Some screenshots show mcp-s.com, which some existing organizations still use. New organizations use withwillow.ai. Copy the URLs shown in your own organization.
Security Settings
Security Settings is where you set organization-wide security defaults: which clients may reach your MCP, what happens to unapproved MCP servers and skills found on developer devices, how OAuth flows and web AI agents are treated, and how long user tokens stay valid.
Open Admin > Settings > Security. Security Settings has five sections:
| Section | Settings |
|---|---|
| Client Access | Approved Clients Only, Blocked AI Tool Message, Restrict Toolkit Access, IP Access Filter |
| Tokens & Approvals | Require Approval for API Tokens, Token Expiration, Approval Validity Period |
| Device & Browser Policies | Enforcement Policy, OAuth Policy, Web AI Agent Policy, Claude in Chrome |
| Guards & Risk | Guard Evaluation, Databricks AI Gateway, User Risk Score, Guards |
| Network & Secrets | Proxy Servers, Vault, MCP Registry |
Most blocks in this section are a control plus a link into the feature that owns it. The settings you change here are defaults. The linked pages manage individual servers, secrets, guards, and rules.
What you can do
- Require authenticated clients to reach your MCP, and turn off STDIO access
- Make group assignment binding on toolkits, so a slug alone does not grant access
- Restrict MCP access to an allowlist of IP addresses and CIDR ranges, for the whole organization or per gateway
- Word the message an AI tool sees when it calls an MCP server restricted to other AI tools
- Hold new API tokens until an admin approves them
- Set how long users stay signed in, and how long an approval stays valid
- Set enforcement defaults for unapproved MCP servers and skills on developer devices
- Set the default action for OAuth flows and web AI agents detected by the browser extension
- Choose where Claude in Chrome can work, what it can read, and which requests wait for approval
- Choose where agent hooks, the browser extension, and the guard API send content to be checked
- Score users on identity risk and block the riskiest ones
- Route an integration's outbound traffic through a forward proxy
- Reach the Vault, where encrypted secrets for MCP configurations live
- Inspect the MCP registry Willow serves to clients like GitHub Copilot
Client Access
Approved Clients Only
A toggle. When enabled, only authenticated clients can access the MCP.
Enabling Approved Clients Only disables STDIO access. Confirm your users are not relying on STDIO connections before you turn it on.
Select Manage AI Agents to work with the individual clients.
Blocked AI Tool Message
What users see when an AI tool calls an MCP server that is restricted to other AI tools. Background agents are never blocked by that restriction. Use the message to point people at a tool that works, for example {{mcpName}} is not allowed in {{aiTool}}. Please use {{allowedAiTools}}.
Select a placeholder to insert it; Willow fills it in when the call is refused. The placeholders are {{mcpName}}, {{aiTool}}, and {{allowedAiTools}}, described in the Allowed AI Tools reference. Leave the message empty for the default, <MCP server> is not available in this AI tool. A message set on an individual MCP server takes precedence over this one.
Restrict Toolkit Access
A toggle that decides whether group assignment on a toolkit is binding.
A toolkit is addressed by its slug (?toolkit=<slug>), and that slug resolves organization-wide. By default, assigning a toolkit to a group only decides who sees it listed in the portal, and anyone who knows the slug can connect through it. Enable this setting to make the assignment authoritative: a user can then only use toolkits assigned to one of their groups, toolkits they own, or toolkits shared with them. Any other toolkit is refused, even if the user knows its slug.
The restriction applies to the toolkit's tools, prompts, and resources alike.
Willow evaluates access to the MCP servers themselves separately, and this setting leaves that check unchanged. See Toolkits › How access is evaluated.
Select Manage Group Toolkits to assign toolkits to groups.
IP Access Filter
A toggle that restricts MCP access to a trusted allowlist of IP addresses and CIDR ranges. Requests from any other address are rejected.
Where Approved Clients Only asks which client is this, the IP Access Filter asks where is it calling from. Use it when your MCP traffic should only ever originate from known networks — an office range, a VPN egress address, or the static IP of a build system.
Turning the switch on reveals the allowlist. Each entry has two fields:
| Field | Description |
|---|---|
| IP address or CIDR range | A single address (203.0.113.7) or a range in CIDR notation (10.0.0.0/8). Both IPv4 and IPv6 are supported. |
| Label | Optional note describing the entry, such as HQ VPN. |
Select Add for each entry, then Save Changes.
The filter checks the caller's own address, the same one the Device column in Logs shows. Willow reads it past Cloudflare, load balancers, and private network hops, so your allowlist holds your users' public egress addresses. A hybrid gateway behind a public proxy of its own can list that proxy in TRUSTED_PROXY_CIDRS (comma-separated) so it is skipped too.
An enabled filter with an empty allowlist restricts nothing, and the card says so: The allowlist is empty, so no addresses are being restricted yet. Add at least one entry to start filtering. The filter only begins rejecting traffic once it has at least one entry.
Per-gateway rules
Per-gateway rules lists each of your gateways by host. Choose a rule for each one:
| Rule | Effect |
|---|---|
| Use organization filter | The gateway follows the organization filter above. This is the default. |
| No IP restriction | The gateway accepts requests from any address, even with the organization filter on. |
| Custom allowlist | The gateway uses its own allowlist. |
Willow matches a gateway by the host the client connects to. Requests that arrive on any other host, such as the shared MCP endpoint, follow the organization filter.
Confirm your own address is covered before saving. An allowlist that omits the network you are calling from will lock you out of your own MCP.
Tokens & Approvals
Require Approval for API Tokens
Turn on Require approval for new API tokens to hold each new API token in a pending state. A pending token can't authenticate until an admin with the api_tokens:approve permission reviews it. When the admin who creates the token already holds that permission, Willow skips the approval.
Select Manage API Tokens to review tokens.
Token Expiration
Sets how long users stay signed in to Willow before they have to sign in again, through the Expires after field.
Shorter expiry limits the window in which a leaked token is useful, at the cost of users reauthenticating more often.
The lifetime applies to sign-ins through the Claude Desktop extension and other STDIO clients too: the session they create lasts the longest group override the user has, otherwise this setting, otherwise 24 hours. If those users are asked to sign in every day, check this value and their groups' overrides.
When your SSO provider has passthrough refresh tokens enabled, the setting shows a note: set the refresh token lifetime in your identity provider to at least this value, or the refresh token expires first and passthrough MCP servers stop authenticating while the user is still signed in to Willow. The note links to Authentication Settings.
After changing it, select Save Changes. Your change does not take effect until you save.
Approval Validity Period
Sets how long an approval stays valid before the user has to approve again. It applies to tool calls held by Require approval on a tool and by guards whose action is Require Approval. The setting is Approval Validity Period under Security > Tokens & Approvals, and the Valid for menu offers 30 Minutes (the default), 2 Hours, 1 Day, 1 Week, 1 Month, or Custom from 1 minute up to 30 days.
An approval covers one user calling one tool with the same arguments. Within the validity period, repeating that exact call runs without a new prompt. A call with different arguments needs its own approval. A longer period means fewer prompts for people who repeat the same calls, at the cost of leaving each approval open for longer. See Manage Tools and Require Approval.
Device & Browser Policies
Enforcement Policy
Controls how unapproved MCP servers and skills are handled on developer devices. The two settings are independent:
| Setting | Applies to |
|---|---|
| MCP Servers | Unapproved MCP servers discovered on devices |
| Skills | Unapproved skills discovered on devices |
Select View Shadow AI Dashboard to see what has been discovered.
OAuth Policy
Sets the default action for OAuth authentication flows detected by the browser extension. The card shows the current action. Select Edit and choose one:
- Allow lets the OAuth flow continue.
- Warn shows a warning before continuing.
- Block stops the OAuth flow.
The default applies to all OAuth providers unless a provider overrides it. Select Manage Extension Allowlist in the dialog to set the overrides, then select Save.
Web AI Agent Policy
Sets the default action for web-based AI agents such as ChatGPT, Claude, and Gemini when the browser extension detects them. Select Edit and choose one:
- Allow lets users open the agent.
- Warn shows a warning before continuing.
- Block blocks the agent.
The default applies to all web AI agents unless an agent overrides it. Select Manage Web AI Agent Rules in the dialog to set per-agent rules, then select Save.
Claude in Chrome
Guards what Claude in Chrome reads and does while it works in a tab. The card shows On or Off. Select Manage to open the Claude in Chrome page.

