Marvin Toolkit
real-case/marvin-toolkit
Toolkit completo de dev lifecycle para Claude Code com workflows estruturados
Ver no GitHubMCP server que integra um ciclo de desenvolvimento completo no Claude Code através de um único plugin. Oferece 60 prompts, 16 ferramentas MCP, 10 agentes e 10 widgets interativos organizados em comandos estruturados para commit, review, segurança, documentação e deploy de código.
Instalação
npx convez-ai real-case/marvin-toolkit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar fluxo de commit com validações de segurança e credenciais
- Executar análises de segurança e refatoração de código assistidas
- Orquestrar pipeline de tarefas e board visual para gestão de projeto
- Gerar documentação e conformidade durante o desenvolvimento
Veredito

Marvin é uma solução madura e bem documentada que consolida ferramentas essenciais de DevOps em um único MCP server. Destaca-se pela segurança built-in (bloqueio de credenciais, force-push em branches protegidas) e triple entry point (chat, slash command, MCP prompt), apesar de exigir Node.js 20+ e estar fortemente acoplado ao ecossistema Anthropic.
README
A tradução para o português deste README ainda está na fila.
Marvin
Claude Code toolkit for those who don't panic.
Marvin is a Claude Code plugin that
packages the whole development lifecycle as one plugin, one MCP server, and one slash
prefix — /marvin:. Install it and you get structured, repeatable workflows for
committing, reviewing, securing, documenting, and shipping code, all inside Claude Code.
Under the hood it ships 60 prompts, 16 MCP tools, 10 agents, and 10 interactive widgets
across seven command groups, built on a TypeScript MCP server that runs on Node.js 20 or
later.
Call it your way
Each workflow is authored once at plugins/marvin/skills/<command>/SKILL.md, and three
independent entry points reach the same body:
- Chat. Describe what you want —
commit my changes,сделай коммит— and Claude Code matches the skill by its frontmatter description. - Markdown slash command. Type the terse form, such as
/commitor/sec-scan. - MCP prompt. Type the namespaced form, such as
/marvin:commit, served by the bundled server.
Pick whichever feels right; they all run the same skill. The architecture tour explains how the three entry points resolve, with diagrams.
Install
Run these two commands inside Claude Code:
/plugin marketplace add real-case/marvin-toolkit
/plugin install marvin@marvin-toolkit
The plugin registers one MCP server named marvin, and its commands appear as
/marvin:<group>-<command>. No clone or build step is needed, because the server ships
bundled.
Installing also lets Marvin block a shell command. Two hooks arrive with the plugin and
are armed with no enablement step: one refuses a commit that skips your local gates
(git commit --no-verify) and a force-push or deletion aimed at a protected branch, the
other refuses a commit whose added lines carry a credential-shaped string. Everything else
runs untouched — a quoted --no-verify in a commit message is not a match, a dry-run push
is never blocked, and any internal error allows the command rather than blocking it. To turn
them off, set "hooks": { "enabled": false } in .marvin/config.json, or export
MARVIN_HOOKS_DISABLED=1 for one session. The
configuration reference has the details.
Documentation
- Getting started — install the plugin, then run the guided
/marvin:onboardwalkthrough. - Usage guide — worked walkthroughs for committing, the task pipeline, the task board, security, and refactoring.
- Configuration — the
.marvin/working directory,.marvin/config.json, and theMARVIN_*environment variables. - Command reference — every
/marvin:command with a synopsis and the phrases that invoke it. - Architecture — how the plugin is built, with diagrams.
The command groups
Commands follow the pattern /marvin:<group>-<command>, and singletons stay bare. The 60
prompts divide into eight groups:
| Group | Purpose | Count |
|---|---|---|
| (bare) | Core developer tools | 18 |
adr-* | ADR lifecycle | 6 |
pr-* | Pull-request operations | 4 |
task-* | Spec-driven task pipeline | 7 |
sec-* | Security scanners | 11 |
refactor-* | Code-health family (read, plan, apply) | 4 |
audit-* | Formal project audit (plan, run one, consolidate) | 3 |
track-* | Lightweight task tracker | 7 |
The tables below give a one-line synopsis per command. The command reference adds the natural-language phrases that invoke each one from chat.
