diff --git a/content/docs/features/approvals.mdx b/content/docs/features/approvals.mdx index b81db51..95fc6c6 100644 --- a/content/docs/features/approvals.mdx +++ b/content/docs/features/approvals.mdx @@ -1,10 +1,10 @@ --- title: Approvals and permissions sidebarTitle: Approvals -description: Choose when Fluso asks before network requests and connector actions, review approval cards, and manage saved rules. +description: Choose when Fluso asks before connector and MCP tool actions, review approval cards, and manage saved tool rules. --- -Approvals let you decide how much Fluso can do without stopping for confirmation. You can keep every external action behind a prompt, allow ordinary network requests while reviewing connector tools, or let both continue automatically. +Approvals let you decide how much Fluso can do without stopping for confirmation. You can keep every connector and MCP tool action behind a prompt, or let Fluso run them automatically. Your choice belongs to a chat. Existing chats keep their current mode. A choice made before a new chat starts is pending until you send the first message and the chat is created. @@ -14,23 +14,12 @@ Your choice belongs to a chat. Existing chats keep their current mode. A choice ## Choose an approval mode -| Mode | Network access | Connector tools | -|---|---|---| -| **Ask for everything** | Ask before a new hostname is reached | Ask before a tool runs | -| **Ask for critical** | Approve new hostnames automatically | Ask before a tool runs | -| **Don't ask permission** | Approve new hostnames automatically | Approve tools automatically | +| Mode | Connector and MCP tools | +|---|---| +| **Ask for everything** | Ask before a tool runs | +| **Don't ask permission** | Approve tools automatically | -Use **Ask for everything** when you want to inspect each new external action. **Ask for critical** removes routine network prompts but still pauses before a connector can read or change data. **Don't ask permission** is best for work you trust Fluso to finish without waiting for you. - - +Use **Ask for everything** when you want to review each action before a tool can read or change data. **Don't ask permission** is best for work you trust Fluso to finish without waiting for you. ### Set the mode from the composer @@ -39,7 +28,7 @@ Use **Ask for everything** when you want to inspect each new external action. ** Start a new chat and click the shield control beside the model mode in the composer. - Pick **Ask for everything**, **Ask for critical**, or **Don't ask permission**. The menu explains which requests will still pause. + Pick **Ask for everything** or **Don't ask permission**. The menu explains which requests will still pause. Before a thread exists, the composer labels the choice as **Pending**. Sending your first message creates the chat and confirms its mode. @@ -48,24 +37,6 @@ Use **Ask for everything** when you want to inspect each new external action. ** You can change the mode between messages. The updated choice applies to later actions in that chat. It does not change other existing chats. -## Network approval cards - -With **Ask for everything**, Fluso pauses the first time work needs an HTTPS hostname that has no saved rule. The card shows: - -- The exact hostname Fluso wants to reach. -- The blocked HTTP method and path, so you can see what prompted the request. -- **Keep blocked** and **Allow** actions. - -Choose **Allow** to let the current work continue and save an allow rule for that hostname. Choose **Keep blocked** to stop the request and remember the deny rule. Fluso will not keep asking about a hostname you denied until you change or remove that rule in Settings. - - - A network rule covers the whole HTTPS hostname, not only the path shown in the card. Allowing `api.example.com/v1/report` also allows other HTTPS paths on `api.example.com`. A subdomain such as `files.example.com` is a separate hostname and asks separately. - - -Redirects are checked again at their destination. Allowing the first hostname does not silently allow a redirect to another one. - -If the card expires or the run is canceled before you decide, ask Fluso to try the action again. An organization-managed deny is locked and cannot be approved from chat. - ## Connector tool approval cards Connector and MCP tools can read or change data in another service. When a tool needs approval, the card shows the connector or server, tool name, a short argument summary, and an expandable view of the exact arguments. @@ -78,55 +49,25 @@ You can choose: If the required app is not connected, the card asks you to **Connect** first. Complete sign-in and Fluso resumes the waiting action. If the request changes while you are deciding, it fails closed instead of applying your decision to different arguments. +If the card expires or the run is canceled before you decide, ask Fluso to try the action again. An organization-managed deny is locked and cannot be approved from chat. + Use **Allow once** for a sensitive write or an unusual set of arguments. Use **Always allow** for a tool you expect this chat to call repeatedly. ## Manage approvals in Settings -Open your account menu, choose **Settings**, then open **Approvals & permissions**. This panel controls the mode for your next new chat and lists the network rules saved for your account. - - - - - Fluso Approvals and permissions dialog with three auto-approval modes and the network access rule editor. - +Open your account menu, choose **Settings**, then open **Approvals & permissions**. This panel controls the mode for your next new chat. ### Set the default for the next chat -The **Auto-approval** section has the same three modes as the composer. A Settings change is staged for your next new chat. It does not rewrite the mode of chats that already exist. - -### Add a network rule - - - - In **Network access**, enter a public HTTPS hostname such as `api.github.com`. You can paste an HTTPS URL; Fluso saves its hostname. - - - **Allow** lets Fluso reach the hostname. **Deny** blocks