People don't understand your product, so they don't care.

GMaestro is the G-stack for technical content. Brief it on your company — a website URL, a file, a folder, anything — and pick where to publish: your website, Reddit, X, LinkedIn, you name it. Ten specialists research your space, draft in your voice, optimize for SEO and GEO, and land an approval card on the dashboard. You approve once; the dispatcher ships it.

Requires Node 22+ (.nvmrc) and pnpm.

git clone https://github.com/sebtsang/gmaestro
cd gmaestro
pnpm install
pnpm db:migrate              # creates ~/.gmaestro/gmaestro.db
pnpm gmaestro setup          # interactive — keys + Composio
pnpm dev                     # dashboard at http://localhost:3000

Fill the 3-input form on the dashboard (company URL + docs URL + destination), hit Run, watch the team work.

Skip every cloud setup. The dashboard ships with a mock driver that produces a fake DAG, fake personas, and a fake approval card so you can click through the full flow without burning a token:

NEXT_PUBLIC_USE_MOCKS=1 pnpm dev

Three paths to give GMaestro an LLM. All work; pick one.

The OAuth path is the cheapest first run if you already have Claude Pro or Max. The token is long-lived, billed against your subscription, and the SDK picks it up via the bundled claude-code binary. Composio integrations (Firecrawl, Reddit, LinkedIn, etc.) need a separate COMPOSIO_API_KEY — pnpm gmaestro setup walks you through it.

One Conductor plans the workflow. Three department heads sub-agent the planning. Ten specialists do the work. Composio executes — but only after the founder approves.

flowchart TB
    Prompt(["company URL · docs URL · destination"])
    Prompt --> Conductor

    Conductor[/"<b>Conductor</b>"/]
    Conductor --> CM[content-mgr]
    Conductor --> DM[distribution-mgr]
    Conductor --> IM[insight-mgr]

    CM --> Researcher
    CM --> Strategist
    CM --> Writer
    CM --> GEOEditor[GEO-Editor]
    CM --> Formatter

    DM --> PipelineReporter[Pipeline Reporter]
    DM --> SlackDigest[Slack Digest]

    IM --> FeedbackTagger[Feedback Tagger]
    IM --> ThemeSynth[Theme Synthesizer]
    IM --> LinearFiler[Linear Filer]

    Formatter -.->|approval gate| Composio
    Composio[("Composio: GitHub · Reddit · LinkedIn · Notion · Slack")]

    classDef mgr fill:#1f2937,stroke:#60a5fa,color:#fff
    classDef gate fill:#0f172a,stroke:#fbbf24,color:#fff
    class CM,DM,IM mgr
    class Composio gate

Personas reason in pure LLM over local data — your docs, your blog (for voice), Reddit/X/Perplexity (for signal). External writes only happen post-approval through the deterministic dispatcher. No persona has live tool access; this eliminates tool-selection hallucination on smaller models.

MIT.