Utopik Core

The host supplies the integrations

Utopik Core connects model calls, tools and durable state through injected interfaces. The host supplies the chat endpoint, session store, stream store, usage ledger and tool providers. The reference host currently speaks OpenAI-compatible Chat Completions; provider-specific behaviour remains outside the core engine.

One authorisation path for every tool

Tools can arrive through stdio MCP servers, confined skill providers or in-process providers supplied at composition time. All three sources pass through the same name-collision checks and capability gate. An ungranted tool is rejected before its handler runs, including a model-requested tool not admitted to the turn.

A small dependency surface

@utopik/core declares one direct runtime dependency: the official MCP SDK. That SDK brings its own transitive dependencies. Core uses no model-provider SDK. The reference HTTP and server-sent-event adapters use Node’s built-in networking APIs.

Typed streams, interruption and durable state

The runtime accepts injected storage interfaces. The reference host stores sessions, streams, usage, budgets and interruption state in one configured SQLite file, so no separate database server is required.

Core’s recovery tests reopen session history and persisted tool outcomes in a new engine, and preserve ordered stream replay across processes. A process-level test reassigns ownership of an unfinished turn-control record after its owner is killed. Repeated interrupt requests trigger cancellation only once in the interruption test.

The runtime emits typed text and progress events. Usage is recorded per model call; missing token usage is treated as unknown rather than silently counted as zero.

Reference-host synthetic memory measurement

On 3 September 2026, a synthetic tool-using workload measured 138.3 MiB mean observed combined peak RSS across 30 concurrent logical sessions on one reference host, with five sequential fresh-process repeats.

Each simulated-model session used a 150 ms response delay and made real read_file, write_file and node --version tool calls. The workload excluded real model-provider traffic, MCP servers, streaming and multi-turn conversations. This is a reference-host memory measurement for that workload; it is not a Core-only latency, cost, throughput or production-capacity result. Full measurement notes.

Applications own their workflow layer

Core owns model and tool execution and provides a generic goal runner. Applications retain their product-specific workflow, approval interfaces and operating model.