Turn on Guard Claude in Chrome to apply the settings below. The Willow Guard browser extension enforces them whenever Claude in Chrome is working in a tab, and users can't change them.
| Setting | Effect |
|---|---|
| Sites | Where Claude can work. All sites except these blocks the sites you list. Only these sites allows the sites you list and nothing else. On any other page, Willow covers the page so Claude can't read it or act on it. |
| Blocked sites or Allowed sites | The list for the chosen mode. Enter a site and select Add. Use * as a wildcard, for example *://example.com/admin/*. A bare host like example.com covers every page on it. |
| Hide sensitive data | Checks each page against your runtime guards and hides what they catch before Claude reads it. Willow keeps pages that a blocking guard catches away from Claude. |
| Require approval | Holds Claude's requests until the person approves them in a Willow window. Applies on every site Claude can work on. |
Select Save to apply your changes, or Discard to drop them.
Guards & Risk
Guard Evaluation
Chooses where Willow sends content to be checked against your guards. The card shows the current choice, Willow portal or Run gateway.
- Willow portal (the default): content goes to your Willow portal, which runs your guards.
- Run gateway: content goes straight to your default run gateway and is checked there. With an external gateway it stays in your network. Choosing it shows the endpoint that devices and browsers will call.
The setting covers agent guard hooks (Claude Code, Cursor, Codex, Pi, Copilot), Claude inference hooks, the browser extension's Prompt Guard and Claude in Chrome scan, and the guard API. Devices and browsers pick up a change on their next sync. Hooks you deployed yourself through MDM or a plugin must be re-exported. An external gateway must run a version that serves guard evaluation, and Databricks keeps calling the Willow portal either way.
Databricks AI Gateway
Connects Databricks Unity AI Gateway to your runtime guards, so an External service policy in Databricks asks Willow for an allow/deny decision on every governed model and MCP tool call.
Select Enable to mint the OAuth M2M credential and reveal the host, API path, and token endpoint the Unity Catalog connection needs. Expand Setup afterwards to see those values again, add a credential, or revoke one.
The feature is in Private Preview at Databricks, so it also has to be enabled on your Databricks workspace. Setup and the fail-closed behavior are documented in Databricks AI Gateway.
User Risk Score
Scores every user from 0 to 100 on identity risk, so you can target guards at the risky ones or refuse their tool calls entirely. Badges on the card report whether the CrowdStrike sync is on and what threshold, if any, is blocking. Select Manage User Risk Score to open the page.
Scores can come from an admin editing a user, the Admin API, or a daily import from CrowdStrike Falcon Identity Protection. The page itself carries the score distribution, the CrowdStrike configuration, and the blocking threshold — see User Risk Score.
Guards
Shows whether guards are configured, along with a status badge such as Inactive. Select Manage Guards to create or edit them.
Guards are documented in full under Guards.
Network & Secrets
Proxy Servers
Defines HTTP(S) forward proxies for the gateway. Add a proxy once here, then choose it for an integration under that integration's Settings > Request Settings. The gateway routes the integration's outbound MCP and API traffic through the chosen proxy.

