Pricing
These are beta prices. Hosted plans keep your full history on object storage, billed at cost.
Developer
For building and testing locally.
- –The CLI, dev loop, and Studio
- –Unlimited scenarios and invariants
- –Full provenance on every run
- –Agent skills for Claude Code, Codex, and Devin
Team
For running expert systems in production.
- –Managed cloud runtime and deploys
- –Full observation history at object-storage rates
- –Receipts and audit views for operators
- –Verification gates in CI
Enterprise
For regulated and high-stakes domains.
- –Dedicated runtime cells, SSO, and private networking
- –Retention, redaction, and legal hold declared against the ontology
- –Audit exports with full provenance
- –Support from named engineers
Beta prices are locked for design partners until general availability.
Questions
What is an expert system on Tarski?
An app whose decisions come from rules that encode your experts' knowledge. LLMs can read input and suggest actions, your rules decide what to accept, and invariants block outcomes your experts have ruled out. Every decision links back to the inputs that caused it.
Does Tarski replace my coding agent?
No. Claude Code, Codex, Devin, and other coding agents write the Tarski rules against your scenarios and invariants. Tarski checks the result and runs it.
Why is storage effectively unmetered?
The observation log is stored as sealed, content-addressed segments on object storage. The ontology determines the layout, and semantic indexes let queries skip segments entirely. Cold history costs S3 prices and still answers queries, so Tarski keeps everything by default and passes storage through at cost.
Can I migrate an existing Postgres or Firestore app?
Not as a whole. Tarski derives state from an observation log and does not emulate a mutable-state database. Teams usually start with one new decision process, such as an LLM feature that needs hard limits, and expand from there.
What does a passing tarski verify mean?
Every scenario in your set produced its expected outcome exactly, and every named invariant held at every fixpoint, on a deterministic engine. It covers the cases you have written down. It is not a proof about all possible inputs.
What happens when an invariant refuses an action in production?
The action does not run. Tarski records a receipt that names the invariant and the observations involved. Operators see it in Studio, and you can promote the timeline to a scenario in one step.