Agent skill
Teach your agent MCPBytes
A skill is a folder of instructions an agent loads when a task calls for it. This one covers the whole MCPBytes workflow: uploads, waiting for jobs, reading results in chunks, and every tool's options and errors.
Install
Most agents read skills from one shared folder, ~/.agents/skills/: Codex, Gemini CLI, GitHub Copilot, Cursor, OpenCode, Goose, Zed, Amp, Windsurf, pi, oh-my-pi and DeepSeek Harness. Install the skill there once and they all find it. Claude Code and Cline only read their own folders (below). Agents load the skill by themselves when you ask for a 3D model to be split or a PDF to be read.
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesOr let the skills installer place it for the agents you choose. It finds the skill through this site's /.well-known/agent-skills/ index and checks the download against its SHA-256 digest.
# asks which agents to install for; -g installs for your user instead of the project
npx skills add https://mcpbytes.com -gYour agent's own folder
For a project instead of your user, use the same folder inside the repository (for example .agents/skills/mcpbytes/) and commit it: that is also how cloud agents get it. The archive is mcpbytes.tar.gz (or .zip).
Claude Code~/.claude/skills/
Claude Code reads only its own folder (or .claude/skills/ in a project). Check with /skills. Its docs
mkdir -p ~/.claude/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.claude/skills/mcpbytesCodex~/.agents/skills/
Codex reads ~/.agents/skills (the older ~/.codex/skills still works but is deprecated). Check with /skills, or type $ to pick a skill. Cloud tasks only see skills committed to the repository, in .agents/skills/. Its docs
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesGemini CLI~/.agents/skills/
Also reads ~/.gemini/skills. Check with /skills list. Gemini CLI asks for your consent the first time it activates a skill. Its docs
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesGitHub Copilot (VS Code and CLI)~/.copilot/skills/
Both also read ~/.agents/skills, and .github/skills/ in a project. Check with /skills in VS Code chat, or /skills list in the CLI. Its docs
mkdir -p ~/.copilot/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.copilot/skills/mcpbytesCursor~/.cursor/skills/
Also reads ~/.agents/skills. Check in the sidebar under Customize > Skills. Cloud agents do not get the skills in your home folder: commit the skill to the repository (.agents/skills/) for those. Its docs
mkdir -p ~/.cursor/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.cursor/skills/mcpbytesOpenCode~/.config/opencode/skills/
The same folder on Windows (under your user folder, not AppData). Also reads ~/.agents/skills and ~/.claude/skills. Its docs
mkdir -p ~/.config/opencode/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.config/opencode/skills/mcpbytespi~/.pi/agent/skills/
Also reads ~/.agents/skills. pi has no MCP support, so the skill uses its script and MCPBYTES_API_KEY. Invoke it with /skill:mcpbytes, or just ask. Its docs
mkdir -p ~/.pi/agent/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.pi/agent/skills/mcpbytesoh-my-pi (omp)~/.omp/agent/skills/
Also reads ~/.agents/skills (not pi's folder). Invoke it with /skill:mcpbytes, or just ask. Its docs
mkdir -p ~/.omp/agent/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.omp/agent/skills/mcpbytesCline~/.cline/skills/
Cline does not read ~/.agents/skills. Turn the skill on in the Skills tab (the scale icon at the bottom of the Cline panel). Its docs
mkdir -p ~/.cline/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.cline/skills/mcpbytesGoose~/.agents/skills/
~/.agents/skills is the folder Goose recommends. Check with goose skills list. Its docs
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesZed~/.agents/skills/
Zed reads only ~/.agents/skills (and .agents/skills/ in a project), without a restart. Check with the agent: manage skills command. Its docs
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesAmp~/.agents/skills/
Check with amp skills list. Its docs
mkdir -p ~/.agents/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.agents/skills/mcpbytesWindsurf~/.codeium/windsurf/skills/
Also reads ~/.agents/skills. Cascade picks the skill by its description; @mcpbytes invokes it. Its docs
mkdir -p ~/.codeium/windsurf/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.codeium/windsurf/skills/mcpbytesDeepSeek Harness (dsh)~/.dsh/skills/
Also reads ~/.agents/skills. With DeepSeek models inside another agent (Claude Code, OpenCode, ...), install the skill for that agent. Its docs
mkdir -p ~/.dsh/skills/mcpbytes
curl -fsSL https://mcpbytes.com/skills/mcpbytes/mcpbytes.tar.gz \
| tar -xz -C ~/.dsh/skills/mcpbytesClaude (desktop and web)upload
No local folder: download the zip, then Customize > Skills > + > Upload a skill. Skills there do not sync with Claude Code. Its docs
mcpbytes.zipWithout MCP, the skill uses your API key from MCPBYTES_API_KEY (create one in the console). With the MCP server connected (setup for every agent) it needs no key of its own.
What is in it
- SKILL.mdThe workflow: which interface to use, uploads, waiting, reading results, errors and limits.
- scripts/mcpbytes.pyREST client for agents without MCP: list tools, run a job, read and download results.
- references/split_3d_model.md3D model splitting: options, output files, how to use the manifest, errors.
- references/extract_pdf.mdPDF text extraction: options, reading text.txt in chunks, jumping to a page, errors.
The skill tells the agent to ask the API which tools exist (GET /v1/tools), so it keeps working as tools are added. It never asks for your key in the chat, and it sends only the file you asked about.
Agents without MCP
Some agents, pi for example, have no MCP support by design. The skill's script is a small REST client (Python 3.9+, standard library only) that gives them the same tools from the command line:
python scripts/mcpbytes.py tools
python scripts/mcpbytes.py run extract_pdf paper.pdf --out out/
python scripts/mcpbytes.py run split_3d_model chair.glb -o detail=high --out parts/
python scripts/mcpbytes.py read j_... text.txt --offset 0 --limit 20000