Core developer tools
These are language-agnostic and used by every engineer. 👤 = human-run: the model never auto-triggers these, you invoke them yourself.
| Command | Description |
|---|---|
/marvin:onboard | Guided first session in this project — every write behind an explicit yes. |
/marvin:commit | Draft a Conventional Commits message with sensitive-file detection. |
/marvin:debug | Systematic root-cause analysis with hypotheses. |
/marvin:adr | Draft an Architecture Decision Record; drafts land proposed. |
/marvin:changelog | Generate a changelog from git history. |
/marvin:readme | Generate or update README.md from codebase analysis. |
/marvin:migration-plan 👤 | Plan a migration with risks and a rollback strategy. |
/marvin:explain | Explain code, architecture, and execution flow. |
/marvin:docs-search | Search and synthesize the project documentation. |
/marvin:handoff | Capture the session context into .marvin/handoff/ to continue later. |
/marvin:handoff-list | List the session-continuation handoff documents, newest first. |
/marvin:lessons | Browse the lessons-learned store — search, add, stats, and prune. |
/marvin:help | Show the project dashboard and the full command index. |
/marvin:dashboard | Report the whole-toolbox state at a glance. |
/marvin:reports | Browse every generated .marvin/ report, newest first, with freshness. |
/marvin:report-export | Export a report to PDF, print-ready HTML, or a Markdown digest (ADR-0033). |
/marvin:widget-preview | Open a widget as a rendered panel with this project's data, on any host (ADR-0034). |
/marvin:autopilot | Deliver a task to a merge-ready PR through the autonomous pipeline; this session only answers its judgments. |
The marvin-guide, marvin-researcher, and marvin-debugger agents support these
commands. The plugin also registers two external MCP servers: context7 for library docs
and gitmcp for GitHub repository docs.
ADR lifecycle — adr-*
These wrap the full decision-record lifecycle around the bare /marvin:adr create command
(ADR-0027). Ratification, rollback, and
project-memory sync are reserved for a person, marked below with 👤.
| Command | Description |
|---|---|
/marvin:adr-review | Deeply review one proposed record and return a verdict. |
/marvin:adr-accept 👤 | Ratify proposed to accepted through the readiness gate. |
/marvin:adr-audit | Lint the corpus read-only, with remediation guidance. |
/marvin:adr-coverage | Analyze gaps between recorded decisions and the actual stack. |
/marvin:adr-supersede 👤 | Roll a decision back through a paired successor record. |
/marvin:adr-sync 👤 | Regenerate the accepted-decisions digest in CLAUDE.md. |
Pull-request lifecycle — pr-*
These cover the pull request from open to merge.
| Command | Description |
|---|---|
/marvin:pr-create | Open a PR with a structured description and pre-flight checks. |
/marvin:pr-review | Review a PR and post severity-tagged inline comments. |
/marvin:pr-resolve | Work through unresolved review threads — fix, push, reply, resolve. |
/marvin:pr-merge | Merge a PR, then check out the base branch and pull. |
Spec-driven task pipeline — task-*
These separate the human decisions in a spec from the automated execution that follows.
| Command | Description |
|---|---|
/marvin:task-start | Co-create a spec through a dialogue, then a Definition-of-Ready gate. |
/marvin:task-implement | Execute a ready spec in the current session. |
/marvin:task-verify | Run the quality gates with stack auto-detection. |
/marvin:task-deliver | Commit and open a PR, gated on verification passing. |
/marvin:task-summary | Aggregate a finished task into one delivery summary. |
/marvin:task-audit | Lint the spec corpus for consistency; read-only, with remediation notes. |
The marvin-tm-writer, marvin-tm-executor, marvin-tm-spec-critic,
marvin-tm-diff-critic, and marvin-tm-review-fixer agents support this pipeline.
