Configuration¶

Rampart has two kinds of configuration:

Runtime/client config — where Rampart should find its local service
Policy config — what agents are allowed to do once Rampart evaluates a tool call

Most users only need one runtime setting:

# ~/.rampart/config.yaml
url: http://127.0.0.1:9090

Runtime Config (`~/.rampart/config.yaml`)¶

If you do not want to keep exporting environment variables, Rampart can load persistent local defaults from ~/.rampart/config.yaml.

url: http://127.0.0.1:9090
# serve_url: http://127.0.0.1:9090   # compatibility alias for url
# api: http://127.0.0.1:9091         # optional advanced override for split control-service setups

Setting	Use it for	Notes
`url`	Primary Rampart base URL	Canonical setting for hook, watch, plugin, and service-backed flows
`serve_url`	Backwards-compatible alias for `url`	Kept for compatibility; prefer `url` in new configs
`api`	Optional API base URL override for approval/control commands	Advanced only; usually unnecessary unless you split the API away from the main serve endpoint

Resolution order¶

Rampart resolves service addresses in this order:

flag → environment → config file → auto-discovered state → default

That means:

--api or --serve-url wins when you pass it explicitly
environment variables such as RAMPART_URL, RAMPART_SERVE_URL, and RAMPART_API override file values
~/.rampart/config.yaml is the persistent local default
if nothing is configured, Rampart falls back to discovered local state and then localhost defaults

For rampart protect, the resolved service URL is used consistently for startup checks, host configuration, and behavioral verification. An explicit non-default endpoint must already be reachable; Rampart will not silently start the default service instead. OpenClaw's native plugin further requires a loopback HTTP(S) URL and rejects credentials, paths, queries, and fragments.

Which setting should I use?¶

Use url unless you have a specific reason not to.

url is the normal setting for local Rampart service discovery
serve_url exists for compatibility with older setups
api is not the normal rampart serve setting — it is an advanced client-side override for approval/control flows

Two different meanings of --api

Client-side --api flags expect an API base URL such as http://127.0.0.1:9091.

Daemon/server --api flags refer to an API listen address such as 127.0.0.1:9091.

Policy File Location¶

Rampart policies are YAML files that define what your AI agent can and can't do. Policies are evaluated in microseconds and hot-reload when you edit them.

By default, Rampart looks for policies in ~/.rampart/policies/. You can specify a custom location:

rampart serve --config /path/to/policy.yaml
rampart wrap --config /path/to/policy.yaml -- agent

Basic Structure¶

version: "1"
default_action: allow  # allow | deny | ask | watch

policies:
  - name: my-policy
    match:
      tool: ["exec"]  # Which tool types this applies to
    rules:
      - action: deny  # deny | allow | ask | watch | webhook
        when:
          command_matches: ["rm -rf *"]
        message: "Destructive command blocked"

Actions¶

Action	Effect
`deny`	Block the tool call. Agent receives an error.
`allow`	Permit the tool call.
`watch`	Permit but log with elevated visibility. (Previously called `log` — `log` still works as an alias.)
`ask`	Block until a human approves or denies. Also valid as `default_action: ask` to surface unmatched calls for review.
`webhook`	Delegate the decision to an external HTTP endpoint.

Deny always wins. If any matching policy says deny, the call is denied regardless of other policies.

Tool Types¶

Common normalized tool types include:

Tool	What It Matches
`exec`	Shell commands (Bash, terminal)
`read`	File read operations
`write`	File write/edit operations
`fetch`	HTTP requests, network operations

For MCP tools, the tool name from the MCP tools/call is used directly (e.g., mcp__filesystem__read_file).

Matching Rules¶

Command Matching¶

rules:
  - action: deny
    when:
      command_matches:
        - "rm -rf *"        # Glob pattern
        - "mkfs.*"           # Filesystem format
        - "dd if=*"          # Raw disk write
        - ":(){ :|:& };:"   # Fork bomb

Path Matching¶

rules:
  - action: deny
    when:
      path_matches:
        - "**/.ssh/id_*"         # SSH private keys
        - "**/.aws/credentials"  # AWS credentials
        - "**/.env"              # Environment files
      path_not_matches:
        - "**/*.pub"             # Exclude public keys

URL Matching¶

Match the full request URL for fetch tool calls. Useful when you need more precision than domain matching — for example, to allow a domain but restrict specific paths:

rules:
  - action: deny
    when:
      url_matches:
        - "https://api.example.com/admin/*"   # Block admin endpoints
        - "http://*/*"                          # Block non-HTTPS requests
        - "https://*/v1/secrets*"              # Block any host's secrets path

Use domain_matches when you only care about the hostname; use url_matches when path, scheme, or query matter.

Session Matching¶

Filter rules by session identity — the repo path, branch, or project label Rampart associates with the current run. Useful for per-project policies without separate YAML files:

rules:
  - action: deny
    when:
      session_matches:
        - "*/production"      # Any repo on the production branch
        - "infra-*"           # Sessions labelled infra-*
  - action: allow
    when:
      session_not_matches:
        - "*/main"            # Exclude main branch from this rule

Sessions are set by the agent integration (Claude Code sets them from the project path; MCP proxy uses the client-provided session ID).

Tool Parameter Matching¶

Match arbitrary MCP tool input parameters by name and glob pattern. Each key is a parameter name; any matching parameter triggers the rule (OR logic across params):

rules:
  - action: deny
    when:
      tool_param_matches:
        path: "**/.env*"           # deny if the "path" param targets .env files
        url: "*webhook.site*"      # deny if the "url" param points to webhook.site

This is particularly useful for MCP tools where the tool name alone doesn't give enough context — for example, a generic file_read MCP tool where you want to restrict specific path arguments.

