# ollaya [Health: Active]

**Category:** 🗣️ Conversational AI  
**Repository:** https://github.com/ollaya-dev/ollaya  
**GitHub Stars:** 1472  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/ollaya

## Description
Jev-compatible API server for open decision models. Ollama for decision models: pull and serve 29 model families (Laya, pplx-decider, Clef, Winnow, coba...) locally or self-hosted behind /v1/systemone. The TypeSafe SDK works unchanged: set its base URL and pick an open model. CLI, desktop app, Docker, MCP server, public bench.

## Claude Desktop Quick Installation
No generated config for this server: it installs through its own setup steps.

The project's own setup steps (from https://github.com/ollaya-dev/ollaya), copied verbatim. Run them in order; they are not an mcpServers config.

1. Install:

```sh
curl -fsSL https://ollaya.dev/install.sh | sh
```

## Documentation

## What ollaya does

ollaya runs open decision models locally and makes their outputs available through MCP. These models inspect a state—such as a message, ticket, email, JSON object, or, for supported models, an image—and answer typed questions. Supported question types include choices, scores, and yes-or-no-style decisions. Results are calibrated probabilities returned in a single forward pass rather than generated prose.

The MCP interface is intended for agents that need a decision layer for routing, guardrails, triage, and tool approval. Running locally keeps the input on the machine or self-hosted server. The project includes multiple model families, including Laya, Winnow, Kev, Decider, NLI classifiers, GLiClass, and Qwen3Guard.

## How it works

Start the Ollaya daemon and expose it to an MCP client with:

```sh
ollaya mcp
```

The client launches the command as an MCP process, while Ollaya loads and serves the selected local model. Models are downloaded by name and can be managed with the Ollaya CLI. The daemon can also start automatically when a CLI command needs it.

Ollaya models use small ONNX graphs that reference the original weight files hosted by their authors. Model sources are pinned to commits and checked with SHA-256 verification. Some families, including Winnow, JevK5, and Jeb, run author-published GGUF files through llama.cpp instead.

The runtime supports CPU execution through ONNX Runtime, CUDA for NVIDIA GPUs, and several llama.cpp backends for GGUF models, including CPU, AMD ROCm/HIP, Windows Vulkan, NVIDIA CUDA, and Apple Metal. `laya` can select an English or multilingual model based on the request language.

## Setup and configuration

Install the Ollaya binary using the project installer, then download a model and start the MCP process. A typical local setup is:

```sh
curl -fsSL https://ollaya.dev/install.sh | sh
ollaya pull winnow:e4b
ollaya mcp
```

Configure Claude Code or another MCP client to run `ollaya mcp`. Cursor provides an installation option, and the project documents the equivalent MCP command for other clients.

The CLI also supports `run`, `pull`, `list`, `ps`, `show`, `rm`, `cp`, `stop`, and `create`. A Modelfile can package a question set into a named model. For example, a model can be built from a base model and a local `QUESTIONS` file, then started with `ollaya create` and `ollaya run`.

## Tools and capabilities

- Serve local open decision models through MCP.
- Answer typed choice, score, and yes-or-no-style questions.
- Process text states and supported image inputs with compatible models.
- Route requests between English and multilingual Laya variants.
- Run safety, intent, triage, and classification decisions locally.
- Calibrate a model on labelled JSONL examples and write calibration settings for a Modelfile.
- Expose a Jev-compatible HTTP API at `/v1/systemone`, `/v1/decisions`, and `/v1/models` for TypeSafe SDKs and existing Jev clients.
- Provide native HTTP endpoints such as `/api/decide`, `/api/pull`, `/api/tags`, `/api/show`, and `/api/ps`.

## Limitations and notes

Ollaya models do not generate text; they return decisions and probabilities. Model capabilities vary: for example, multilingual support, image handling, context length, and question behavior depend on the selected model. The recommended Winnow model is a 4B-class model that benefits from an NVIDIA GPU, while `laya:multilingual` is intended for CPU use.

Ollaya does not re-host model weights. It retrieves the required author-provided files from their Hugging Face repositories, while publishing the small model graphs needed to run them. Model performance and calibration can differ by model, device, quantization, and media type.

_Full upstream README: https://allmcps.com/mcp/ollaya/readme_

