Arcanum

One harness, every surface

Sustained agentic work from one source of truth.

Arcanum runs as a long-lived HTTP host, with terminal commands and other clients kept thin over the same API-first core. Its progress-driven model-and-tool loops keep useful work moving until the task completes, an authoritative boundary stops it, or the operator cancels.

Bring the providers you choose

Keep provider choice at the edge.

Configure the OpenAI-compatible endpoints that fit your work, including Ollama through its /v1 endpoint. When installed locally, Claude Code and Codex CLI transports can be added as optional choices.

Arcanum’s /v1 surface is an OpenAI-compatible compatibility subset: a practical way to connect compatible clients, not a promise to reproduce every provider feature.

More than a chat wrapper

A harness for sustained, context-aware, testable work.

Keep work moving with — Persistent agent loops

Let models and tools make progress without an arbitrary turn or hop ceiling, while explicit budgets, safety boundaries, cancellation, and no-progress detection remain authoritative.

Perceive the workspace with — Eye of the World

Give the harness situational directory perception so the work begins with a legible view of its local context.

Delegate bounded jobs with — Apprentices and MCP tools

Hand defined work to sterile child turns and purpose-built tools while the operator's limits stay in force.

Organize and prove the work with — Campaigns, Sessions, Spells, and Trials

Carry durable context forward, reuse deliberate instructions, and evaluate results with Proving Grounds Inquisitors.

Advanced retrieval and layered memory

Reach exact details and the shape of the whole corpus.

Arcanum combines Retrieval-Augmented Generation (RAG), semantic search, hierarchical retrieval, associative memory, and fast local full-text search instead of betting every question on one retrieval strategy.

Semantic RAG with — The Weave and Divination

Embed and retrieve relevant workspace files and session attachments, with explicit material kept authoritative over derived context.

RAPTOR-style hierarchical retrieval with — The Tapestry

Build summary trees over indexed workspaces, attachments, and session history for corpus-level and multi-hop questions beyond flat top-k search.

Associative memory with — Saga

Retrieve durable, provenance-aware memory alongside the exact context already attached to the work.

Durable facts with — Lexicon

Store durable key-value facts with typed provenance and authority lanes that remember nothing without saying where it came from or who may use it.

Local full-text search with — SQLite FTS5

Accelerate inspection and exact text discovery inside Arcanum's encrypted persistence architecture, alongside the Weave's semantic retrieval.

Fully Native AOT compatible

A production harness built for direct deployment.

Native AOT is a hard architectural constraint across the Arcanum host, not a packaging afterthought. The result is fast startup, predictable self-contained deployment, and no separately installed .NET runtime for Native AOT distributions.

Local-first by design

Production control is designed into the harness.

Arcanum keeps durable state in an encrypted SQLCipher Grimoire, protects external artifacts independently, binds to loopback by default, and authenticates every route. Wards can require an operator decision for high-risk work, while Sanctum constrains where Campaign tools may act.

Budgets, audit records, backup, retention, and recovery make sustained agentic work observable and governable without claiming that any system eliminates risk.

Compendium

A visual editor for the configuration.

Compendium is the visual editor for arcanum.json. It helps you configure Arcanum without hand-editing the file.

It does not run inference, open the Grimoire, or own server logic. Those responsibilities remain with Arcanum.