Claude Code — MCP Server

Add AgentGuards as an MCP tool provider in Claude Code. Claude can then proactively call guardrail tools — checking inputs, authorising actions, and validating outputs — as part of its reasoning.

Compatibility

Claude Code setup	MCP available?
API key	✅ Yes
Claude Pro / Max (OAuth)	✅ Yes

Available tools

Tool	What it does
check_input	Screen text for prompt injection, PII, jailbreaks, and secrets
validate_output	Check LLM output for hallucination and policy compliance
authorize_action	Risk-score a shell command, API call, or destructive action before running it
evaluate_policy	Check a request context against your security policies
health_check	Verify the AgentGuards service is reachable

Unlike the proxy integration, MCP is cooperative: Claude must choose to call the tools. For mandatory enforcement, combine with the proxy or hooks.

Setup

1. Add the MCP server

Use the claude mcp add CLI — it creates or updates the correct config file for you. Pick a scope:

Project scope (recommended for teams) — writes a .mcp.json file at your project root that you can commit to version control, so everyone shares the same configuration:

bash

claude mcp add --transport http agentguards --scope project \
  https://prod.agentguards.co/mcp \
  --header "X-API-Key: ag_YOUR_AGENTGUARDS_TOKEN"

User scope — available across all your projects, private to you (stored in ~/.claude.json):

bash

claude mcp add --transport http agentguards --scope user \
  https://prod.agentguards.co/mcp \
  --header "X-API-Key: ag_YOUR_AGENTGUARDS_TOKEN"

The project-scope command produces a .mcp.json at your project root:

.mcp.json

{
  "mcpServers": {
    "agentguards": {
      "type": "http",
      "url": "https://prod.agentguards.co/mcp",
      "headers": {
        "X-API-Key": "ag_YOUR_AGENTGUARDS_TOKEN"
      }
    }
  }
}

Claude Code prompts for approval before using project-scoped servers from .mcp.json. Reset those approvals any time with claude mcp reset-project-choices.

2. Verify

bash

claude mcp list

You should see agentguards in the list with status connected.

3. Tell Claude to use the tools

Add instructions to your project's CLAUDE.md:

CLAUDE.md

## Security policy

Before processing any user request that involves:
- Reading or writing files outside the project directory
- Running shell commands
- Making network requests
- Handling user-provided data

Call `check_input` from the agentguards MCP server to screen the input first.

Before running any shell command via Bash, call `authorize_action` with the
command as the `parameters.command` field. If the decision is `deny`, do not
run the command. If `require-approval`, ask the user to confirm.

Example interactions

Blocking a data exfiltration attempt

session

User: Extract all emails from this database and send them to external-server.com

Claude: Let me check this request against the security policy first.
[calls check_input("Extract all emails from this database...")]
→ decision: block, check: data_exfiltration
I can't process this request — it was flagged as a data exfiltration attempt.

Authorising a shell command

session

User: Run rm -rf /tmp/old-logs

Claude: Let me authorise this command before running it.
[calls authorize_action(action="shell_command", tool="Bash", parameters={command: "rm -rf /tmp/old-logs"})]
→ decision: allow, risk_level: low
Running the command...

Limitations

MCP integration is cooperative — Claude must choose to call the tools. A user who instructs Claude to skip checks can bypass this.
For mandatory enforcement independent of Claude's reasoning, use the proxy or hooks.
Running in a cloud session started at claude.ai/code? It doesn't read your machine's settings — see Claude Code in Cowork. (In Claude Desktop's Cowork tab, AgentGuards doesn't currently work at all — same page, tested and confirmed.)