組み立て · assemble

Build a composition.

Add rolls from the shelf. The board surfaces what they need and resolves it from the catalog, separates what a licence closes from what nobody has determined yet, and tells you what the finished stack provides.

Try the free Local Chat starter

Delivery model

84 of 93 rolls are open both ways, so the choice stays open until you make it. Nothing is hidden from the shelf — the board warns instead.

Requirements

All met

most rolls require nothing, which is normal

Availability

Not chosen

both models stay open until you choose

Footprint

3 cores · 5 GB RAM

GPU required

Review RecipeReview the open Recipe facts below. Open the Kitchen portal when you want to save the Recipe or work with Tsugiita.

The shelf · 95 rolls

Hosomaki

Core orchestration and inference

Model serving, gateways, automation, and the core services that other recipes build around.

AnythingLLM

Retrieval-augmented chat workspace bundling document ingestion, embedding, vector storage and multi-provider model connectivity into one container. Answers cite the source passages they were drawn from.

2 GB

Gemma

Open-weight text generation models from Google DeepMind, built from the research behind Gemini and released in sizes that run on a single machine. Executed here by the Ollama runtime.

9 GB

LiteLLM

Gateway fronting more than 100 model providers behind a single OpenAI-compatible API, adding routing rules, fallback chains, retries, per-key spend limits and request logging.

1 GB

n8n

Workflow automation platform with 400+ service integrations, extensible in JavaScript. Runs event-driven and scheduled flows across APIs, databases and model endpoints.

closed0.5 GB

Ollama

Local runtime that downloads, quantises and runs open-weight language models, exposing them over its own HTTP API and an OpenAI-compatible endpoint. Also serves embedding models.

GPU4 GB

Ray

Distributed execution framework for Python that spreads tasks, stateful actors and machine-learning workloads across a cluster, with libraries for tuning, training, data processing and inference.

8 GB

Temporal

Durable execution engine running long-lived workflows as replayable code, persisting every step so processes survive worker restarts, retries and version changes without losing their place.

2 GB

Text Generation Inference

Inference server from Hugging Face for text generation models, using tensor parallelism, continuous batching and paged attention to hold throughput steady as concurrency rises.

GPU4 GB

Triton Inference Server

Inference server from NVIDIA hosting models from several frameworks — TensorRT, PyTorch, ONNX Runtime, Python — behind one HTTP and gRPC endpoint, with dynamic batching and per-model metrics.

GPU8 GB

vLLM

Inference engine for large language models built around PagedAttention, which pages the attention cache to raise batch sizes and throughput. Exposes an OpenAI-compatible API.

GPU8 GB

Futomaki

Databases and knowledge stores

Relational, vector, graph, and object storage for retrieval, memory, and durable application data.

ArangoDB

Multi-model database holding documents, graphs and key-value pairs in one engine, queried through a single declarative language, AQL, that joins and traverses across all three in one statement.

closed4 GB

Chroma

Vector database aimed at retrieval-augmented applications, storing embeddings alongside their source text and metadata and answering similarity queries with metadata filters applied.

2 GB

Infinity

Inference server for embedding, reranking and CLIP models, loading checkpoints from Hugging Face and answering over an OpenAI-compatible API with dynamic batching across concurrent requests.

2 GB

MariaDB

Relational database server forked from MySQL and community-developed since, with pluggable storage engines, wire compatibility for MySQL clients and replication built into the server.

1 GB

Memgraph

In-memory graph database speaking Cypher, built for graphs that change constantly — streaming updates, real-time traversals and analytics over relationships as they form.

closed4 GB

MinIO

Object storage server with an S3-compatible API for on-premises and edge deployments, providing erasure coding, object versioning, write-once retention locks and bucket replication.

1 GB

RustFS

Maintained Rust object storage with a tested S3-compatible API subset for bucket and object workflows.

2 GB

Neo4j

Graph database built on the property graph model and the Cypher query language, used for knowledge graphs, network analysis and relationship-heavy queries that join badly in SQL.

