Enterprise AI Agents: When to Use MCP, CLI, or Skills and How to Combine Them

When an AI Agent moves into a production environment, the focus shifts from simply having tools to extending the agent’s capabilities by connecting to various enterprise IT systems. MCP, CLI, and Skills each address a different problem, and only a thoughtful combination can unlock their full power.

Skills provide reusable organizational knowledge by packaging SOPs, templates, scripts, and validation rules into a knowledge package that solves the "how‑to‑complete‑a‑task" problem. MCP standardizes connections to external systems through a host/client/server model, exposing tools, resources, and prompts via stdio or HTTP, handling remote access, authentication, and governance. CLI stays close to the work site, leveraging mature command‑line tools such as git, gh, jq, kubectl, and python, which are lightweight and token‑efficient.

These three layers can overlap but their core positioning remains unchanged. For example, querying an ERP system can be done with an MCP‑wrapped tool, a direct CLI command, or a Skill that encapsulates the query method, parameters, and script.

The article proposes a three‑question decision framework:

Does the task require connecting to a remote system with identity semantics? Use MCP.

Does the task need to execute commands in the local environment (e.g., read files, run scripts)? Use CLI.

Does the task rely on internal knowledge, templates, or process standards that should be reusable? Use Skills.

Complex tasks rarely need a single solution; they benefit from a combination. The article illustrates this with a code‑release‑review agent that uses MCP to fetch change‑request data and approvals, CLI to analyze the local repository, and Skills to enforce company‑specific release rules and output formats.

When designing MCP tools, avoid exposing every REST endpoint as a separate tool. Instead, provide aggregated tools that align with specific task goals to keep context size manageable. For instance, rather than exposing getIssue, getApproval, and getDeploymentWindow as separate tools, offer a single "read release context" tool that aggregates the necessary data.

The article also debunks the claim that CLI "replaces" MCP. While CLI commands are shorter and can reduce token usage, token savings depend on controlled output, result summarization, and filtering. CLI excels at execution‑side tasks (local repo operations, build tests, script calls), whereas MCP excels at connection‑side tasks (CRM, OA, finance systems requiring authentication and governance).

Skills should not become massive knowledge bases. They should be concise SOP capsules containing trigger conditions, task goals, key steps, constraints, and output formats. Supporting materials such as references, templates, and scripts belong in separate files ( references, template, scripts).

The article introduces Programmatic Tool Calling (PTC), also called Code Mode, where the agent generates orchestration code that calls multiple MCP tools, CLI commands, or internal APIs in a sandboxed environment. This reduces round‑trip latency, minimizes context noise, and is suited for stable, well‑defined multi‑tool workflows with controllable risk.

PTC is appropriate when the workflow is stable, tool dependencies are clear, results are verifiable, and permissions are tightly controlled. It must run in a sandbox with explicit permission boundaries.

Emerging "Skills over MCP" patterns show MCP servers delivering associated Skills that describe how to combine tools. Future MCP servers may not only list available tools but also provide Skills that guide tool composition.

In summary, MCP solves standardized system connections and governance, CLI provides efficient local execution, Skills encapsulate reusable domain knowledge and SOPs, and PTC enhances multi‑tool orchestration. The key is to place each capability in its proper layer and combine them flexibly to build robust, production‑grade enterprise AI agents.