Changelog

Fix a space’s logs going missing when written before its agent connects to the server, or while it reconnects

Fix space CPU, memory and disk usage, and CPU history, showing the host’s figures rather than the space’s own

Fix knot secret get failing with an i/o timeout when 1Password takes more than a few seconds to answer, as it does for a mistyped reference Pro

Fix the usage history chart showing disk at 0% for spaces without a disk limit: it now shows the space used, on its own axis

Add file storage Experimental : buckets replicated across the cluster, with sharing, quotas, the Files page, knot file commands, the knot.files library and an API (docs)

Add S3 access to file storage at <server>/s3 Pro Experimental (docs)

Add the Backup Server permission and a backup API token scope; a new server creates a Backup User role and gives it, with Admin, to its first user

Improve usage visualization for the Usage and Users pages

Add the buckets each user can reach to the Users access overview Pro

Add Grok (xAI) support: set the chat provider to grok, or use provider=ai.GROK with scriptling.ai in scripts (docs)

Add PostgreSQL as a database; it stores sessions too, so they survive a restart without Redis / Valkey, which still takes the sessions when enabled (docs)

Add two-way and continual sync between a directory and a bucket with knot file sync --two-way and --watch (docs)

Add knot file move (alias mv), and move files on the Files page by dragging them onto a folder

Add the template custom field option Always Show When Creating a Space (docs)

Cache API tokens and groups to cut database reads

Suggest the closest commands, with their aliases, when the CLI is given an unknown command

Fix new installations failing to create their tables on MySQL 8 and later

Fix tunnel clients reconnecting after a tunnel was deleted on a busy server

Fix knot admin ignoring server.redis.hosts

Fix creating templates through the API without an idle timeout unit

Fix the Users page showing no users when one was created without roles or groups

Fix the chat ignoring provider and never sending reasoning_effort to the model

Fix a slow direct connection blocking a port forward from accepting further connections

Fix syslog lines without a priority, such as web access logs, being logged as errors

Fix space health checks ignoring the template’s timeout and skip SSL verification settings

Fix knot.template.update() turning off VNC on KVM templates

Fix knot.role.get() leaving out plugin permissions, so knot.role.update() cleared them

Fix knot.stack.list() reporting every space as not running

Fix the OAuth provider scopes setting being ignored Pro

Fix listing sessions through the API failing when authenticated with an API token

Fix the chat temperature and max_tokens settings never reaching the model; with temperature unset, the provider’s default now applies

Fix the AI assistant refusing on-demand tools: every approved execute_tool or tool_search call failed with “not callable from this context”

Fix a mistyped CLI subcommand running a different command, such as knot file buket list running knot file list

UI stacks: stack name auto completion on the space form offered no suggestions to users without permission to set space dependencies.

UI port forwarding: the target picker lists other users’ spaces that share a port, narrowed to those sharing the remote port entered Pro .

Apple Containers: fixed an issue where it wasn’t possible to delete the container.

The spaces page broke for users with only the share permission: it requested the user list without being allowed it, and the resulting error stopped the page loading anything. The share permission now grants the user list read the share dialog needs, and the page keeps working if that request fails.

UI improvements: the Clients page command block is readable in dark mode, and share/transfer dropdown avatars load only when opened instead of on every page view.

Outbound network restrictions for server-side scripts: a new server.script_net_policy setting points MCP tool and event sink scripts at a network policy file — the network equivalent of server.script_fs_allowed_paths. See Network Policy.

MCP over stdio (knot mcp): run knot as a local MCP server — the command proxies the remote server’s /mcp endpoint over stdio, authenticating from the stored connection (--alias) or --server/--token, never from the host’s config. Also built into the agent binary, so inside a space knot mcp connects through the agent socket with no configuration. See knot mcp.

MCP Apps in the AI chat: a tool call linked to a ui:// resource (the MCP Apps extension) now renders its app view inline in the conversation instead of staying buried in the tool-call disclosure. Remote MCP servers’ tools and resources are exposed to the chat too, and the MCP servers management page badges app tools and shows their icons.

