Skip to content
← All Posts
Build notes

Self-hosted SolScribe now uses local AI models by default

Self-hosted SolScribe 1.8.7.0 uses local AI models unless you tell it otherwise. A server that isn’t the hosted service now refuses to send transcripts to Anthropic or OpenAI, and it checks the address a “local” model server points at, not just its name. SolScribe Cloud accounts can also delete themselves now, along with everything they own.

The release was tagged on October 10. It also puts Brain projects and kept daily digests, which have run on Cloud since September 28, into a tagged release for self-hosted servers.

Local models are the default

Running your own server usually means you care where the audio and text go. Until this release, a self-hosted server with an Anthropic key in its settings would use it anyway.

Now a self-hosted install runs in self_hosted mode. In that mode SolScribe makes no cloud LLM calls. Anthropic and OpenAI are refused, and Settings stops checking their keys. LM Studio, Ollama and vLLM setups need no change.

Transcription is a separate switch. A server that turns on RunPod offload with CLOUD_ENABLED still sends audio there, because that’s what the setting is for.

If you want a cloud model for summaries on your own server, set SOLSCRIBE_DEPLOYMENT_MODE=cloud and it works as before. Sending transcripts out is now something you opt into.

Checking the address of a local model server

Local-only mode used to trust a provider’s name. Now it also checks the address, since a provider called “Ollama” is only local if it points somewhere local.

Loopback and private addresses pass, and so do names ending in .lan, .local or .internal, which are accepted without a lookup. Any other hostname is resolved, and every address it returns must be private. A name that resolves to a public address is refused, and so is one that doesn’t resolve at all.

That catches a hosted API typed into the Ollama field by mistake. It also refuses your own GPU box when you reach it through a public hostname, which is stricter than some people will want. I think a clear rule beats a clever one here, and cloud mode is there for setups that need the reach.

The check limits which addresses SolScribe will send to. It can’t prove where the hardware sits. A VPN can carry a private address to another building, and a .lan name is taken on trust, so where those addresses lead is still up to your network.

SolScribe also stops following redirects from AI providers, since some redirects can forward the transcript to another address. If your model server sits behind a redirect, point the setting at its final address.

Deleting a Cloud account

SolScribe Cloud accounts now have a Delete account control on the Account page, behind a typed confirmation.

Deleting an account removes transcripts, notes and highlights, summaries, Brain, digests, voiceprints, webhooks, API keys, staged audio, unused credits and the credit ledger. If a transcription is still processing, the delete is refused with a note to try again once it finishes. A few things land after a job completes, such as voiceprints and Brain enrichment, so an hourly sweep deletes anything written for an account in the day after it was deleted.

One thing stays: audit log rows, kept under the account’s ID. On Cloud a daily job now deletes them after six years. An audit trail that vanishes on request isn’t much of a trail.

Deleting a single transcript on Cloud now removes its notes and highlights too, and deleting a recording removes what Brain learned from it.

Brain reaches self-hosted servers

Brain creates projects from names that recur across recordings, keeps an hourly status note for each one and cites the entries behind every line. Daily digests are stored, can follow your own time zone and carry structured Brain and to-do data in the daily.digest webhook.

The MCP server gains brain_search, brain_projects, brain_project and get_digest, so an assistant can answer questions from your own recordings with the entries to back it up.

Upgrading

SQLite installs create what they need on boot. PostgreSQL deployments need the migrations in migrations/neon/ up to 45 applied first. The server now refuses to start without migration 45, which is better than starting with AI features quietly refused for every account.

If your self-hosted server uses Anthropic or OpenAI today, set SOLSCRIBE_DEPLOYMENT_MODE=cloud before upgrading, or summaries will stop until you do.

Pricing is unchanged: $49 Professional, $499 Compliance & Legal and custom Enterprise. More at solscribe.ai.