4 GB

pgAdmin

Web-based administration and query tool for PostgreSQL, covering schema browsing, SQL editing, graphical query plans, session and lock monitoring, and role and permission management.

0.5 GB

PostgreSQL

Relational database with strong SQL conformance, MVCC transactions and an extension system. This image ships the pgvector extension, adding vector columns and approximate nearest-neighbour search.

2 GB

Qdrant

Vector database written in Rust, pairing dense and sparse vectors with structured payloads so that metadata filters and nearest-neighbour search are resolved together in a single query.

4 GB

Rclone

Sync and transfer tool for more than 70 cloud and object storage backends, with copy, sync, move and verify operations, and a browser interface served by its own remote-control daemon.

0.25 GB

Redis

In-memory data store used as a cache, message broker and ephemeral database, offering strings, hashes, lists, sets, sorted sets and streams, with optional persistence to disk.

1 GB

Redis Insight

Graphical client for Redis covering keyspace browsing, command execution, memory analysis, slow-log inspection and live profiling, with per-type viewers for streams, JSON and search indexes.

0.25 GB

Supabase

Backend platform assembled around PostgreSQL, adding generated REST and GraphQL APIs, authentication, authorisation through row-level security, file storage and realtime change streams.

4 GB

Weaviate

Vector database with a typed schema that fuses nearest-neighbour and keyword search into one hybrid query, with optional modules that generate embeddings at ingest time.

4 GB

Gunkanmaki

Security, identity and governance

Identity, secrets, policy, and protective controls for systems that need a stronger security boundary.

authentik

Identity provider and single sign-on server supporting OAuth2, OIDC, SAML, LDAP and proxy-based authentication, with a flow engine that composes login, enrolment and recovery as editable stages.

2 GB

Docker Socket Proxy

HAProxy-based filter placed in front of the Docker daemon socket that permits only whitelisted API endpoints, so a container can read runtime metadata without gaining control of the host.

0.0625 GB

The Honorable Halt

Safety mechanism that lets an automated process stop itself at a boundary it should not cross, holding a registry of invariants and escalating through four graduated responses with the reasoning recorded.

0.5 GB

Infisical

Secrets and configuration platform with per-environment scoping, versioning, access control and injection into running processes, plus scanning for credentials committed to source control.

1 GB

Keycloak

Identity and access management server implementing OIDC, OAuth 2.0 and SAML, with user federation, identity brokering, fine-grained authorisation and per-realm isolation of tenants.

4 GB

Governance Ledger

Append-only, hash-chained record of governance decisions in which each entry links to its predecessor by hash, so the sequence cannot be altered afterwards without the break becoming detectable.

0.25 GB

Portainer

Management interface for Docker, Docker Swarm and Kubernetes covering container, image, volume, network and stack lifecycle, with role-based access across several attached environments.

0.5 GB

Sealed Secrets

Kubernetes controller that decrypts SealedSecret resources into ordinary Secrets inside the cluster, so encrypted secret manifests can be committed to a Git repository safely.

1 GB

SOPS

Command-line editor for encrypted files that encrypts values in YAML, JSON, ENV and INI while leaving keys readable, using age, PGP or a managed key service, so diffs stay meaningful in Git.

0.5 GB

Uptime Kuma

Self-hosted uptime monitor that checks HTTP endpoints, TCP ports, DNS records, certificates and container health on a schedule, with public status pages and a wide set of notification integrations.

0.5 GB

Vault

Secrets management server providing static and dynamic credentials, encryption as a service, PKI issuance and leasing, with pluggable authentication methods and policy-based access control.

closed2 GB

Vaultwarden

Rust implementation of the Bitwarden server API, compatible with the official Bitwarden clients and extensions, covering vaults, organisations, collections, sends and file attachments.

0.5 GB

Inari

Observability

Metrics, logs, traces, and operational visibility so you can see what the recipe is doing after it runs.

cAdvisor

Container resource monitor that discovers running containers automatically and exports per-container CPU, memory, filesystem and network statistics in Prometheus format and a built-in live view.

