine and the adapters you selected. Then doctor checks the installation and reports how many checks passed. Updates keep the files you customised: BMAD+ tracks which files it manages. A version check only looks. Automatic updates need an explicit project policy. Uninstall removes only files that still match what was installed. Your memory, your edits and your configuration stay where they are. It never deletes a folder recursively, so hidden files and nested content survive.
How Nexus finds independent tasks, gives each its own write scope, records real results and runs independent checks before accepting.
Transcript
Nexus coordinates the work of the other roles. It looks for independent tasks, the ones that touch no shared file. Each task gets its own write scope, and scopes may not overlap. Parallel agents come from your host tool. Without them, the tasks run in sequence. Nexus can also launch a planned local command or a Codex CLI process, under a foreground supervisor. It records the real result, then runs independent checks before anything is accepted. There is no background scheduler, and host permissions still apply. In autopilot, the same pipeline runs from idea to delivery, with the checkpoints you configure.
BMAD+ is a framework of agent instructions, skills, workflows and a CLI installer. It works inside compatible AI coding tools. You use the model, account and tool access you already have. It does not include a model subscription.
Do I need a special initialization command after installation?
Install BMAD+ in your project, then open your AI tool in that folder so it can read the generated instructions. Ask for bmad-help, call an agent by name, or describe your task. bmad-help, autopilot and parallel are conversation instructions for your AI assistant, not standalone shell commands.
What does “ask first” mean in practice?
A pack can declare an intake: questions it needs answered before it acts. The assistant asks them a few at a time, bmad-plus intake records each answer as answered, defaulted or unknown, and the rendered brief is what the work is held to. Today the Design pack ships one; the mechanism is available to any pack.
Does every tool execute agents in parallel?
The framework describes how to divide and coordinate work. Actual parallel execution and tool use depend on the host assistant. Installing an adapter does not add a background worker service or guarantee that every host supports the same execution features.
What is a swarm, and what does it cost?
A swarm is a set of sub-agents, each given one written brief, whose answers are validated and tallied by the CLI. The default budget is 26 sub-agents per run; a larger run is refused until you approve it. Your own AI tool does the model work under your own plan, and paid image, video or audio generation always asks first.
Can I trust the numbers a simulated panel gives me?
Treat them as an upper bound to confirm with real people. Simulated units are not people, their quotes are not testimonials, and every tally opens with that disclosure. A valid answer means well-formed, not true.
Can I use it in an existing project?
Yes. Run the installer from the existing project directory and select the relevant adapters and packs. Keep your usual version control workflow. Managed-file tracking helps the updater distinguish shipped files from local customizations.
What does BMAD+ send off my machine?
There is no telemetry. Every flow that can reach the network or start a program is declared, each on a stated basis: when you run a command, once you add a key, when you opt in, or on by default with one switch to stop it, as for the update check. A build gate fails when code makes a call that is not declared.
What happens to my files if I uninstall?
Uninstall removes only files that are still byte-identical to what BMAD+ shipped, plus its manifest. Your configuration, project memory, outputs and anything you edited are kept.
Which AI tools does it support?
Claude Code, Gemini CLI, Antigravity, Cursor, Codex CLI, OpenCode and Aider, through one AGENTS.md spine plus each tool’s own instruction file. Execution stays with the tool.
A team in your repository.One command away.
Open a terminal in your project folder. The installer helps you select packs and configure the adapters for your AI tools.
Run in your project directoryCurrent release v0.23.1
npx bmad-plus@0.23.1 install
Requires Node.js 20 or newer. Some optional packs have additional runtime or API requirements.
01
Install your foundation
Run the command, select your tools and choose the packs you need. Core is included.
02
Open your usual AI tool
Start a session in the project so your tool can load its generated adapter and the agent instructions.
03
Ask for a role or a plan
Say “Atlas, help me scope this project” or ask for bmad-help to discover the available agents and skills.