Security scanners — sec-*
These cover OWASP-aligned scanning, threat modeling, and remediation.
| Command | Description |
|---|---|
/marvin:sec-scan | Run a comprehensive OWASP Top 10:2025 audit. |
/marvin:sec-secrets | Scan for leaked secrets across code, config, and git history. |
/marvin:sec-deps | Audit dependencies for CVEs, license risks, and maintenance health. |
/marvin:sec-gate | Run a fast, diff-scoped pre-commit security check. |
/marvin:sec-threat-model | Build a STRIDE threat model for a feature or system. |
/marvin:sec-iac | Review Infrastructure-as-Code across Terraform, Kubernetes, and Docker. |
/marvin:sec-ci | Audit CI/CD pipelines for supply-chain risks. |
/marvin:sec-fix | Generate and verify a tested patch for a vulnerability. |
/marvin:sec-compliance | Check code against OWASP ASVS at L1, L2, or L3. |
/marvin:sec-pentest | Generate an application-specific penetration-testing checklist. |
/marvin:sec-report | List the structured scanner findings by severity. |
The marvin-auditor agent supports these commands.
Refactoring — refactor-*
These form the code-health family (ADR-0029), split by mutation into read, plan, and apply.
| Command | Description |
|---|---|
/marvin:refactor-audit | Run a whole-project structural audit into a findings register. |
/marvin:refactor-smells | Scan a path, module, or diff for smells and anti-patterns. |
/marvin:refactor-plan | Sequence selected findings into small, risk-annotated steps. |
/marvin:refactor-apply | Execute one behavior-preserving step behind the verify gate. |
The marvin-refactor-auditor agent supports these commands.
Task tracker — track-*
These drive a lightweight per-project board with interactive forms
(ADR-0025). New tasks branch off following the
convention <type-prefix>/<seq>[-<tracker>]--<slug>, and committing or opening a PR for a
board task goes through the board-aware /marvin:commit and /marvin:pr-create, which
pick up the linked task automatically.
| Command | Description |
|---|---|
/marvin:track-menu | Open the main menu — every board action, including link-pr and archive. |
/marvin:track-new | Create a task — bug, feature, chore, or spike. |
/marvin:track-list | List the board: all tasks by status, the work-in-progress view, or the tracked tasks linking out. |
/marvin:track-show | Show one task in full. |
/marvin:track-start | Pick a todo task, branch off, and mark it in progress. |
/marvin:track-move | Move a task — to review, done, or any configured status. |
/marvin:track-config | Show or edit the board configuration. |
The board dashboard and command reference stay one call away as /marvin:help track
(ADR-0032 records the seven-command
surface).
Statuses are project data (ADR-0026):
configure your tracker's vocabulary through /marvin:track-config, and the lifecycle
commands drive it by role. The configuration reference documents
every setting.
Interactive widgets
On an MCP host that supports the Apps widget layer, ten widgets render an interactive
panel in addition to the text output
(ADR-0024). The panel is additive, so a
text-only host shows the same information as text and no command depends on a rich host.
The widget-backed commands are /marvin:track-list (which fronts both the board list and
the tracked link-out view), /marvin:track-show, /marvin:task-summary,
/marvin:sec-report, /marvin:handoff-list, /marvin:dashboard, /marvin:help, and
/marvin:reports, and /marvin:autopilot shows its run as a status card at least every five
minutes (ADR-0051).
Development lifecycle
The core, pr-*, and sec-* commands map onto the everyday flow from planning a change to
shipping it.
flowchart LR
Plan --> Code --> Review --> Secure --> Document --> Ship
| Phase | Commands |
|---|---|
| Plan | adr, migration-plan |
| Code | debug, explain, docs-search, refactor-* |
| Review | pr-review, refactor-smells |
| Secure | sec-scan, sec-secrets, sec-deps, sec-gate |
| Document | readme, changelog |
| Ship | commit, pr-create, pr-resolve, pr-merge |
Layer the track-* tracker on top of any of these for day-to-day tracking, and run
/marvin:dashboard to see the whole toolbox's state at a glance.
Architecture decisions
Decisions with long-lived consequences are recorded as ADRs under docs/adr/. The two pre-consolidation ADRs, covering the source format and the MCP-first stance, were retired in the publication cut, and their still-relevant rationale is folded into 0001, 0013, and 0018.
