Release · v3.6.0
Per-run cost, wider audio and video support, and a readable audit trail
Every run now shows what it cost, transcription handles more formats and longer recordings, and the Superadmin audit log is easy to read and export.
This release makes MagOneAI more transparent. Every run now shows its real cost, step by step. Audio and video transcription accepts more file types and handles long recordings, and the Superadmin audit log is easier to read and can be exported. We've also strengthened security around outbound API calls.
Highlights
Token usage and cost for every run
AI models charge by the token, a small chunk of text roughly three-quarters of a word long. Every step of a workflow now records the tokens it actually used and what they cost. The execution log shows these figures for each step, plus totals for the whole run.
Why it matters: You can see exactly what each run cost, down to the individual step, which makes it easy to spot expensive steps and plan budgets.
Learn more: Usage & quotas · Analytics
Audio and video transcription, end to end
MagOneAI can turn the speech in audio and video files into text that your workflows can then work with. Transcription now:
- Accepts MKV, M4A, WebM and untagged MP3 files, as well as inputs that take a list of audio files.
- Handles long recordings by splitting them into overlapping pieces, so no words are lost where the pieces join.
- Explains failures clearly. If a file has no audio track, is over the 120-minute limit, or no transcription model is set up, the message says so and tells you how to fix it. Silence, or text the model probably made up, shows as a warning rather than an error.
- Is more reliable. Transcripts survive a server restart, and a retry carries on from where it stopped instead of re-processing (and re-billing) the whole file.
Why it matters: You can drop in far more recordings, such as meetings, calls and interviews, and get complete, trustworthy transcripts.
Learn more: Audio and video
A readable Superadmin audit trail
The audit log records administrative actions taken on the platform. Each entry now shows the values before and after the change, and the names of the things that changed instead of internal IDs. You can filter by the type of item changed and export up to 10,000 entries as a CSV file. Secret values are never recorded, only the name of the field that changed.
Why it matters: Reviewing who changed what, for example in a compliance review, is quick, and it never exposes sensitive values.
Learn more: Audit logging
Improvements
- Find the right model faster. In the model catalog, ticking Supports Vision or Supports Audio sorts models into those that match and everything else. Test Config now sends a real image or audio clip, so a model that can't see or hear fails when you save it, not halfway through a run. See Model capabilities.
- Deleting a model configuration removes it completely. It no longer leaves an "Inactive" card behind. The confirmation also lists any workflows that still rely on it.
- Disposable email addresses blocked. Sign-ups from throwaway email services are blocked by default, and administrators of self-hosted deployments can allow them if needed.
- Newest members first. Organization and project member lists show the people you've just added at the top, and refresh by themselves.
- One cookie choice across portals. Your cookie consent now carries across MagOneAI Studio, MagOneAI Hub and the Superadmin portal, and Hub and Superadmin offer an "Essential only" option.
- The right legal pages for each deployment. Each MagOneAI deployment now shows the Privacy Policy and Terms that apply to it.
Bug fixes
- Re-runs reproduce the original. Re-running an execution now uses the version of the workflow that originally ran, not whichever version is published now.
- Super Agent works with choice fields. A Super Agent no longer fails when a workflow it uses starts with a drop-down (choice) input.
- Secrets work in parallel branches. Steps inside a Parallel node can now use stored secrets, for example in an API call's headers.
- Transcription settings save correctly, and using Claude as the main model with OpenAI for transcription no longer fails.
- Lighter notifications. Open browser tabs no longer hold a database connection for as long as the tab stays open.
Security
- A hardened API node. The API node calls external web services. Outbound calls are now limited to an approved list of network ports. Custom headers and stored secrets are removed if a request is redirected to a different domain. Oversized compressed responses and deliberately slow servers are cut off, and response headers carrying credentials are hidden from logs. See API node.
- No hardcoded passwords. A database password used by our secrets vault is now stored as a secret instead of in the code.