0.5 GB

Grafana

Dashboard and visualisation server that queries time-series, log and SQL data sources directly and renders panels, with alert rules, notification routing and dashboards defined as JSON.

1 GB

Homepage

Configuration-driven start page for a self-hosted deployment, showing service links with live status and widgets that pull counts and health directly from each service's own API.

0.25 GB

Langfuse

Observability platform for language-model applications, recording traces of nested calls with prompts, completions, latency, token counts and cost, plus prompt versioning and evaluation runs.

2 GB

Loki

Log aggregation system that indexes only a small set of labels per stream and stores log content compressed in object storage, keeping cost proportional to volume rather than cardinality.

2 GB

Node Exporter

Prometheus exporter for host-level metrics, reading CPU, memory, disk, filesystem, network and kernel statistics from the operating system and exposing them on a scrape endpoint.

0.125 GB

OpenTelemetry Collector

Vendor-neutral pipeline for traces, metrics and logs, receiving telemetry in several protocols, processing it through batching, filtering and attribute rules, and exporting to one or more backends.

1 GB

Prometheus

Time-series database and monitoring server that scrapes metrics endpoints on a schedule, stores samples locally, and evaluates PromQL queries, recording rules and alert conditions against them.

2 GB

Promtail

Log shipping agent for Loki that discovers log files and container streams, attaches labels, applies parsing stages and pushes lines to an ingest endpoint. Reached end of life in March 2026.

0.125 GB

Otsumami

Utilities, publishing and embedded stores

Useful supporting tools that extend a recipe without becoming its primary architectural center.

DuckDB

In-process analytical database that runs OLAP queries over Parquet, CSV and its own columnar files without a server, embedding directly into the process that queries it.

2 GB

Ghost

Publishing platform for blogs, newsletters and paid memberships, combining a block editor with built-in email delivery, subscriber management and native subscription billing.

1 GB

HedgeDoc

Real-time collaborative Markdown editor where several people edit one note simultaneously in the browser, with live preview, slide mode, diagrams and per-note access control.

0.5 GB

Hugo

Static site generator written in Go that renders Markdown content into a complete HTML site, with layouts, taxonomies, multilingual support and asset processing built in.

0.5 GB

Material for MkDocs

Documentation site generator built on MkDocs with the Material theme, producing a static site from Markdown with client-side search, versioned navigation and code annotation.

0.5 GB

SearXNG

Metasearch engine that queries many upstream search engines on the user's behalf and merges the results, forwarding no identifying information and storing no search history.

1 GB

SQLite

Embedded relational database that stores an entire database in one file and runs inside the calling process, with no server, configuration or administration to manage.

0.5 GB

Wiki.js

Wiki platform with a WYSIWYG and Markdown editor, page versioning, granular permissions and optional two-way synchronisation of content with a Git repository.

1 GB

WordPress

Content management system for sites, blogs and stores, extended through themes and a very large plugin ecosystem, with a block editor and a REST API for headless use.

1 GB

Chirashi

Data science and ML operations

Notebook, experiment, analytics, and machine-learning workflow tools for exploratory and repeatable work.

BentoML

Python framework for turning trained models into HTTP inference endpoints, handling request batching, multi-model composition, dependency capture and a versioned, runnable build artifact.

4 GB

Dask

Parallel computing library for Python that partitions arrays, dataframes and custom task graphs across cores or a cluster, mirroring the NumPy and pandas APIs on larger-than-memory data.

8 GB

Docling

Document conversion toolkit parsing PDF, Office, HTML, image and audio inputs into a structured representation, recovering layout, reading order and table structure for retrieval pipelines.

3 GB

JupyterLab

Browser-based development environment presenting notebooks, code editors, terminals and data files in one tabbed workspace. This image variant ships a Python data-science stack preinstalled.

