Release highlights: September 24, 2026
Released September 24, 2026. This page also covers what shipped on September 10, 2026. Together, these releases let coding agents read your workspace through MCP, make workspace API keys available to every workspace, and add alerts for successful runs.
Breaking changes
- Upgrade the CLI to 0.28.5. Since September 10, 2026, CLI versions older than 0.28.4 are rejected with an upgrade message on every surface that checks the CLI version: the control-plane API and auth, and the data-plane API, logs and telemetry. Upgrade with
uv pip install -U "dlt[hub]", see Installation.
Workspace access for agents
A coding agent can now ask what is deployed in a workspace, what ran, what it printed, what it configured, and what the pipelines did, through dlt's MCP server. Installing dlthub-client makes the tools available, and they stay off until you ask for them:
dlthub ai mcp run --stdio --features=+dlthub.jobs,+dlthub.logs,+dlthub.config,+dlthub.telemetry
See dlthub ai mcp run for the full command, and the AI harness for how the pieces fit together.
Workspace API keys for every workspace
Workspace API keys are available to every workspace. Use one to run dlthub commands in CI or scripts, where there's no browser to log in with. The CLI signs in as the workspace instead of as you. See Use a workspace key.
Alerts on success
Alerts are no longer only about failure. You can now also be notified when a job run succeeds, configured from the Alerts tab in workspace settings. Each alert covers all pipelines or only the ones you pick, and goes to the members you choose. See dltHub Platform alerts.
Notification emails are also safer: content you supply can no longer inject HTML into them.
Workspaces and settings
An archived workspace is now clearly read-only. A banner appears on every one of its pages, and controls that would change data are disabled in workspace settings and on the jobs page.
Owners can still clean up environment variables in an archived workspace.
Permission cues across settings are consistent. Where an action is unavailable, a note names who can perform it, instead of the control being silently disabled, hidden, or replaced by an empty panel. The organization Members page explains the same way why inviting is unavailable until a region is set: owners get a note linking to General settings, where the region is chosen.
Smaller improvements
- Any key-value field can be copied, with a copy button appearing on hover or keyboard focus.
- Interactive apps such as Streamlit, marimo and MCP show a self-reloading "Starting your app" page while they boot, instead of a 502.
- Collaborators no longer reach a dead end in onboarding: the skip action and the organization setup wizard are hidden from them, rather than prompting them to create a workspace they have no permission to create.
- Usage totals follow the range you selected. Switching the chart from Daily to Weekly or Monthly used to widen the period behind "Total Credits by Workspace" and "Credits by Instance Size". Only the chart snaps its start to a bucket boundary now.
- Redeploying no longer drifts an
every:schedule. Shorteningevery:1htoevery:5mleft a job silent for the rest of the old hour, and redeploying without a schedule change pushed the next run one period further out each time. Cron (schedule:) jobs were never affected. - Unsaved variable rows stay in place when their scope changes.
- Table cells keep their state across live refreshes.
- Misrouted workspace pages redirect immediately instead of flashing the wrong page.
- The legacy per-workspace and organization settings pages, replaced by the centralized settings tree, now render the 404 page. Nothing in the product linked to them.