QOMPASS AI

LOCAL MODEL SERVER — GO

Rose

Rose

Version 1.00Source on GitHub

OVERVIEW

Rose is Qompass AI's fork of Ollama, the local model server — upstream Ollama at tag v0.40.1 imported verbatim with the fork delta re-applied — built to serve phlow's local specialists (planner / coder / reviewer) on local hardware with its own identity, registry and security posture. It is not rose.nvim; the plugin talks to this server as a backend. The README tagline is "Start building with open models."

HOW IT WORKS

  1. 1Install/run the `rose` binary (macOS/Windows/Linux/Docker routes in the README); running `rose` prompts to run a model or connect Rose to existing agents/applications. Source: GitHub rose README.md, "Download" / "Get started".
  2. 2`rose serve` starts the local server, loopback by default at `127.0.0.1:11434`; the HTTP API remains unauthenticated, as upstream's is. Source: GitHub rose `docs/rose.md`; `envconfig/config.go` (`Host()` default).
  3. 3Models are pulled, run and managed from the CLI (`rose run`, `rose pull`, `rose ls`, `rose ps`, `rose stop`, `rose rm`), and customized by writing a `Modelfile` and running `rose create -f Modelfile`. Source: GitHub rose `docs/cli.mdx`.
  4. 4Applications call the REST API — e.g. `POST http://localhost:11434/api/chat` with a model and messages — or the Python/JavaScript client libraries shown in the README. Source: GitHub rose README.md, "REST API".
  5. 5Coding/agent integrations are launched and configured through `rose launch` (e.g. `rose launch claude`, `rose launch claude --model qwen3.5`, `rose launch droid --config`). Source: GitHub rose `docs/cli.mdx`.
  6. 6Fork identity is enforced in configuration: only `ROSE_*` environment variables are read (`OLLAMA_*` are not aliases), the default model store is `$XDG_DATA_HOME/rose/models` (or `~/.local/share/rose/models`), `ROSE_MODELS` always wins, and unqualified model names resolve against `harbor.qompass.ai`. Source: GitHub rose `docs/rose.md`; `envconfig/config.go`.
How it works, step by step — each step as the project's own documentation states it.

CONCEPTS

ONE REQUEST THROUGH THE ROSE SERVERHTTP APIPOST /api/chatrose serve · 127.0.0.1:11434SCHEDULERqueue — max 512 waitingkeep-alive 5m · load timeout 5mRUNNER LANESplanner · coder · reviewerthe specialists phlow callsloads weights fromMODEL STORE — GGUF FILES ON DISK$XDG_DATA_HOME/rose/modelsfallback: ~/.local/share/rose/modelsROSE_MODELS overrides the store path — it always winsnames resolve via harbor.qompass.aiTHE SAME SERVER FROM THE CLIrose pull · rose run · rose ls · rose ps · rose stop · rose rm · rose createONLY ROSE_* VARIABLES CONFIGURE ROSE
A request's path through the rose server: the HTTP API (loopback 127.0.0.1:11434 by default) hands work to the scheduler, which queues requests (ROSE_MAX_QUEUE 512), loads models from the GGUF store at $XDG_DATA_HOME/rose/models, and runs them in runner lanes — the planner, coder, and reviewer specialists phlow calls. The CLI drives the same server, and the configuration values shown are the server's own envconfig defaults.

CONFIGURATION

OptionDefaultWhat it does
ROSE_HOST127.0.0.1:11434IP address / bind for the rose server
ROSE_MODELS$XDG_DATA_HOME/rose/models`, else `~/.local/share/rose/modelsPath to the models directory; no fallback to a stock `~/.ollama/models` store
ROSE_KEEP_ALIVE5mDuration models stay loaded in memory
ROSE_LOAD_TIMEOUT5mHow long to allow model loads to stall before giving up
ROSE_MAX_QUEUE512Maximum number of queued requests
ROSE_NUM_PARALLEL1Number of parallel model requests
ROSE_MAX_LOADED_MODELS0 (as coded)Maximum number of loaded models per GPU
ROSE_MAX_TRANSFER_STREAMS4Maximum parallel transfer streams for safetensors model pulls/pushes
ROSE_KV_CACHE_TYPE`f16` (default per map description)Quantization type for the K/V cache
ROSE_CONTEXT_LENGTH0 (as coded)Context length override
ROSE_DEBUG(unset)Show additional debug information (e.g. `ROSE_DEBUG=1`)
ROSE_NO_CLOUD(unset / false)Disable Rose cloud features (remote inference and web search)
default registryharbor.qompass.aiRegistry that unqualified model names resolve against (fork deviation from `registry.ollama.ai`)

CONTROLS

ActionControl
Start Rose (interactive)rose
Run and chat with a model`rose run <model>` (e.g. `rose run gemma4`)
Launch integrations menurose launch
Launch a specific integration`rose launch claude` (also `--model <model>`, `--config` variants in docs)
Download a modelrose pull <model>
Remove a modelrose rm <model>
List modelsrose ls
List running modelsrose ps
Stop a running modelrose stop <model>
Create a customized modelrose create -f Modelfile
Start the server`rose serve` (`rose serve --help` lists settable environment variables)
Sign in / sign outrose signin` / `rose signout
Chat via REST APIcurl http://localhost:11434/api/chat -d '{"model":"gemma4","messages":[...],"stream":false}'

NOT COVERED

- The GitHub README is upstream-Ollama-shaped: install commands and several links still point at ollama.com / Ollama-named artifacts, and community/library lists are inherited upstream content. The fork-specific facts above come from `docs/rose.md` and `envconfig/config.go`, not the README. - The sandbox clone's README (`~/workspace/repos/rose/README.md`) is older organisational boilerplate (Responsible Open Science Engine framing, FAQ, licensing essay) with no install, CLI or configuration content; it disagrees with the current GitHub README and was not used for CONFIG/CONTROLS. - Not in this branch, per `docs/rose.md`: the hybrid-TLS transport layer lives on a separate branch and is not ported forward yet; the HTTP API is unauthenticated. - Full `ROSE_*` variable list was not exhaustively transcribed; `rose serve --help` is the documented way to view settable variables, and only variables verified in `envconfig/config.go` are listed above.

REFERENCES

  1. GGUF format specification (ggml)The model file format in rose's store: the metadata and tensor layout every GGUF file the server loads follows.
  2. llama.cpp (ggml-org)The ggml-org inference project in rose's lineage: rose's own import guide names its llama-quantize tool as the way to prepare and quantize GGUF models for import.
  3. XDG Base Directory SpecificationDefines $XDG_DATA_HOME — the variable rose's default model store, $XDG_DATA_HOME/rose/models, is built from.