Select Add Proxy and fill in the form:
| Field | Description |
|---|---|
| Name | A label for the proxy, such as Corporate egress. |
| Proxy URL | The proxy address, such as http://proxy.internal:3128. |
| Username (optional) and Password (optional) | Credentials for a proxy that requires authentication. |
| Bypass hosts (optional) | Comma-separated hosts that skip the proxy, such as *.example.com, .internal, localhost, 10.20.*. * is a wildcard, so *.internal covers every subdomain. A bare host also covers its subdomains. * on its own bypasses the proxy entirely. |
| Skip certificate verification (insecure) | Turns off verification of the destination's certificate. Use it only when the proxy's TLS-inspection certificate is structurally invalid (for example, missing CA:TRUE) and a custom CA can't fix it. Traffic stays encrypted, but Willow no longer verifies the destination's identity. Fix the proxy's certificate as soon as you can. |
| Test connectivity to | A URL to reach through the proxy. Select Test proxy to check the connection. |
Default proxy for integrations applies to every integration that hasn't picked a proxy of its own. It defaults to No default (direct connection).
Select Save. With no proxy servers defined, integrations connect directly.
Vault
Store and manage encrypted secrets for MCP configurations. Select Manage on the Vault card to open it.
MCP Registry
Shows the registry Willow serves to MCP clients, and what they can reach through it. Select View Registry to open it.

The page opens with the registry endpoint — the base URL you give a client, such as https://your-org.withwillow.ai/registry. GitHub Copilot appends /v0.1/servers itself, so the base URL is all you enter. A badge reports whether the registry is Live and how many entries it serves. View raw JSON returns the registry document itself.

Beneath it, one card per registry entry:
| Entry | Required header | What it reaches |
|---|---|---|
MCP (your-org/mcp) | x-mcp | Your organization's MCP servers. |
Toolkit (your-org/toolkit) | x-toolkit | Your organization's toolkits. |
Each card lists the transport (streamable-http), the gateway URL clients connect to, and its version and status.
What each entry reaches
Each entry has a row reporting how many MCP servers, or how many toolkits, are reachable through it. Expand that row to see exactly which slugs a client can reach. The table lists the slug to send as the header, the display name, and the status. A search box filters it, and Include drafts and disabled widens the list beyond published entries.

Toolkits work the same way, and personal toolkits are marked with a Personal badge.

Being listed here means the slug is reachable through the registry. Whether a given user can actually use it also depends on group assignment and, where per-user authentication is required, on that user having connected their account.
This page is read-only. It reflects your MCP servers, toolkits, and gateway configuration. To enforce that Copilot uses only this registry, see Restrict GitHub Copilot to the Willow Gateway.