Restricted knowledge bases, secrets in more nodes, a human-task inbox, and MCP v2

This release focuses on control and safety. You can now restrict a knowledge base to specific users and groups, and use secrets in more kinds of workflow step without their values ever showing up in outputs or error messages. People who are asked to review or approve something get a human-task inbox and a notification when work is assigned to them. Support for MCP v2 makes it easier to connect agents to outside tool servers. We've also made workflows easier to save and import, and fixed a number of bugs.

Highlights

Restricted knowledge bases

A knowledge base is a collection of your documents, such as policies, contracts or product manuals, that agents search when they answer questions. Until now, every knowledge base was available to everyone in its project.

You can now restrict a knowledge base to specific users and user groups. Only the people on its access list can use it. The list is managed by MagOneAI's access-control system, the same one that governs the rest of your permissions.

We've also tidied up how files are handled: when you delete or replace a file in a knowledge base, the old version no longer appears in the file tools agents use.

Why it matters: Sensitive material, like HR records or board papers, can live in MagOneAI without being exposed to everyone in the project. Agents also stop seeing files you've already removed.

Learn more: Knowledge bases · Groups

Secrets in Tool, Human-in-the-loop and Code nodes

In MagOneAI, a workflow (also called a use case) is a series of steps, called nodes, laid out on a visual canvas. A secret is a sensitive value, such as an API key or password, that MagOneAI stores securely so you never have to paste it into a workflow.

Secrets can now be used in three more kinds of node:

• Tool nodes, which call an outside system, for example a CRM or an internal API.
• Human-in-the-loop nodes, which pause the workflow until a person reviews, approves or adds information.
• Code nodes, which run a short piece of your own code.

Once a secret has been used, its real value is removed from the node's output, from error messages and from the screens shown to the person handling a human-in-the-loop step.

Why it matters: You can connect workflows to more systems without hard-coding credentials, and without the risk of a password turning up in a log, an error or someone's review screen.

Learn more: Secrets management · Tool node · Human task node · Code node

A human-task inbox, with notifications

Some workflows need a person to step in, for example to approve a payment or check an AI-drafted reply before it goes out. These steps create a human task for whoever is responsible.

Human tasks now have an inbox where the people involved can see the tasks waiting for them, and MagOneAI sends a notification as soon as a task is assigned.

Why it matters: Nobody has to keep checking whether a workflow is waiting on them. Tasks are picked up sooner, so workflows don't stall.

Learn more: Human task node · Notifications

MCP v2 support

Tools let agents act in other systems, such as sending email or updating a ticket. Many tools reach MagOneAI through tool servers that speak the Model Context Protocol (MCP), an open standard for connecting AI agents to outside tools.

MagOneAI now supports version 2 of MCP:

• Connect by URL. Tool servers are found from their URL, so they no longer have to be registered first.
• More external servers work. Outside tool servers that keep a session (a connection that remembers earlier requests) can now be reached.
• Custom headers. Servers that rely on the x-mcp-header header are supported.
• More flexible tool lists. MagOneAI reads a server's list of available tools whichever of the two common formats it uses.

Why it matters: Connecting agents to third-party and in-house tool servers takes fewer steps, and many more of them work out of the box.

Learn more: MCP protocol · External MCP servers

Improvements

• Save and run workflows without a default model. A workflow no longer needs a default AI model (LLM) set before you can save or run it. Instead, each LLM node must name the model it uses before it can be saved, so it's always clear which model a step runs on.
• Imports and store installs publish straight away. When you import a workflow or install one from the store, it now picks up your project's model automatically. It's ready to publish immediately instead of landing as a draft you have to fix.
• A warning when the default model has been deleted. If the default LLM saved on a workflow's Start node has been deleted, the Start node now tells you, so you can pick a new one.
• More reliable MCP connection tests. Test Connection for tool servers now uses MCP v2, and errors from tool servers keep their original error code, which makes problems easier to diagnose.
• Faster form filling. New input fields are ready to type in as soon as you add them.

Bug fixes

• Human tasks no longer time out unexpectedly. If someone answers after a task has closed, they now see a message explaining that the answer was refused, instead of it being dropped silently.
• Active runs are no longer cancelled by mistake. The background job that cleans up stuck executions (individual workflow runs) no longer cancels runs that are still working, including runs started by a webhook that are still taking in files.
• No more duplicate or empty saves when you press Cmd+S on the canvas.
• Files with Arabic, Chinese, Japanese or Korean names can now be downloaded.
• Tool server connections are more reliable. A tool server connection is never shared between organizations, streamed responses are no longer cut off, and a progress update is no longer mistaken for an empty successful result.
• Re-uploaded knowledge base files open correctly. If you deleted a file and then uploaded exactly the same file again, the new copy now opens as expected.
• Superadmin portal fixes: audit-log times now show in the right time zone, password rules accept the full intended set of characters, marketplace fields reset correctly, and replies to support queries work again.
• Querying spreadsheets works again. Agents can once more run data queries over CSV and Excel (XLSX) files.

Security

• Secrets stay hidden. As described above, secret values are removed from node outputs, error messages and human-in-the-loop screens.
• Tool server sessions are kept separate per organization, so one organization's connection can never be reused by another.
• Restricted knowledge bases let you keep sensitive documents limited to the people who need them.