Tool Integration

Tool Integration is how agents touch the world: authenticated APIs, browser computer-use, MCP-shaped tool servers, and catalogs of prebuilt actions. It sits under Agentic next to Frameworks (which call tools) and Agent Protocols (which standardize discovery/calling). Without it, frameworks only talk to themselves.

Technically the cell cares about auth scoping, schema stability, idempotency, and audit of side effects. Failure modes: tools that mute errors, OAuth tokens wider than the agent’s mandate, and browser-use paths that cannot be replayed for dispute. In a company OS, every tool call should be attributable to a case and reversible or compensatable under policy.

Members (few but sharp):

• Composio catalogs SaaS actions agents can call with less per-app glue.
• Arcade MCP leans into MCP-shaped tool hosting.
• Browser Use exposes browser computer-use for agents that must operate web UIs.

Prefer catalog connectors when the job is many SaaS verbs; prefer browser-use when the system of record has no API; prefer protocol-first packing when you standardize tool surfaces across runtimes. Feature maps such as Integrations & Tool Calling, APIs, SDKs & MCP Extensibility, and Browser & Computer Use name the contract.

Navigate to Frameworks for the control loop, Agentic Gateways for the front door, and IAM when identity of the acting principal is the missing piece. Stay here when the bottleneck is binding agents to real external systems safely.

Every tool should declare blast radius: read-only, draft, or mutate. Mutating tools need idempotency keys and compensation paths; read tools need pagination and timeouts. Prefer MCP or stable OpenAPI-derived schemas so frameworks do not hard-code vendor quirks. Browser computer-use is a last resort when APIs are absent — record the session artifact so disputes are evidence, not he-said. Review OAuth scopes quarterly; agents accumulate privileges the way humans accumulate Monday boards.

For this research catalog the acceptance test is boring and strict: after wiring the pack, can a second person replay the same scenario and get the same measurable outcome? If not, the page is still a brochure. Document the scenario, the expected number, and the log line that proves it. That habit is what separates a deepened taxonomy cell from a landscape screenshot.

Also record what you refused: tools not connected, corpora not ingested, metrics not collected. Refusals keep the cell honest when marketing pages claim universality. A taxonomy essay that only lists winners teaches less than one that names the skip conditions next to the install path.

Keep the essay technical enough that an engineer can argue install order — and short enough that a reader finishes it before opening the right-rail member list.