4 GB

Kubeflow

Machine-learning platform for Kubernetes covering pipelines, distributed training operators, hyperparameter search, model registry and multi-user notebook servers under namespace isolation.

8 GB

Metabase

Business intelligence server for querying databases through a visual question builder or SQL, assembling results into dashboards with filters, scheduled subscriptions and row-level permissions.

4 GB

MLflow

Platform for tracking machine-learning experiments, logging parameters, metrics and artifacts per run, with a model registry holding versions, stage transitions and lineage back to the producing run.

2 GB

Panel

Python framework for building interactive dashboards and applications from analysis code, connecting widgets to plots and tables from the main visualisation libraries without front-end code.

2 GB

Temaki

Builder, authoring and prototyping

Developer workspaces, visual builders, scaffolding, and rapid-prototyping tools for making new systems.

Dify

Platform for building language-model applications, combining a visual workflow canvas, prompt orchestration, retrieval datasets, tool-calling agents and observability in one interface.

2 GB

docassemble

Expert system for guided interviews that gathers answers question by question and assembles them into PDF, DOCX or RTF documents, with signatures, conditional logic and multilingual support.

2 GB

Flowise

Visual builder for language-model applications, wiring retrieval, memory, tools and model nodes on a canvas and exposing the finished flow as an API endpoint or embeddable chat widget.

2 GB

Gitea

Self-hosted Git service with repository browsing, pull requests, issues, releases, package registries and a built-in CI runner, in a single binary with modest resource needs.

1 GB

OpenUI

Experimental tool that turns a natural-language description of an interface into live-rendered HTML, iterating on the result through follow-up instructions and converting it to framework code.

1 GB

SonarQube

Static analysis server that inspects source code for bugs, security issues, duplication and maintainability problems across many languages, gating pull requests on configurable quality thresholds.

4 GB

code-server

VS Code running as a server and reached through a browser, so editing, terminals, debugging and extensions happen on the remote machine while only the interface is local.

2 GB

Windmill

Workflow and job platform that turns scripts in Python, TypeScript, Go, Bash and SQL into scheduled jobs, APIs and approval flows, with auto-generated forms and a flow builder.

closed2 GB

Uramaki

Visual and image tooling

Image-generation and visual-media tools for creative and automated production workflows.

Stable Diffusion web UI

Web interface for Stable Diffusion image generation, covering text-to-image, image-to-image, inpainting, upscaling and model switching, with a large third-party extension ecosystem.

GPU8 GB

Lychee

Self-hosted photo and visual-asset manager for uploading, organizing, browsing and sharing media.

2 GB

ComfyUI

Node-based interface for diffusion image and video generation, where sampling, conditioning and model loading are explicit graph nodes that can be saved, shared and re-executed.

GPU8 GB

ExifTool

Command-line metadata engine for reading, writing and normalizing EXIF, IPTC, XMP, GPS, ICC, maker notes and other embedded metadata across photographs, illustrations and many other file types.

0.5 GB

FFmpeg

Command-line suite for decoding, encoding, transcoding, filtering and streaming audio and video, supporting a very wide range of codecs and container formats.

2 GB

ImageMagick

Command-line image toolkit reading and writing over 200 formats, with conversion, resizing, composition, colour correction and effects, suitable for scripted batch processing.

1 GB

Rclone

Command-line tool for copying, syncing and mounting data across more than 70 cloud and object storage backends, with checksum verification, encryption and bandwidth limits.

0.5 GB

Remotion

Framework for producing videos from React components, where compositions are rendered frame by frame in a headless browser so video output can be generated from application data.

closed4 GB

Nigiri

Speech and conversational interfaces

Speech recognition, synthesis, and interaction tools that give a recipe ears, a voice, or a chat surface.

Open WebUI

Self-hosted chat interface for local and remote language models, adding document retrieval, model switching, per-user workspaces, tool calling and an extensible pipeline layer.

1 GB

OpenVoice

Voice cloning system that reproduces a reference speaker's tone from a short sample and applies it to synthesised speech, including in languages absent from its training data.

