TECHNIQUE

Tool registries & MCP

Tool Use & Structured Output

2APPLICATIONS
5OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 7 OPERATORS

Across the seven deployed operators, tool registries and MCP are used as a controlled tool-access layer: MCP servers/gateways, tool managers, plugins, playbooks, and skills expose internal systems, third-party apps, and retrieval indexes to agents while narrowing tool surfaces and adding auth, policy, and observability.

Observed Practices

Operators put a named tool layer between agents/LLMs and systems: MCP servers or gateways, tool managers, plugin frameworks, local MCP servers, or tool definitions.

7 of 7 operators
AirbnbArcade.devDropboxGrabIberdrolaLinkedInUber

Operators use tools to give agents access to internal systems, external APIs, third-party apps, knowledge sources, or live product/design data rather than relying on model-only context.

7 of 7 operators
AirbnbArcade.devDropboxGrabIberdrolaLinkedInUber

Several operators deliberately reduce the visible tool surface with single tools, meta-tools, consolidated frameworks, or gateways in front of many underlying systems.

5 of 7 operators
DropboxGrabIberdrolaLinkedInUber

Operators attach authentication, authorization, ACLs, credential handling, identity, or policy enforcement to tool access.

6 of 7 operators
Arcade.devDropboxGrabIberdrolaLinkedInUber

Operators let agents or planners decide which tools to invoke and in what sequence for multi-step work.

6 of 7 operators
AirbnbDropboxGrabIberdrolaLinkedInUber

Operators instrument tool, playbook, agent, or gateway execution for observability, audits, production operation, or roadmap decisions.

4 of 7 operators
AirbnbIberdrolaLinkedInUber

Some operators package reusable procedural knowledge as tools, playbooks, skills, workflows, or instruction files with schemas, rules, or templates.

4 of 7 operators
AirbnbGrabLinkedInUber

Where Operators Converge

Every observed operator uses an explicit tool abstraction layer rather than exposing raw systems directly to the model; the layer is implemented as MCP, a gateway, a tool manager, plugins, playbooks, or skills depending on the operator.

Every observed operator uses the tool layer to bring proprietary or live operational context into agent workflows: support systems, work content, internal services, design files, ITSM tickets, developer systems, or external APIs.

Where Operators Diverge

Operators differ in how they shape the tool catalog exposed to the agent.

APPROACH 01

Collapse many retrieval or backend options behind one tool or a small set of meta-tools.

DropboxLinkedIn

APPROACH 02

Consolidate plugins or scattered capabilities into managed tool frameworks while preserving modular app-building patterns.

AirbnbGrab

APPROACH 03

Use a central gateway that fronts existing APIs/tools and becomes the integration or policy point.

IberdrolaUber

Operators differ in where the MCP or tool server runs.

APPROACH 01

Local developer or desktop integration: IDEs or local apps connect to a local MCP bridge/server.

LinkedInUber

APPROACH 02

Central server or gateway integration: the MCP server/gateway sits in the platform path to tools, retrieval, or external services.

Arcade.devDropboxGrabIberdrolaUber

APPROACH 03

Platform-level tool manager without an MCP-specific server claim in the evidence.

Airbnb

Operators differ in where authentication, authorization, and sensitive credential handling live.

APPROACH 01

Browser/user-consent flow with the server managing tokens and scopes for external APIs.

Arcade.dev

APPROACH 02

Enterprise identity and gateway enforcement using OAuth, JWTs, provenance, or tool-level policy checks.

IberdrolaUber

APPROACH 03

Local developer authentication with credentials stored in the OS keychain.

LinkedIn

APPROACH 04

User-scoped retrieval or user-defined endpoint credentials inside the app platform.

DropboxGrab

Watch Items

Tool and context bloat is an admitted failure mode: Dropbox says adding tools creates context-management tradeoffs and emphasizes trimming context; LinkedIn says a giant tool list creates context bloat and hurts accuracy.

Credential and identity handling around MCP/tool calls remains a production risk: Arcade.dev reports MCP lacks a secure third-party credential-gathering path, and Uber reports missing attribution makes security controls, audits, and incident response harder.

Operators repeatedly add guardrails, redaction, ACLs, or policy checks because tool-enabled agents can touch sensitive systems and data.

Freshness and grounding remain operational concerns: Grab added auto-updates for Knowledge Vaults, Dropbox built unified indexes/graphs over third-party work content, and Uber’s uSpec reads live Figma data because manual specs drift out of date.

Production operators treat tool invocation as something that must be observable; Airbnb, LinkedIn, Iberdrola, and Uber all cite tool/agent observability, telemetry, traces, or invocation logging.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
Model Context Protocol serverspatternestablished
OpenAPI-derived toolsetspatternestablished
03

Observed in Production

2 APPS