Dyad LogoDyad
Releases

Dyad Free Pro model option, per-version diffs, and safer agent auto-approvals

Dyad v1.5.0 Beta 1 introduces a Dyad Free Pro model option, adds per-version commit diffs, improves agent approvals for SQL and MCP tools, and includes several fixes for app versioning, Git workflows, Supabase functions, and Linux support.

Highlights

Dyad Free Pro model option

Dyad now includes a Dyad Free Pro model option in the model picker.

This gives eligible Dyad Pro users a free model path with visible quota information directly in the picker. Dyad now shows remaining free model usage, displays a Data sharing chip for this model, and provides clearer messaging when the daily free model limit is reached.

The free model option works for all modes except for Build mode.

Compare changes with per-version diffs

Dyad now makes it easier to inspect what changed in an app version.

When you select a version in the version panel, the code editor now focuses on the files changed in that version and shows each file as a side-by-side diff. This lets you compare the previous version against the selected version without browsing the full app codebase.

Fewer prompts for safe SQL in Agent Mode

Agent Mode now reduces approval prompts for routine SQL.

SQL that neither changes the database schema nor drops data can run without a consent prompt. Schema-changing SQL and destructive SQL that can delete or wipe data still ask for approval. Dyad treats SQL it cannot safely classify as risky, so those queries continue to require confirmation.

This behavior is enabled by default and can be turned off from Advanced settings. Explicit per-tool approval choices still take precedence.

Improved MCP approvals

MCP tool approvals now use the same agent permission flow as other agent approvals.

Instead of showing MCP approval as a separate toast, Dyad now queues MCP approvals in the shared permission banner, shows one request at a time, and includes a count when multiple approvals are pending. This makes MCP approvals more consistent with the rest of Agent Mode.

Dyad also adds an opt-in experiment for auto-approving safe MCP tool calls. When enabled, Dyad can skip approval for MCP calls that look safe, such as read-only actions, while still asking for approval for risky or unclear actions. If the safety check fails, times out, or cannot classify the call, Dyad asks for approval.

When an MCP call is flagged for review or auto-approved, Dyad now surfaces the reason in the approval prompt or tool-call card.

Enhancements

  • Improved MCP tool discovery in search mode by listing available MCP tool names by server and adding a way for the agent to fetch a selected tool’s full schema only when needed.
  • Improved Supabase function deployment after shared module changes so Dyad deploys only affected functions when it can determine the impact, while falling back to a full deploy when analysis is ambiguous.
  • Fixed version checkout and revert when a Neon database snapshot is outside the retention window. Dyad now restores the code, leaves the live database in place, and shows a clear warning that the database snapshot could not be restored.
  • Fixed Git HTTPS operations on some RHEL-based Linux setups, including GitHub sync, push, pull, and template clone flows.
  • Fixed Git staging failures caused by global line-ending settings by applying app-local Git line-ending configuration before staging changes.
  • Refreshed the uncommitted changes banner as soon as a chat stream ends, reducing cases where the banner lagged behind recent agent changes.
  • Improved cancellation behavior for agent todos so canceled responses no longer leave new in-progress todo tasks visible.
  • Improved error handling for pending MCP approvals when a stream is canceled or ends, preventing tool calls from hanging after the approval banner is cleared.

Feedback

This is a beta release, so feedback is especially helpful—particularly around the Dyad Free Pro model option, per-version diffs, SQL auto-approval, MCP permission prompts, and the safe MCP auto-approval experiment.

If you run into issues or have suggestions, please help us improve by reporting a bug or uploading a chat session.

On this page