GPU4 GB

Piper

Fast neural text-to-speech engine that runs entirely on CPU, producing natural speech from small ONNX voice models at real-time speed on hardware as modest as a single-board computer.

1 GB

Whisper

Speech recognition service wrapping OpenAI's Whisper models, transcribing and translating audio across many languages with timestamps, served over an HTTP endpoint.

GPU4 GB

Sashimi

API surfaces and developer docs

Tools for exposing, exploring, documenting, and working directly with the APIs inside a running system.

Docusaurus

Documentation site generator built on React and MDX, with versioned docs, i18n, automatic sidebars and search, producing a static site that can embed interactive components.

1 GB

FastAPI

Python web framework for HTTP APIs that derives validation, serialisation and OpenAPI documentation from type annotations, running async on an ASGI server.

3 GB

Redoc

Renderer that turns an OpenAPI description into a three-panel reference page, with a navigable operation list, schema expansion and per-operation request and response samples.

0.25 GB

Swagger UI

Renderer that turns an OpenAPI description into browsable documentation with an execution panel, so each operation can be called against a live server from the page itself.

0.25 GB

Shoyu

Backup and recovery

Backup, restore, and resilience tooling for the data and state a recipe cannot afford to lose.

borgmatic

Configuration-driven front end for BorgBackup that runs deduplicated, compressed, encrypted backups on a schedule, adding retention pruning, consistency checks and database dumps.

0.5 GB

Duplicati

Backup client with a browser interface that produces encrypted, incremental, deduplicated backups to a wide range of cloud and network storage backends on a schedule.

1 GB

pgBackRest

Backup and restore tool for PostgreSQL supporting full, differential and incremental backups, parallel transfer, delta restore and point-in-time recovery from archived write-ahead logs.

0.5 GB

restic

Backup tool producing encrypted, deduplicated, content-addressed snapshots to local, SFTP, S3 and other backends, with fast restores and repository integrity checking.

0.5 GB

Dragon

Networking and service mesh

Edge proxies, routing, ingress, and service-to-service networking for connecting the parts safely.

Caddy

Web server and reverse proxy that obtains and renews TLS certificates automatically, supports HTTP/3, and is configured either by a short text file or entirely over an API.

0.25 GB

Istio

Service mesh that puts a proxy alongside each workload to handle routing, retries, mutual TLS, policy enforcement and telemetry without changes to application code.

4 GB

Traefik

Reverse proxy and ingress controller that builds its routing table from live service discovery, so containers declare their own routes through labels rather than static configuration.

0.25 GB

Current Recipe

2 rolls

Ollamahosomaki

Local runtime that downloads, quantises and runs open-weight language models, exposing them over its own HTTP API and an OpenAI-compatible endpoint. Also serves embedding models.

2 cores · 4 GB

Open WebUInigiri

Self-hosted chat interface for local and remote language models, adding document retrieval, model switching, per-user workspaces, tool calling and an extensible pipeline layer.

1 core · 1 GB

Open Recipe Review

Build first. Save when you are ready.

The public Builder stays open: choose Rolls, resolve declared requirements, compare delivery posture, and inspect the footprint without an account. Durable Recipes and Tsugiita live inside the Sushii Kitchen portal.

Every declared Roll requirement is currently satisfied.

This draft is structurally ready for deeper Sushii Intelligence guidance.

Current footprint: 5 GB RAM · 3 CPU cores · 55 GB storage · GPU required.

Ready to ship

Nothing left to resolve.

2 rolls · every requirement met · both delivery models open · 5 GB RAM · GPU required

Review Recipe

The open Builder does not charge for composition facts or the right to use the stack. Durable account work lives in the Kitchen portal; Tsugiita is the optional paid guidance layer.

This composition provides

Chat interfaceEmbeddingsLanguage model API

A running kitchen usually wants

DatabaseNot yet

Chat interfaceIn this build

MonitoringNot yet

Reverse proxyNot yet

BackupsNot yet