Back to dollama.net
What works today
Dollama is an experimental developer tool. This page lists what is available, what is available but experimental, what is in testing and not yet available, and what is only planned.
Dollama is an experimental developer tool. This page lists what is available, what is available but experimental, what is in testing and not yet available, and what is only planned.
Last verified 1 October 2026 · applies to CLI v0.72.0 and relay v0.64.0
| Feature | Status | Notes |
|---|---|---|
| Local endpoints for coding tools and applications | Available | Anthropic Messages, OpenAI Chat Completions and Responses |
| Private, Group and Open Network modes | Available | Your own machines are always tried first |
| Worker, planner and helper roles | Available | Assigned to each machine from its benchmark |
| Recovery: retries, continuation, tool-call repair | Available | Measured per model on the Models page |
| Web search and page fetch for the model | Available | Runs from your machine |
| Contribution-based priority | Available | Orders the queue when the network is busy |
| Speech-to-text | Available | Runs on CPU or GPU. See Capabilities |
| Embeddings | Available · experimental | Lightly benchmarked so far |
| Personal hosting of static sites | Available · experimental | See the limits on the Hosting page |
| Direct links between your own machines | Available · experimental | Opt in with dollama peer enable; falls back to the relay |
| Models outside Ollama (Bonsai 27B on llama.cpp) | Available · experimental | Your own machines only; slow to first token on a 12 GB GPU |
| Evaluating a candidate model and submitting the result | Available · experimental | Results feed the list of models the network accepts |
| Several capabilities on one machine | Available · experimental | For example speech-to-text next to a language model |
| Decision model for routing choices (Julia-1) | In testing · not yet available | Benchmarked for deciding when a request needs planning; not used in the product yet |
| Offline reference search (Kiwix) | In testing · not yet available | Local-only lookup of offline reference collections; not enabled yet |
| Client-to-client encryption | Planned | Would encrypt each request from your machine to the machine that runs it. Today each hop is encrypted in transit, but the relay can read requests. The serving machine always has to read it. |
| Sending less of each request to the relay | Planned | Built, but off by default |
| Self-service account deletion | Planned | By email until then |
| Separate settings for where your requests run and who may use your hardware | Planned | Today one mode sets both |
| Public source code and licence | Planned | No date yet. See Source release plans |
The aim is that models, relays and applications stay replaceable, so your access does not depend on any one provider, including Dollama.
Work that already exists in some form and is being finished.
Intended directions. None has a date.