Skills and MCP

Skills and MCP servers give your agents extra know-how and tools, such as documentation lookups, databases or your team's own checklists. This page shows you how to find them, choose what each agent gets, and add your own.

Open the library

Go to Settings → Skills & MCP. You can also open it from the sidebar, or from the ⚡ Skills & MCP button on the composer.

Discover new servers and skills

The Discover panel has two tabs, MCP servers and Skills. Type what you are looking for and press Enter or click Search.

MCP servers come from the GitHub MCP Registry, ranked by stars. Each one is labelled by how it runs:

Label	Meaning
Remote	Connects directly over the web. Nothing to install.
npm or PyPI	Runs inside each sandbox.
Not for sandboxes	Needs Docker or a desktop bundle, so it can't run in a sandbox. Threadrun shows the reason.

Click Add, or Add… if the server needs a key or a setting. Fill in the fields and click Add server. The entry then shows Added ✓ and Remove.

Skills come from skills.sh, the open Agent Skills directory, ranked by installs. Click Install to download a skill's folder from its GitHub repository. Installed skills show Reinstall.

If a remote server's address is blocked by your network policy, Threadrun shows a Network access needed banner. An admin can click Allow for that host. See Governance.

See what you already have

Threadrun also finds the servers and skills you set up in Claude Code, Codex and OpenCode on this computer. They appear in the same lists, with a badge that shows where they came from.

MCP servers come from your Claude Code and Codex settings, and from opencode.json or opencode.jsonc.
Skills are folders that contain a SKILL.md file. Threadrun looks in the skill folders of each tool.

Some servers depend on programs that exist only on this computer. These are hidden until you click Show local MCP servers. They are marked Unavailable in the cloud, with the reason, and you can't tick them.

Choose what each agent gets

Ticked servers and skills are the default selection. Untick an item to leave it out of that default.

You can also choose for each agent:

⚡ Skills & MCP on the composer, when you start a thread.
The same button on each factory stage.
The same button on each loop.

Tick what this agent should use, then click Done. Each agent gets only what you ticked there.

Add an MCP server by hand

Use the Add MCP server form under MCP servers:

Enter a name.
Choose HTTP, SSE or stdio.
Enter the server URL for HTTP or SSE, or the command for stdio.
Enter headers or environment values in the last field, then click Add server.

Headers look like Authorization: Bearer …. Environment values look like KEY=value, separated by commas.

Stdio servers run inside each sandbox, where Node.js and npx are available. HTTP and SSE servers must be reachable from the sandbox, so their address must be on the network allowlist.

Servers that need a key

When you open the page, Threadrun checks each remote server. If a server asks for authentication, its entry shows Needs API key.

Paste the header from the provider's docs, for example x-api-key: your-key.
Click Save key.

The key is stored in your OS keychain. Threadrun sends it only to threads that use this server, and only when those threads start.

If the server is set up in Claude Code, Codex or OpenCode, add the key there instead, then reopen the page.

If a server still refuses, the thread's Activity panel says why. A server that needs authentication shows as an error in the thread's MCP chips.

Skills, instructions, agents and commands

Add skill folder… adds a skill folder you already have on this computer. It must contain a SKILL.md file.
Global instructions, agents and commands copies your CLAUDE.md or AGENTS.md, custom agents and commands to the thread's agent. Each file is capped at 512 KB, and the total is capped at 8 MB.

Remove or switch off

Every item has a checkbox. Untick it to switch it off.

Items you added in Threadrun, by hand or from Discover, also have Remove. Removing an item deletes its stored keys. For skills you downloaded, it also deletes their files. A folder you added with Add skill folder… stays on disk.

Items from your Claude Code, Codex or OpenCode setup can only be switched off here. To remove them, change them in that tool.

Web search

Agents can search the web when the Web search checkbox is on. You find it in Settings → Governance, under Network egress. It is on by default.

OpenCode uses Exa (exa.ai) for its web search tool. Turning the checkbox off blocks that.
Claude Code and Codex search through their own model providers.

Searching and reading are separate. An agent can read a page only if the page's domain is on the allowlist. For example, add wttr.in to let agents read weather pages. See Governance.