| ADR | Decision | Status |
|---|---|---|
| 0001 | Single-plugin consolidation under one /marvin: prefix | Accepted |
| 0002 | Tool-backed verification gate | Accepted |
| 0003 | Tool-backed Definition-of-Ready gate | Accepted |
| 0004 | Traceable spec contract and gate reordering | Accepted |
| 0005 | Portable, host-adaptive spec contract | Accepted |
| 0006 | All subagents on Opus; economy via deterministic tools | Accepted |
| 0007 | Unified .marvin/ working directory | Accepted |
| 0008 | MCP-door plugin-resource resolution | Accepted |
| 0009 | Config-first gate resolution for verify | Accepted |
| 0010 | Tool-backed contract-seal verification | Accepted |
| 0011 | Tool-backed scope-allowlist gate | Accepted |
| 0012 | Tool-backed delivery gate | Accepted |
| 0013 | Self-contained committed server bundle | Accepted |
| 0014 | Distribution and release model (git tag to GitHub Release; no npm) | Accepted |
| 0015 | verify shell-execution trust boundary | Accepted |
| 0016 | Bundled external MCP dependencies (context7, gitmcp) | Accepted |
| 0017 | Adversarial critic gates in the task pipeline | Accepted |
| 0018 | Three doors and the instrument taxonomy | Accepted |
| 0019 | Branching model: release main, integration dev, changes via PRs | Accepted |
| 0020 | Root-cause analysis as the marvin-debugger agent | Accepted |
| 0021 | Tool-backed lessons-learned feedback loop | Accepted |
| 0022 | Numeric-prefixed spec filenames | Accepted |
| 0023 | Unified pr-* pull-request command family | Accepted |
| 0024 | MCP Apps widget layer: data-first staging and shared data contracts | Accepted |
| 0025 | Kanban goes board-only; git ops fold into the commit and pr-create skills | Accepted |
| 0026 | Configurable status model: statuses are project data, roles stay closed | Accepted |
| 0027 | Tool-backed ADR lifecycle | Accepted |
| 0028 | Lessons v2: hygiene surface and recall or capture expansion | Accepted |
| 0029 | Refactoring command family: read, plan, apply under hard rails | Accepted |
| 0030 | Toolbox dashboard and local usage log | Accepted |
| 0031 | Rename the kanban-* command group to track-* | Accepted |
| 0032 | Reduce the track-* surface to seven commands | Accepted |
| 0033 | Report export is template-only (Claude fills a shipped print template) | Accepted |
| 0034 | A local preview door renders widgets on hosts that cannot | Accepted |
| 0035 | A verification is bound to the tree it verified | Accepted |
| 0036 | Acceptance oracles are executed and journalled, not narrated | Accepted |
| 0037 | Spec corpus mechanics are tool-backed | Accepted |
| 0038 | A finding has an identity that survives the next scan | Accepted |
| 0039 | A critic's verdict is a receipt on disk, not a sentence in a transcript | Accepted |
| 0040 | Enforcement runs before the call, as a plugin hook | Accepted |
| 0041 | A widget-bound tool's text is gated on the client's capabilities | Accepted |
| 0042 | Bounded critique, and verification before review | Accepted |
| 0043 | Task workflow metrics are a committed per-task series, derived at delivery | Accepted |
| 0044 | The metrics record is written by two gates, not by prose | Accepted |
| 0045 | By-product files are exempted from a task's scope by project configuration | Accepted |
| 0046 | A spec keeps only the sections a reader uses, and states each fact once | Accepted |
| 0047 | A feature records one red phase, and the delivery gate reads it | Accepted |
| 0048 | A test_one placeholder is shell-quoted by marvin, not by the template author | Accepted |
| 0049 | The seal stage reads test_one for a whole file instead of refusing its per-test parts | Accepted |
| 0050 | One calibration file per finished run, not one appended log | Accepted |
| 0051 | The autopilot reports at least every five minutes, as one status card | Accepted |
| 0052 | The delivery gate checks the contract's declared names against the code | Proposed |
Contributing
Contributions are welcome. Every change must pass the same quality gates CI runs:
npm run lint # ESLint over the TypeScript source
npm run format:check # Prettier
npm run lint:manifests # marketplace and plugin manifest structure
npm run lint:docs # ADR coverage and working-directory paths
npm run eval:self-test # trigger-eval harness guard (no network)
npm run eval:trigger # mock sweep over every trigger dataset
npm run build # build every workspace
npm run test # Node.js native test suites
npm run verify-dist # committed dist/ matches a fresh build
See CONTRIBUTING.md for the full setup and workflow, and CLAUDE.md for the architecture reference.
Security
If you find a vulnerability, please report it privately, as described in SECURITY.md.
License
MIT © Yurii Anichkin