Skill vs MCP vs API

Agent Analytics exposes one analytics surface through four real access paths:

Skill for agent environments that already support skills and command execution
MCP for chat-native and editor-native tool use
CLI for shell-oriented agent workflows
API for raw HTTP control

The product model does not change between them. Projects, analytics reads, and experiment operations stay the same; only the native entrypoint changes.

Recommended path by environment

Environment	Recommended path	Why
Claude Code	Skill first	Keeps the workflow agent-native without requiring raw MCP
Claude Desktop / Cowork	Hosted MCP	Best fit for connector-style chat tools with native tool calls
OpenWork	Skill first	Best fit when you want workspace-local skills, automations, and optional MCP fallback in the same app
Cursor	Skill + CLI first	Usually lower overhead than MCP when the agent can already run commands
OpenAI Codex	Hosted MCP or skill + CLI	Use native tool calls and supported views, or keep a terminal workflow; see Codex
ChatGPT	Hosted MCP in developer mode	Account and workspace policy must allow developer mode; see ChatGPT
OpenClaw	Skill first	Cleanest path when OpenClaw owns the scheduled analytics job from chat
Instinct	CLI with detached approval	Reuse a saved session for scheduled analytics briefs; see the Instinct guide
Custom runtime or internal agent	API	Best fit when you own retries, parsing, and orchestration

When to use each path

Skill

Use a skill when your agent already supports skills and can execute commands in the same environment.

A skill is usually the best fit when:

you want a guided workflow layer around common analytics tasks
your agent already has terminal access
you want to stay in the agent’s native loop instead of switching to tool-call-heavy MCP flows

MCP

Use MCP when your AI agent already runs inside a tool that supports connectors or MCP servers.

MCP is usually the best fit when:

you want the install to feel native inside chat
you want tool calls instead of shell commands
you do not want to hand-roll auth headers or request payloads
you want quick project or account summaries through structured tool responses
you want agent-readable reports such as analytics_paths, where the tool response includes both compact text and structured data

Tradeoff:

MCP often adds more latency and token overhead than skill + CLI flows because the model has to manage more tool-call round trips and tool result payloads.

Install the hosted MCP server

Server name: Agent Analytics
Server URL: https://mcp.agentanalytics.sh/mcp
Transport: Streamable HTTP
Authentication: browser sign-in with GitHub or Google

In an MCP-compatible app, add a remote server or custom connector using the URL above. Complete the browser sign-in with the identity that owns your Agent Analytics projects. Your client must support remote Streamable HTTP and OAuth authentication; you do not need to paste an account API key into the connector.

For Claude Code, register the server with:

claude mcp add agent-analytics --transport http https://mcp.agentanalytics.sh/mcp

Then open /mcp in Claude Code and complete authentication for Agent Analytics. For the connector UI walkthrough, see Claude Desktop / Cowork installation.

For Codex’s OAuth commands or ChatGPT’s developer-mode steps, use the dedicated Codex and ChatGPT guides. Interactive views require host support for MCP Apps. Local plugin testing is separate from the public OpenAI Plugins Directory, where Agent Analytics is not currently listed.

Verify the connection by asking your agent:

“List my Agent Analytics projects.”
“Show me a seven-day analytics overview for my project.”

If no projects appear, check that you signed in with the correct account. If you have not created a project or sent events yet, continue with First Project in 5 Minutes.

CLI

Use the CLI when your AI agent already has terminal access and is comfortable executing commands.

CLI is usually the best fit when:

your AI agent already lives in a shell-first environment
you want predictable command output
you prefer command composition over tool integration
you want lower overhead than MCP in editor-style agents like Cursor
you want simple local auth helpers like login and logout around the same API
you want shell-readable commands such as paths that summarize entry pages, exit pages, terminal labels, and next-step analysis

For install, login flow, common commands, and CLI-to-API mapping, continue to the dedicated CLI page.

API

Use the API when you want strict control over requests, retries, and response parsing.

API is usually the best fit when:

you are integrating from your own code
you need exact HTTP-level behavior
you are debugging auth or payload shape directly

Agent Analytics publishes auth.md discovery and user-claimed agent authentication for custom agent runtimes. Compatible agents can start from an unauthenticated API call, follow the WWW-Authenticate metadata pointer, ask the human for browser approval, then use a scoped aas_* bearer token.

Quick rule of thumb

Choose skill + CLI first in shell-capable or workspace-driven environments like Claude Code, OpenWork, Cursor, or Codex.
Choose MCP when the agent already lives in a connector-style chat environment and you want native tool calls.
Choose API when you need full control, custom integration, or lower-level debugging.

Paths Across Access Paths

Session paths are available through the same product surface:

CLI: agent-analytics paths <project> --goal <event>
MCP: analytics_paths
API: POST /paths

Use paths when the agent needs to connect entry pages and exit pages to goal behavior before deciding whether to run a funnel query, retention check, or experiment.

The report is intentionally bounded and session-local. It is not a long-cycle identity-stitching report.