Install

Add the guard to your agent

Before your agent pays, AsAsked checks the cart against what the person asked for. In apps like Claude Code or Cursor it runs as an MCP server; in your own code it sits in front of checkout. Pick your app below.

1. A model key

The guard reads the person's request with a language model, then compares it with the cart in plain code. It needs one key, as an environment variable or in the app's config. Without a key it cannot read the request, and every check answers ask.

model	variable	extra
Claude	`CLAUDE_API_KEY (or ANTHROPIC_API_KEY)`	`asasked[anthropic]`
Gemini	`LLM_API_KEY`	`none`
Any OpenAI-compatible API	`OPENAI_COMPAT_BASE_URL, OPENAI_COMPAT_API_KEY, OPENAI_COMPAT_MODEL`	`none`

The configs below use Claude. For Gemini, drop ,anthropic and set LLM_API_KEY instead of CLAUDE_API_KEY. A key written into a config file is stored there in plain text.

2. Your app

The MCP configs start the server with uvx, which downloads and runs the package itself, so there is nothing to install first. It comes with uv. No uv? See “Other MCP apps”.

Run once in a terminal, with your key in place of your-key:

terminal

claude mcp add --env CLAUDE_API_KEY=your-key --transport stdio asasked -- uvx --from "asasked[mcp,anthropic]" asasked-mcp

Add --scope user to have it in every project. To share it with a project through .mcp.json, keep the key in your environment instead of the file:

.mcp.json

{
  "mcpServers": {
    "asasked": {
      "command": "uvx",
      "args": ["--from", "asasked[mcp,anthropic]", "asasked-mcp"],
      "env": {"CLAUDE_API_KEY": "${CLAUDE_API_KEY}"}
    }
  }
}

The plugin adds the same server plus a skill that tells Claude to call check_order before it pays. Type in Claude Code:

Claude Code

/plugin marketplace add ginjikurata/asasked
/plugin install asasked@asasked

The server reads the key from the environment Claude Code starts in: set CLAUDE_API_KEY in your shell before you run claude.

Settings → Developer → Edit Config opens claude_desktop_config.json. Add the server, with your key in place of your-key:

claude_desktop_config.json

{
  "mcpServers": {
    "asasked": {
      "command": "uvx",
      "args": ["--from", "asasked[mcp,anthropic]", "asasked-mcp"],
      "env": {"CLAUDE_API_KEY": "your-key"}
    }
  }
}

Quit Claude Desktop completely and start it again. If the server does not appear, replace "uvx" with the full path that which uvx prints; on macOS the log is in ~/Library/Logs/Claude/mcp-server-asasked.log.

~/.cursor/mcp.json for every project, or .cursor/mcp.json in one. The key is read from your environment:

mcp.json

{
  "mcpServers": {
    "asasked": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "asasked[mcp,anthropic]", "asasked-mcp"],
      "env": {"CLAUDE_API_KEY": "${env:CLAUDE_API_KEY}"}
    }
  }
}

Restart Cursor after saving. If the server does not appear, the app may not see uvx: replace "uvx" with the full path that which uvx prints.

For Copilot agent mode: .vscode/mcp.json in a project. The key is read from your environment: a server that prompts for it (${input:...}) is not passed on to Copilot agent sessions.

.vscode/mcp.json

{
  "servers": {
    "asasked": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "asasked[mcp,anthropic]", "asasked-mcp"],
      "env": {"CLAUDE_API_KEY": "${env:CLAUDE_API_KEY}"}
    }
  }
}

Run once, with your key in place of your-key:

terminal

codex mcp add asasked --env CLAUDE_API_KEY=your-key -- uvx --from "asasked[mcp,anthropic]" asasked-mcp

Or add it to ~/.codex/config.toml, passing the key through from your environment:

~/.codex/config.toml

[mcp_servers.asasked]
command = "uvx"
args = ["--from", "asasked[mcp,anthropic]", "asasked-mcp"]
env_vars = ["CLAUDE_API_KEY"]

In ~/.gemini/settings.json, or .gemini/settings.json in a project. This one reads requests with Gemini, so it needs LLM_API_KEY in your environment:

settings.json

{
  "mcpServers": {
    "asasked": {
      "command": "uvx",
      "args": ["--from", "asasked[mcp]", "asasked-mcp"],
      "env": {"LLM_API_KEY": "$LLM_API_KEY"}
    }
  }
}

Most apps that take MCP servers ask for the same three things:

command: uvx
args:    --from "asasked[mcp,anthropic]" asasked-mcp
env:     CLAUDE_API_KEY=your-key

Without uv, install the server with pip and use the path the second command prints as the command, with no arguments:

terminal

pip install "asasked[mcp,anthropic]"
which asasked-mcp     # Windows: where asasked-mcp

If payment goes through your own code, put the guard there. The agent then cannot pay around it, and it cannot answer the question for the person.

terminal

pip install "asasked[anthropic]"

python

from asasked import Guard

guard = Guard("Black ASICS in size 42, I need to be able to return them.")

# checkout is your payment function; cart turns its arguments
# into a list of dicts (name, brand, color, size, price, returnable, ...)
pay = guard.protect(checkout, cart=lambda items: items)

# Give the agent pay instead of checkout. If the cart does not match the
# request, checkout is not called: the agent gets "NOT PAID" and a question
# to show the person. Their reply comes from your code, not the agent:
guard.answer("Yes, final sale is fine")

Agent frameworks get a ready hook over the same check:

framework	install	hook
Claude Agent SDK	`asasked[claude-agent]`	`hooks={"PreToolUse": [checkout_hook(guard, cart, tool="mcp__shop__checkout")]}` from `asasked.integrations.claude_agent`
OpenAI Agents SDK	`asasked[openai-agents]`	`@function_tool(tool_input_guardrails=[checkout_guardrail(guard, cart)])` from `asasked.integrations.openai_agents`
LangChain 1.x	`asasked[langchain]`	`create_agent(..., middleware=[CheckoutGuard(guard, cart)])` from `asasked.integrations.langchain`

The cart format and everything else Guard takes are in the docs.

3. Check that it works

Ask the agent:

I asked for black sneakers in size 42. The cart has white sneakers in size 42 for $50. Check the order with check_order.

It should come back with ask and a question about the colour. If it answers ask with “I couldn't automatically check this order”, the server runs but has no working key.

A second opinion, not a lock

An MCP server answers only when the agent calls it, and it has to take the agent's word for the request and the person's replies. It helps an agent that wants to get the order right; it cannot stop one that skips it. To make sure nothing is paid around the guard, put it in front of your checkout, as in “Your Python agent”.

Package: asasked on PyPI · v0.1.10 release · the same configs in docs/install.md