Cowork on 3P feature matrix — what's in, what's out, per route — The Cowork Bible — Tinkso

TL;DR.

First-party hosted Cowork has the full feature set; new features ship here first.
Bedrock and Vertex AI routes lose three things — the Chat tab, Computer Use, and the Skills Marketplace. Everything else (full Cowork agentic loop, Code tab, projects, artifacts, memory, file upload/export, MCP connectors, plugins, customer-deployed skills) is preserved.
Microsoft Foundry route keeps a feature set closer to first-party because inference still terminates at Anthropic infrastructure. Feature parity moves fastest on this route — verify at evaluation time.
Custom gateway route depends on what the customer's gateway routes to.

How to read this matrix#

Symbol	Meaning
✅	Supported and reliable
⚠️	Supported with caveats — read the notes
❌	Not supported in this deployment route
🧩	Supported via customer-deployed extension (skill, plugin, MCP connector)
🔬	Research Preview status; ships first-party first, may lag on 3P

The matrix re-verifies against Anthropic's 3P feature-matrix doc at each last_reviewed cycle. Trust the source over this page if a discrepancy appears.

Tabs in Claude Desktop#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
Cowork tab	✅	✅	✅	✅	✅ (if gateway compatible)
Code tab	✅	✅	✅	✅	✅
Chat tab	✅	❌	❌	⚠️ verify	❌
Computer Use (research preview)	🔬	❌	❌	⚠️ verify	❌

Cowork agentic capabilities#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
Multi-step agentic loops	✅	✅	✅	✅	✅
Sub-agent coordination	✅	✅	✅	✅	✅
File creation in granted folders	✅	✅	✅	✅	✅
File upload / export	✅	✅	✅	✅	✅
Projects	✅	✅	✅	✅	✅
Artifacts	✅	✅	✅	✅	✅
Memory	✅	✅	✅	✅	✅

Extensions#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
MCP connectors (customer-deployed)	✅	✅	✅	✅	✅
Plugins (customer-deployed)	✅	✅	✅	✅	✅
Skills (customer-deployed)	✅	✅	✅	✅	✅
Skills Marketplace (in-app discovery / install)	✅	❌	❌	⚠️ verify	❌
Hooks	✅	✅	✅	✅	✅
Remote connectors	✅	✅	✅	✅	⚠️ verify
Web search / web fetch	✅	✅ (egress allowlist)	✅ (egress allowlist)	✅	✅ (gateway-controlled)

Identity, admin, and observability#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
Anthropic account login	✅	❌	❌	⚠️ verify	❌
Local device identity	⚠️ optional	✅	✅	⚠️ verify	✅
In-app admin console (claude.ai)	✅	❌	❌	⚠️ verify	❌
MDM-managed configuration	✅	✅	✅	✅	✅
SSO / SCIM (Team / Enterprise plans)	✅	❌ (use OS-level SSO)	❌ (use OS-level SSO)	⚠️ verify	❌
Usage analytics in claude.ai	✅	❌	❌	⚠️ verify	❌
OpenTelemetry export of session activity	✅	✅	✅	✅	✅
Anthropic-bound telemetry (crash, analytics, updates)	✅ on	⚠️ disable-able	⚠️ disable-able	✅ on	⚠️ disable-able

Compliance and residency#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
Conversation data leaves customer tenant	Yes — to Anthropic	No	No	Yes — to Anthropic	Depends on gateway
In-region data residency	⚠️ Enterprise contract negotiation	✅ via cloud region	✅ via cloud region	❌ (residency follows Anthropic)	Depends on gateway
FedRAMP / ITAR / sovereign cloud compatible	❌	✅ in authorised regions	✅ in authorised regions	❌	Depends on gateway
Zero-retention contracting	⚠️ Enterprise contract	✅	✅	⚠️ via Anthropic terms	Depends on gateway

Billing#

Capability	First-party hosted	Bedrock	Vertex AI	Microsoft Foundry	Custom gateway
Anthropic seat licensing	✅ Pro / Max / Team / Enterprise	❌	❌	⚠️ via Microsoft	Customer-defined
Consumption-based via cloud provider	❌	✅ AWS	✅ GCP	✅ Azure	Customer-defined
Unified bill with cloud commitments / credits	❌	✅	✅	✅	Customer-defined

Reading the matrix as a buyer#

Three takeaways the matrix is designed to surface.

One. First-party hosted Cowork has the broadest feature set. If the customer's compliance program accepts that posture, hosted is faster, simpler, and ships features sooner. Most general mid-market US/UK customers should stay there.

Two. Bedrock and Vertex AI are functionally equivalent from a Cowork-feature standpoint. The choice between them is a strategic-cloud question, not a Cowork question.

Three. Microsoft Foundry's row in the residency section is the one that catches buyers out. Foundry leaves conversation data flowing to Anthropic — the Compliance row reads "Yes — to Anthropic" for that route, deliberately. Buyers evaluating Foundry for residency reasons have made an early-stage procurement mistake; they should be redirected to Bedrock, Vertex AI, or first-party Enterprise. See Cowork on Microsoft Foundry for the longer write-up.

What this matrix doesn't cover#

Per-feature performance differences between routes (latency, throughput, cost per token). Those depend on model, region, and workload — out of scope for a feature matrix.
Model availability by region on Bedrock / Vertex / Foundry. Verify against the cloud provider's current model-availability documentation at the time of evaluation; the model lineup moves.
The Code tab's specific feature parity with first-party on each route. The 3P feature matrix at Anthropic has the granular Code-tab table; consult it directly if Code-tab parity is decision-critical.

Tinkso's take#

We use this matrix as the artifact in any "should we go 3P?" conversation. The headline call almost always falls out of three rows: the Compliance row (does the customer's program accept hosted?), the Skills Marketplace row (does the pilot rely on marketplace-discovered skills?), and the Computer Use row (does the planned use case need it?). When the answers are yes, yes, no, the customer should stay first-party. When they're no, can ship our own, no, 3P is the right call.

Try this#

Print this page, sit down with the IT lead and a function head, and walk row by row. Every row gets a "this matters / this doesn't / this needs verification" annotation. The result is a one-page deployment-route decision artifact — and it's the input we use for the procurement requirement document on every Tinkso 3P engagement.