Skills served over the MCP skills extension: knot’s skills are now exposed the standard way on /mcp: skills/list and skills/get, with each SKILL.md readable as a skill:// resource, scoped to the requesting user’s access. The web assistant and the OpenAI-compatible endpoints list available skills in the system prompt — including skills from attached remote MCP servers — and pull the full content on demand. See Skills.

Exclusive pool member leases: knot pool acquire checks out one healthy pool member for a caller’s exclusive use — shared routing skips it while held, and the holder works with it by its space name, the same way they would ssh to it. The default is the simple allocate → use → release with no timeout; optional lease limits and extension caps cover shared pools and CI, and release --destroy swaps the member for a fresh one so the next acquire starts clean. See Space Pools.

Scriptling: knot.pool leases: acquire(name, time=None, wait=None), extend(space, time=None), release(space, destroy=False), leases(name) and a leased() context manager; extend and release take the member’s space name or id, durations accept "5m"-style strings or "none". See knot.pool.

Idle auto-stop for spaces: a new template idle timeout stops a running space after it has had no user activity, an open terminal, SSH or web-port session (even idle), method calls, sustained CPU, and on Pro filesystem writes — for the configured time. Held pool leases are exempt. Set it next to Maximum Uptime in the template form, the API (idle_timeout, idle_timeout_unit), template export/import, and knot.template.create/update. See Managing Templates.

Shared ports and cross-user forwarding: template ports gain a shared type (alongside http, https and tcp), not published anywhere but reachable by every user in the same zone through port forwards, addressed as user--space. knot space port forward client 5432 paul--shared-pg 5432. Forward targets resolve as your own space, your own pool (round-robin across healthy members, leased members excluded), or another user’s space/pool on shared ports; access is re-checked on every connection. Templates, pool membership and users are now cached in memory on each node, invalidated on every write. See Space Forwarding.

Template port forward wiring Pro : templates can carry port forwards seeded into every new space. Client spaces connect to their services (or a pool) automatically on start. Edited in the Pro template form; available through the API, export/import and knot.template everywhere.

The /mcp endpoint serves knot’s own tools only: remote MCP servers are no longer federated through the public endpoint, so its tool list describes knot alone and doesn’t churn when a remote server is added, removed or changes. The web chat, the OpenAI-compatible endpoints and knot.mcp in scripts still list and call remote tools under their namespace prefix; external MCP clients that want a remote server’s tools should connect to that server directly. See Remote MCP Servers.

The get_skill MCP tool is removed: skills are no longer exposed as a tool on any surface. External MCP clients use the skills extension instead (skills/list, skills/get, resources/read on the skill:// URI). The CLI, the knot.skill scripting library and the skills API are unchanged.

knot.mcp works again: the scriptling library’s list_tools(), call_tool(), tool_search() and execute_tool() called API routes that were removed when chat moved to the OpenAI endpoints in February, so every call failed. The routes are back, resolving tools exactly like the web chat: knot’s own tools plus remote MCP servers under their namespace prefix. See knot.mcp.

CSI volume deletes survive an in-flight operation: Nomad rejects a concurrent delete with Aborted — an operation with the given Volume ID already exists; knot now retries for up to five minutes until the earlier operation clears, and if it stays wedged says so — with the remedy — instead of failing with the raw error.

The template form could never enable Maximum Uptime: the unit was submitted as disabled for every platform, so the setting silently did nothing regardless of what was chosen.

Persistent forwards created on a stopped space never came up: they stored the target as a space ID, which the proxy rejected, so the forward sat dead after the space started. Targets are now stored by name (user--space for other users’ targets) and the proxy accepts names, qualified names and IDs.

Non-admins got an empty space list from the CLI: knot space list and every other client that calls the spaces API without a user_id filter received nothing back, because the endpoint only allowed admins an unfiltered list. An omitted user_id now means the requester’s own spaces; asking for another user’s id still returns nothing without the manage-spaces permission.