it and suppresses future approval prompts for that hostname. - - - Click **Add**. The saved row is available to every chat on your account. - - - -There is no saved **Ask** rule. Asking is the default when a hostname has no user or organization rule, so remove a user rule to return that hostname to ask-on-use behavior. - -Hostnames approved automatically are marked **Auto-approved**. You can remove auto-approved rows in bulk without removing rules you chose yourself. An origin managed by an MCP connection may return as an ask rule while that connection remains active. +The **Auto-approval** section has the same modes as the composer. A Settings change is staged for your next new chat. It does not rewrite the mode of chats that already exist. ## Which rule wins Fluso applies the most restrictive relevant rule. In practical terms: -1. An organization **Deny** always blocks the hostname or tool. +1. An organization **Deny** always blocks the tool. 2. Your **Deny** can block something the organization allows. 3. An organization **Allow** can make an action available, unless you denied it. 4. Your **Allow** can approve an action when the organization has not blocked it. @@ -151,20 +92,14 @@ See [MCP servers](/features/mcp) for connection setup and credential handling. The chat has not been created yet. Send the first message to commit that mode to the new chat. - - Compare the hostnames. Rules are exact per HTTPS hostname. Subdomains and redirect destinations are checked separately. - - - Check **Settings → Approvals & permissions** for a saved **Deny**. If the rule is locked or marked as managed by your organization, only an organization administrator can change it. - Ask Fluso to retry the action. Approval cards are bound to one pending request and cannot approve a later or changed request. Tool rules apply to the specific tool, not every tool on the connector or MCP server. Review that server's tool permissions under **Plugins**. - - Remove its user rule in **Settings → Approvals & permissions**. With no saved allow or deny, the chat mode controls what happens next. + + Check that server's tool permissions under **Plugins** for a saved **Deny**. If the rule is locked or marked as managed by your organization, only an organization administrator can change it. diff --git a/content/docs/features/chat.mdx b/content/docs/features/chat.mdx index d977665..8abdc65 100644 --- a/content/docs/features/chat.mdx +++ b/content/docs/features/chat.mdx @@ -30,9 +30,9 @@ This only affects the chat you are using in the app. Telegram has its own conver ## Approval mode -Use the shield control beside the model mode to choose when Fluso asks before network access or connector tools. The choice belongs to the current chat. Before a new chat exists, it remains pending until you send the first message. +Use the shield control beside the model mode to choose when Fluso asks before connector and MCP tool actions. The choice belongs to the current chat. Before a new chat exists, it remains pending until you send the first message. -See [Approvals and permissions](/features/approvals) for the three modes, in-chat approval cards, saved network rules, and organization policy behavior. +See [Approvals and permissions](/features/approvals) for the modes, in-chat approval cards, saved tool rules, and organization policy behavior. ## Asking for things well diff --git a/content/docs/release-notes.mdx b/content/docs/release-notes.mdx index 97611ec..6a13364 100644 --- a/content/docs/release-notes.mdx +++ b/content/docs/release-notes.mdx @@ -6,11 +6,11 @@ description: "What's new in Fluso - product updates, improvements, and fixes." **Highlights** -- Fluso asks before it reaches a new site or uses a tool you haven't approved. Your answer sticks, and stays editable in settings. +- Fluso asks before it uses a tool you haven't approved. Your answer sticks, and stays editable in settings. - Granular per-tool [permissions](https://docs.fluso.ai/features/approvals): let an agent read your inbox, but require your approval before it sends anything. - Fluso is now on the [iOS App Store](https://apps.apple.com/in/app/fluso/id6780937477). - Works end to end with any [MCP server](https://docs.fluso.ai/features/mcp): ours, hosted, or your own. -- Kimi K3 becomes the default model. +- Kimi K3 becomes the default model in standard chats, and Confidential Mode runs DeepSeek V4 Flash on the encrypted path. - Share a chat as a public read-only link. - Move in and out of Confidential Mode whenever you want. @@ -21,12 +21,10 @@ description: "What's new in Fluso - product updates, improvements, and fixes." - Every tool on every connection carries its own [allow, ask or deny](https://docs.fluso.ai/features/approvals). - An agent can read your Gmail freely, while sending an email always stops for your approval. - The same split works across Slack, Drive, your trackers and anything else you connect. -- Network access follows the same rule. Reaching a site nobody approved raises a card instead of happening quietly. - Your decisions are durable, and survive reloads and restarts. - Each one shows whether it came from your organisation or from you. - Read-only tools can be auto-approved in one switch. - Everything stays editable in settings whenever you change your mind. -- The first time it is live for your account, a short walkthrough shows how it works. **Any MCP, from anyone, anywhere, end to end** @@ -51,9 +49,10 @@ description: "What's new in Fluso - product updates, improvements, and fixes." **Kimi K3 is the default model** -- Chat now runs on Kimi K3. +- Chat now runs on Kimi K3 for standard, non-confidential work. - It holds its own against the best closed-source frontier models on the long, multi-step work people actually hand to an assistant. - It is fully open source, so the quality you get does not depend on a vendor's roadmap. +- Confidential Mode runs its own model on the encrypted path: every confidential request goes through DeepSeek V4 Flash. **Confidential Mode, on your terms**