Ollaya: Run Jev-Style Decision Models Locally (Ollama-Like)

Ollaya is an open-source local runtime for decision models — not LLM chat, but models that take a state (text, an email, a ticket, or any JSON) plus typed questions (choice, score, or noul — a yes/no probability) and return calibrated probabilities in a single forward pass, in milliseconds. Its CLI and daemon mirror Ollama's UX almost exactly: pull a model by name, then run or serve it locally. Ollaya hit the Hacker News front page around September 25-26, 2026 (462 points, 117 comments), largely because it offers a local, drop-in way to run the kind of decision model TypeSafe AI's hosted Jev "System One" API introduced on September 15, 2026.

A decision model doesn't generate text. It takes a state and typed questions and returns only probabilities (or score distributions) for each — essentially a smart if-statement. That makes Ollaya useful for:

- Support ticket triage — judge refund intent, urgency, and churn risk in one call
- Email and ticket routing — score priority and sentiment to route to the right queue
- LLM guardrails — check inputs and outputs for risky intent or policy violations before or after generation
- Smart if-statements in workflows — replace regex or rule-based branching with a calibrated, probability-backed decision

Ollaya is open source under Apache 2.0 (per-model licenses vary) and free to use, with no per-token fees.

The setup flow mirrors Ollama closely. After installing, start the local daemon with ollaya serve and pull a model with something like ollaya pull laya. To try it out, run ollaya run laya --preset triage "I was charged twice for my subscription this month and want a refund." — a single call returns intent refund 1.00, is_urgent no 0.88, frustration 1.76/3 0.36, refund_requested yes 0.90, and churn_risk no 0.61. From an application, call the HTTP API directly (POST /api/decide, or the Jev-compatible POST /v1/systemone), or point TypeSafe's official SDK (confirmed working at 0.7.1) at Ollaya by setting TYPESAFE_BASE_URL=http://localhost:11435. For Claude Code, Claude Desktop, or Cursor, add it as an MCP server with claude mcp add ollaya -- ollaya mcp so an agent can call decision models directly. Manage models and the server with ollaya list (pulled models), ollaya ps (running models), ollaya show (details), and ollaya stop / ollaya rm. Behavior can be tuned via OLLAYA_HOST, OLLAYA_MODELS, OLLAYA_KEEP_ALIVE, OLLAYA_DEVICE, and OLLAYA_API_KEY.

Ollaya sits between local LLM runtimes and TypeSafe's hosted decision-model API — it's neither.

Worth noting: Ollaya is not an official TypeSafe AI product — it's a third-party open-source project that happens to implement a wire-compatible API against the hosted System One endpoint TypeSafe Jev announced on September 15, 2026. LLM structured output/JSON mode is sometimes used for similar tasks, but that approach forces a large generative model into a fixed output shape, whereas Ollaya's small, purpose-built models return probabilities in a single forward pass — an order of magnitude or more cheaper and faster.

- Good fit: teams triaging support tickets or routing email at volume who want fast, cheap, typed judgments repeated many times over
- Good fit: developers who want to bolt a probability-based guardrail onto an LLM agent's inputs or outputs
- Good fit: teams already using TypeSafe Jev who want to run it offline, self-hosted, or at close to zero cost
- Not a fit: use cases centered on generating or summarizing free text — Ollaya returns probabilities, not prose
- Not a fit: low-volume, low-frequency workloads where calling a hosted API is simpler than running local infrastructure
- Not a fit: teams that need a proven, long-maintained project — Ollaya only became widely known in September 2026

Ollaya keeps Ollama's familiar CLI experience but points it at a different kind of model: one that returns calibrated probabilities from a state and typed questions, instead of generating text. Because it exposes an endpoint wire-compatible with TypeSafe Jev's hosted API, it's a plausible option for teams that want to self-host support-ticket triage or workflow branching logic instead of depending on an external API. For more on the API this project mirrors, see TypeSafe Jev: The System One Decision Model Explained and the TypeSafe Jev Practical Guide.