Features

Every feature reads from the same capability map.

The engine runs local and remote models side by side, with built-in tools, agents that remember and workspaces for teams. You need no subscription to run it, and local mode stays free.

Engine · Capability map

CategoryChosen modelScore

Code generationqwen2.5-coder:14b0.97

Code reviewqwen2.5-coder:7b0.85

Security reviewqwen2.5-coder:7b0.83

Long contextgpt-4o · own key0.88

Creative writinggemma3:4b0.94

Measured, not assumed

The capability map sits under everything on this page.

Every feature here reads from the same place. Soriku benchmarks each model you connect, on your own hardware, across thirteen categories. Each score comes with its spread and the number of tests behind it. What was never measured says so, and what got old is flagged. Routing, verification, gap analysis and team workspaces all read from that map.

One benchmark run per model, thirteen categories, on your hardware
Every score shows its spread and how many tests it came from
Re-run whenever you want, and stale measurements get marked

Chat

Write a short story about a lighthouse keeper who finds a letter in a bottle. Around 300 words, calm tone.

Chosengemma3:4bCreative writing

Highest measured score for creative writing on this machine.

gemma3:4b0.94
qwen3:4b0.86
llama3.1:8b0.79

3 of 9 models compared

Orchestration

Capability benchmark

Measure your models on your machine. 13 categories, re-runnable, yours.

Smart routing

Capability-scored routing per category. Override per prompt or pin globally.

Multi-model verify

Run a prompt through two or three models, return the best or merged answer.

Stack optimiser

Soriku spots gaps in your model coverage and suggests fills.

Models & providers

Ollama (local)

First-class. Auto-discovery of installed models, RAM tuning, mute/reactivate flow.

Claude, OpenAI, Mistral, Groq

Bring your own keys. Per-provider daily budget caps.

OpenAI-compatible endpoint

Drop-in replacement at /api/v1/chat/completions for any SDK or tool.

RAM-aware loading

Soriku knows your hardware and won't load three 8B models at once.

Agents & tools

Agent personas

Persistent identities with their own memory, tone, specialisation.

Pilot mode

Multi-step tool calling. Read files, run shell, fetch web, generate documents, supervised or automatic.

Plan mode

Multi-worker planning with conductor + editor. Cost-balanced or fully local.

MCP integration

Soriku speaks Model Context Protocol. Use it in Cursor, VS Code, or the terminal.

Teams & deployment

Simezu SSO

One login across the Atypisch ecosystem. Shared groups, shared billing.

Shared agents

Roll out an agent persona to your team. Everyone gets the same setup.

Team capability map

Aggregate scorecard across your team's hardware, so the benchmark runs once.

Self-host or hosted

Run on your hardware, or let Soriku Cloud handle it.

In the app

What it looks like at your desk.

Capability map

Summaries0.91

Code0.84

Translation0.78

measured 3 days ago

Routing logRefactor this function

Codeqwen2.5-coder

0.88 · local only

Closed networkNothing left the building6 devices admitted · gate closed for outgoing

New modelgemma3:12b