API Anything

What you need

Requirement	Why
Node 22.13 or newer	Setup and the server run on Node
A Chromium browser: Chrome (default), Brave, Edge, Chromium, Vivaldi or Arc	Only to learn new sites and for the browser fallback. Calling operations someone already shared doesn't need one
An agent: Claude Code, Codex, Cursor, Claude Desktop, Gemini CLI or VS Code	Where the tools show up

To use a browser other than Chrome, set API_ANYTHING_BROWSER to a name like brave or to the browser's path.

Step 1. Check Node

node --version

If it says less than v22.13, or nothing, install the current LTS from nodejs.org first.

Step 2. Decide what you share

Setup asks whether you want to help build the shared catalog. Decide before you run it:

You want	Run setup with
To share what you learn from public pages (cookies and tokens are stripped)	--yes
To share nothing, but still search the catalog	--no-share
Nothing sent to the catalog at all, searches included	--no-share --local-only

For client work, use --no-share --local-only.

A second question, background verification (an idle server testing up to 20 new shared operations a day from your network), is off by default. Turn it on with --verify only if you want it.

Step 3. Run setup

Interactive. It lists the agents it finds and asks which ones to connect:

npx -y api-anything setup

Or name the agent directly:

npx -y api-anything setup --agents claude-code --no-share --local-only

Agent names: claude-code, codex, cursor, claude-desktop, gemini, vscode. Separate several with commas.

To see what it would change without changing anything, add --dry-run.

What setup does:

• Connects the MCP server to each agent you chose. In Claude Code that's claude mcp add --scope user api-anything -- npx -y api-anything mcp.
• Installs the skill (Claude Code: ~/.claude/skills/api-anything/. Codex, Cursor, Gemini CLI, VS Code: ~/.agents/skills/api-anything/). Claude Desktop takes skills only as an upload in its settings.
• Checks for a Chromium browser.
• Makes this install's key at ~/.api-anything/key.json.
• Backs up each config file it edits to <file>.bak-api-anything first. It never touches your other MCP entries.

If the output says "ok": false, look for the agent whose mcp is failed and read its error.

Step 4. Restart your agent

Close and reopen Claude Code (or whichever agent you connected) so it loads the new tools.

Step 5. Test it

npx -y api-anything call hacker-news search query=sqlite

You want "ok": true, "tier": 1 and a list of stories. Tier 1 means plain HTTP, no browser.

Then ask your agent: "What does Goodreads rate Piranesi?" It should call find_sites, then call_operation. For a site nobody has taught it yet, it calls learn_site.

The tools your agent gets

Tool	Does
find_sites	Searches the catalog and this machine's sites
list_operations	Shows an operation's inputs and returned fields
call_operation	Calls a read
call_batch	Up to 20 reads in a row
call_write_operation	Calls a write. Only on a server started with --allow-writes
get_job	The result of a call still running after about 25 s
get_result	The rest of an answer that was cut
site_status	This machine's record of a site's calls and repairs
learn_site	Learns a new read, tests it and answers your request
list_sites, login	What's installed; refresh a session you imported

Signed-in sites

To use a site where you're logged in, import your session from the browser you already use:

api-anything login <site>

Sessions are stored encrypted with your system's keychain. If a login window opens, sign in and close it. It closes by itself after 10 minutes.

Change your sharing answer later

npx -y api-anything setup --share-only --no-share

--share-only records your answers and changes nothing else. For one shell session only, API_ANYTHING_LOCAL_ONLY=1 does the same as --local-only.

Pin a version

Setup points each agent at the latest release, so every start runs the newest version. To stay on the version you tested:

npx -y api-anything setup --pin

Agents setup doesn't know

Print an entry with absolute paths and paste it into your client's MCP config:

npx -y api-anything mcp --print-config

Troubleshooting

Problem	Fix
The client says the server failed to start, or shows no api-anything tools	It can't find node or api-anything. Run api-anything mcp --print-config and use its entry. Check that the node it names is 22.13+
It worked, then stopped after you changed Node versions	The entry points at the old node. Run --print-config again and replace the entry
Startup timeout through npx	The first run downloads the package. Run the npx command once in a terminal, then restart the client
A call answers class: "pending"	It's still running. Call get_job with its job. Don't call the tool again
A tool call times out	Raise the client's tool timeout to 60 s or more. In Claude Code: MCP_TOOL_TIMEOUT (ms)
The client asks approval before every read	Restart it so it loads the current tool list
A call comes back blocked	The site showed a challenge it couldn't clear. There's no CAPTCHA solving

Remove it

API Anything's docs don't list an uninstall command. In Claude Code, removing the MCP entry stops the tools from loading:

claude mcp remove api-anything --scope user

Then delete ~/.claude/skills/api-anything/ and ~/.api-anything/. For other agents, remove the api-anything entry from their MCP config. Setup left a .bak-api-anything copy of each file it edited.

Before you rely on it

• It only knows what someone taught it, and the catalog is young.
• An operation that works for others can fail for you, from another region or network.
• Read each site's terms. You're responsible for how you use it.