Agent and Add-on Compatibility Matrix

Every agent in this table is documented as supporting MCP except Aider, where the pages we read did not confirm it. Beyond that the field splits: Claude Code, Cursor, Antigravity CLI, Kiro and GitHub Copilot cover skills, hooks and custom agents, while the open-source tools lean on local models. A cell reads Yes only if a primary page says so. Checked 31 October 2026.

How to read this

The matrix is built from the documentation and repositories linked at the bottom, as read on 31 October 2026. It records what the vendor says the tool supports. It is not a test result. Where the pages we read were silent, the cell says Not confirmed. That means we did not find it, not that the feature is missing.

Six columns, each explained in the next section:

The matrix

Reading the notes:

• Gemini CLI is in transition. Google's developer blog describes moving to Antigravity CLI, which keeps skills, hooks, subagents and extensions. We read that as the Antigravity row, and left Gemini CLI itself open where its pages were silent.
• Cursor and GitHub Copilot cells come from each product's own documentation navigation and customisation pages, which list skills, subagents, hooks and MCP as separate features.
• Kiro says its agent supports AGENTS.md and a skills format. Treat that as the vendor's statement.
• Roo Code is a fork lineage of Cline, and the pages we read for it were thin. See the Roo Code page.

What each column means

MCP

The Model Context Protocol is an open standard for connecting an agent to outside tools and data. A server exposes tools, the agent calls them. Support means the agent can load servers in its configuration. It does not mean every server works well in every agent. See what are MCP servers and our MCP server hub. Transport is the practical difference: servers run locally over stdio or remotely over HTTP, and agents differ in which they accept.

Skills

A skill is a packaged set of instructions the agent loads when it needs them, typically a folder with a SKILL.md file. OpenClaw's docs describe exactly that shape. Skills are the lightest way to teach an agent a repeatable job. See agent skills and the skills hub. Skill folders are not guaranteed to be interchangeable between agents, even when they share a file name.

Subagents

A subagent is another agent session that a lead agent starts for a sub-task, often in parallel, and then reads the result. OpenCode ships a built-in general subagent. Antigravity CLI describes subagents as a way to delegate background tasks to concurrent sessions. See subagent and the subagents hub.

Hooks

Hooks run your scripts before or after the agent acts: format after each edit, scan for secrets, block a risky command. Claude Code, Cursor and Kiro document them. Cline exposes lifecycle hooks through its SDK, not as a file you drop in.

AGENTS.md

A shared instruction file read by many agents, so a team can keep one set of rules. See AGENTS.md and CLAUDE.md. The practical point: put shared rules in AGENTS.md and keep tool-specific notes in the tool's own file.

Local models

Whether the agent can talk to a model that runs on your hardware, usually through an OpenAI-compatible endpoint. Aider's README says it can connect to almost any model including local ones. Copilot's CLI docs name Ollama and vLLM as working through an OpenAI-compatible endpoint, provided the model supports tool calling and streaming. Local models are slower and weaker on hard tasks, so "Yes" is about access, not quality.

Patterns worth noticing

1. MCP is the common layer. It is the one feature nearly every tool lists, which is why an MCP server is the safest add-on to start with.
2. Skills and hooks cluster in the agents with a configurable harness. If you want to script an agent, check these columns first.
3. Open-source tools lead on local models. The tools that let you bring any provider are the ones that list local endpoints.
4. AGENTS.md is close to a default. Most of the tools that document project instructions read it.

What to pick, by need

Moving between agents

Your portable assets are the instruction file and the MCP servers. The less portable ones are skills, hooks and subagent definitions, whose formats differ.

1. Move shared rules into AGENTS.md.
2. Re-add each MCP server using the new agent's command and test one tool call.
3. Rebuild hooks. Do not assume they carry over.
4. Check where the new tool looks for skills.
5. Run one real task from your backlog and compare the diff.

Add-ons and the matrix

An add-on works with an agent only if the agent loads that add-on's format. MCP servers fit any agent with MCP support. Skills, subagents and hooks depend on the agent's own format. The directory's add-ons hub filters by agent compatibility, and the state of agent add-ons counts what we list by type.

Method and limits

• Built from the primary pages linked above, read on 31 October 2026.
• A cell says Yes only when a page states the feature or lists it as a documented section.
• We did not install or run these tools for this report.
• Features change often. Check the vendor's page before you commit.
• Our methodology describes how we check facts.

Not confirmed cells are open until we find a primary page that states the feature. If you spot an error, use the corrections policy.