Vendors
The extension catalog — every vendor, the products they expose, and the capabilities they contribute
A vendor is a third-party (or first-party) system Scrydon talks to — Google, OpenAI, Salesforce, Mistral, the built-in scrydon core, and so on. Every block, tool, and runtime capability you can use in a workflow is exposed by a vendor.
The model is always the same: Vendor → Product → Capabilities (tools, blocks, triggers, runtime capabilities). There is no flat global tool catalog. If you don't see what you need here, you can author your own extension and upload it.
The model
Vendor e.g. google, openai, scrydon
└── Product e.g. gmail, sheets, drive — one per vendor area
├── Block workflow-editor card (one per product)
├── Tools individual operations the agent can call
├── Triggers webhook + polling triggers
└── Capabilities runtime contracts: LLM, STT, TTS, embeddings, image, video, OCR, webhooks, discoveryA workflow that "sends a Gmail" is using vendor=google → product=gmail → tool=send-mail. The "Gmail" card on the canvas is google → gmail → block. The platform's extension registry resolves all of this — you never hardcode vendor names in workflow code.
Catalog
Built-in (always available)
Productivity & Storage
Workspace + public APIs: Gmail, Sheets, Docs, Drive, Calendar, Search, Translate, YouTube, +11 more.
Microsoft 365: Excel, OneDrive, Outlook, Planner, SharePoint, Teams, Resource Discovery.
Jira, Confluence, Bitbucket Cloud — issues, wiki, source control.
Repositories, issues, pull requests, files, gists, workflows, releases, milestones, projects, search, webhooks.
CRM: accounts, contacts, leads, opportunities, cases, reports, SOQL.
Customer Engagement, Business Central, Finance & Operations.
Read and write SAP S/4HANA Cloud business data via OData.
AI & LLMs
Claude Opus, Sonnet, and Haiku via the Messages API.
GPT family + embeddings, image / video generation, STT, TTS, vision.
Chat, embeddings, Document AI (OCR), and Voxtral audio.
Per-org deployed LLMs, embeddings, images, audio.
Foundation models via the Converse API.
Local open-source LLMs and embeddings.
Self-hosted high-throughput inference.
Cloud
Manage Azure resources through the az CLI, governed at the wire.
BigQuery, Cloud Storage, Pub/Sub, Vertex AI, Speech, Vision, Translate, +5 more.
Observability
Isolation tiers
A vendor definition may declare how the platform isolates its archive at runtime:
defineExtension({
// ...
isolation: "microvm", // optional — defaults to "memory"
})memory(legacy default declaration) — the manifest does not strengthen the deployment/organization minimum. The shared resolver still binds execution tomicrovmormanaged_process; it never means an in-process Worker Thread.microvm— the archive requires a dedicated microVM. If that tier is not permitted, proven, and ready, execution is refused; there is no downgrade.
The manifest field is a minimum-isolation declaration, not a backend selector. Administrators can see the effective execution tier in the audit log.
Where vendors come from
| Source | What it produces |
|---|---|
First-party templates in packages/sdk-authoring/src/extensions/authoring/templates/ | The vendors on this page — shipped with the platform. |
| Custom extensions built with the Extensions Authoring SDK | Uploadable .archive.tar.gz archives that register a new vendor at runtime. No platform redeploy. |
Both paths produce the same kind of artifact (a defineExtension() definition); the only difference is where the archive is stored.
Capability resolution (LLM, STT, TTS, embeddings, …)
Capabilities are resolved through the extension registry, in this order:
- Per-call override (e.g., a block setting that pins one model)
- Org policy (defaults configured by an admin)
- Automatic pick across installed candidates (first that matches the requested capability)
null— the platform answers409 CAPABILITY_NOT_CONFIGURED("no provider configured")
Capabilities resolve through your installed extensions — no vendor is hardcoded and no silent fallback occurs. Speech-to-text requests are brokered server-side by the platform (credentials never leave the server); other capabilities (LLM, embedding, OCR) resolve through your installed extensions at request time. Never import a vendor SDK directly in product code, never hardcode a vendor branch, never silently fall back to a vendor the org didn't install.