Domain Matching¶

rules:
  - action: deny
    when:
      domain_matches:
        - "*.ngrok-free.app"
        - "*.requestbin.com"
        - "webhook.site"

Response Matching¶

Scan tool output for sensitive patterns:

rules:
  - action: deny
    when:
      response_matches:
        - "AKIA[0-9A-Z]{16}"                          # AWS access key
        - "-----BEGIN (RSA )?PRIVATE KEY-----"         # Private keys
        - "ghp_[a-zA-Z0-9]{36}"                       # GitHub PAT

Rate Limiting¶

Use call_count to trigger a rule when a tool is invoked more than N times in a sliding time window. Useful for capping runaway agents or limiting expensive operations:

rules:
  - action: ask
    when:
      call_count:
        gte: 50       # Threshold: 50 or more calls...
        window: 10m   # ...within the last 10 minutes
    message: "High fetch volume — pausing for review"

  - action: deny
    when:
      call_count:
        tool: fetch   # Optional: count only fetch calls (omit for all tools)
        gte: 200
        window: 1h
    message: "Fetch rate limit exceeded"

window accepts standard duration strings: 1h, 30m, 10m, 5m, 1m. The counter increments on every PreToolUse event and resets as calls age out of the window.

Priority¶

Lower priority number = evaluated first. Use priority to ensure critical rules run before general ones:

policies:
  - name: critical-blocks
    priority: 1           # Evaluated first
    match:
      tool: ["exec"]
    rules:
      - action: deny
        when:
          command_matches: ["rm -rf /"]

  - name: general-logging
    priority: 10          # Evaluated later
    match:
      tool: ["exec"]
    rules:
      - action: watch
        when:
          command_matches: ["curl *"]

Approval Flow¶

For commands that aren't dangerous enough to block outright, but risky enough for a human to decide:

policies:
  - name: approve-deployments
    match:
      tool: ["exec"]
    rules:
      - action: ask
        when:
          command_matches:
            - "kubectl apply *"
            - "terraform apply *"
            - "docker push *"
            - "helm upgrade *"
        message: "Deployment — approve or deny?"

  - name: approve-installs
    match:
      tool: ["exec"]
    rules:
      - action: ask
        when:
          command_matches:
            - "pip install *"
            - "npm install *"
            - "brew install *"
        message: "Package install — approve or deny?"

How you'll see the approval prompt depends on your setup:

Claude Code — native permission dialog (the same one Claude uses for ask)
Standalone MCP clients — ask fails closed immediately because the stdio proxy has no safely reachable approval resolver; use an explicit allow/deny rule for now
OpenClaw — sends a chat message you can approve inline
Webhooks — sends a notification with a signed approve/deny link

Manage pending approvals:

rampart pending                          # What's waiting
rampart approve abc123                   # Let it through
rampart deny abc123 --reason "not now"   # Block it

Webhook Notifications¶

Get alerts when commands are blocked or need approval:

version: "1"
default_action: allow

notify:
  url: "https://discord.com/api/webhooks/your/webhook"
  on: ["deny", "ask"]  # Options: deny, log, ask

policies:
  # ... your policies

Example: Standard Policy¶

A sensible default that blocks destructive commands and credential access:

version: "1"
default_action: allow

policies:
  - name: block-destructive
    match:
      tool: ["exec"]
    rules:
      - action: deny
        when:
          command_matches:
            - "rm -rf /"
            - "rm -rf ~"
            - "mkfs*"
            - "dd if=*"
            - "chmod -R 777 /"
        message: "Destructive command blocked"

  - name: block-credential-access
    match:
      tool: ["read"]
    rules:
      - action: deny
        when:
          path_matches:
            - "**/.ssh/id_*"
            - "**/.aws/credentials"
            - "**/.env"
          path_not_matches:
            - "**/*.pub"
        message: "Credential file access blocked"

  - name: block-exfil
    match:
      tool: ["fetch"]
    rules:
      - action: deny
        when:
          domain_matches:
            - "*.ngrok-free.app"
            - "*.requestbin.com"
            - "webhook.site"
        message: "Exfiltration domain blocked"

Example: Anti-Exfiltration Policy¶

Detect encoding, obfuscation, and data staging:

version: "1"
default_action: allow

policies:
  - name: encoding-sensitive-files
    priority: 1
    match:
      tool: "exec"
    rules:
      - action: deny
        when:
          command_matches:
            - "base64 ~/.ssh/*"
            - "base64 ~/.aws/*"
            - "*cat*.ssh*|*base64*"
            - "*cat*.aws*|*base64*"
            - "*xxd*.ssh*"
        message: "Encoding of sensitive files blocked"

  - name: encoded-data-exfil
    priority: 1
    match:
      tool: "exec"
    rules:
      - action: deny
        when:
          command_matches:
            - "*base64*|*curl*"
            - "*base64*|*wget*"
            - "*base64*|*nc *"
        message: "Encoded data sent to network blocked"

  - name: reverse-shell
    priority: 1
    match:
      tool: "exec"
    rules:
      - action: deny
        when:
          command_matches:
            - "*bash -i >& /dev/tcp*"
            - "*python*socket*connect*"
            - "*nc -e /bin/*"
        message: "Reverse shell pattern blocked"

Hot Reload¶

Rampart watches policy files for changes via fsnotify. Edit a YAML file and the policy engine picks it up immediately — no restart needed.

Validation¶

Check your policy file for errors:

rampart policy check

Trace how a specific command would be evaluated:

rampart policy explain "rm -rf /"

Next Steps¶

Policy Engine → — Full evaluation logic
Policy Schema → — Complete YAML reference
Integration Guides → — Hook up your agent