How Sovariq keeps builds fast and cheap

Most of what a model call costs is what the model reads, and in a long build it reads the same things again on every step: its instructions, the definitions of its tools, the playbooks it loaded and the conversation so far. Sovariq is built so that this part stays small and is read from the provider's prompt cache (about a tenth of the normal price), and so that each step does more of the work.

#Tools on demand

Sovariq can use more than 400 tools: tables, rules, storage, functions, auth, the App Builder and more. Sending all of their definitions with every call would cost tens of thousands of tokens, so each call carries a small core (reading the schema, the builder's file tools, its own planning tools) and the rest are found when needed: Sovariq matches your message against every tool's name, description and keywords (and your table names), and names the few that fit. find_tools finds others by description.

With deferred loading (on Anthropic models, switched on by Sovaria staff) a tool's definition arrives inside the result of the call that found it, instead of being added to the list at the start of the prompt. The start of the prompt never changes, so it stays cached for the whole conversation. In Sovaria's token simulator this cut the cost of a long, varied conversation by about 20% and of a one-prompt app build by about 11%, mostly through fewer cache writes.

The playbooks follow the cross-vendor Agent Skills format (SKILL.md: a name and a description up front, the instructions loaded only when they fit). The public ones — the SDKs, schema design, row rules, functions, storage, realtime, vectors, collections, schedules, queues, imports and backups — are packaged as sovaria/agent-skills for Claude Code, Codex, Cursor and other agents that read SKILL.md. Your own project agents can import SKILL.md skills too (Agents › Playbooks › Imported skills): the agent sees each skill's name and description and loads the rest when a request needs it; scripts inside an imported skill don't run. Up to 8 skills per agent, each SKILL.md at most 40,000 characters and 200,000 characters together.

#Playbooks in sections

Sovariq's playbooks (how to design tables, set up sign-in, build a dashboard, call the SDK from Python…) are its know-how. Each message names the ones that fit. A short playbook loads whole; a long one loads its introduction, its outline and the sections that match what you asked (plus the sections it must always follow), and Sovariq asks for any other section by name. On typical requests this sends about a quarter less playbook text, and up to 70% less for the SDK references.

#A conversation that stays short

The model sees the newest steps in full and older ones shortened: a file it read becomes its first lines and a note to read it again before editing, a file it wrote becomes its size, older notes keep their first line. Past a budget, the oldest steps are folded into a short summary by a smaller model. Facts Sovariq writes down with note (a cause it found, a decision and why) stay with it through every fold, so it never works them out twice. Files it already read are not read again unless something changed.

#Ready-made screens

Every app starts with a component kit: data tables, forms with validation, charts, kanban boards, detail drawers, stat cards and sign-in screens, all bound to your tables and following your theme (Component kit: tables, forms, charts and boards, ready to place). Sovariq places and configures them instead of writing each screen by hand, so a list page is a few dozen lines instead of several hundred: fewer tokens written, and fewer build errors to fix afterwards.

#Auto-fix after a build

When a build changed your app, Sovariq doesn't stop until the app builds and its preview shows no errors. If the last build failed, or the preview logged an error, or the files changed since the last check, it starts a fix round on its own: it builds, reads the errors, fixes the first one where it starts and builds again. An Auto-fix card in the chat shows each round and how it ended.

Auto-fix only fixes errors. It doesn't add features or restyle anything, and Stop ends it like any run.

#Seeing what it costs

The meter in the message box shows the credits a run is using as it goes and what is left; Sovariq credits and limits explains how runs are estimated, held and charged. Sovaria's own prices are set so that each part of a call (input, output, cache reads and writes) costs you at least twice what it costs Sovaria, and cache reads are charged at their lower price.

#Settings for Sovaria staff

Staff switch these in Control Center › Platform › AI models › Sovariq › Efficiency, with a reason that is kept and audited; every replica applies a change within 30 seconds.