English-language public beta · MCP-ready

Verifiable execution
for AI agents

Agents propose. Rulith verifies evidence and policy, gates external actions, and records an execution receipt you can inspect.

Start building free Read the quickstart

Built for teams worldwide. Hosted primarily in the United States. Connect Claude Code, Codex, or any MCP client.

The missing execution layer

Turn model output into controlled work

A model can suggest the next step. Production systems still need to know what evidence supports it, which policy applies, whether the action is allowed, and what actually happened.

Before execution

Rulith evaluates facts against installed Knowledge and Constitution packages. Missing evidence remains a visible gap instead of being guessed away.

After execution

Workers return structured results. Rulith records what ran, the evidence it produced, and whether the task satisfied its acceptance criteria.

How it works

One execution path, five accountable steps

  1. Submit a real taskYour agent proposes facts, goals, or the next allowed action through MCP.
  2. Check the evidenceModel assertions stay distinct from source-attested data. Weak inputs cannot silently become strong conclusions.
  3. Apply Knowledge and ConstitutionThe installed capability determines what can be concluded and which conditions must hold.
  4. Gate and execute the actionA worker performs the approved external call with bounded credentials and returns a structured result.
  5. Inspect the receiptThe run keeps its conclusions, evidence chain, actions, outcomes, and acceptance result together.
Agent configuration

Install the capability. Keep the agent generic.

Workflow behavior lives in explicit configuration, not in a one-off system prompt. Each layer has one job and one owner.

Knowledge + Actions

Vocabulary, rules, and the governed operations the agent can request.

Sources

Where evidence comes from and which facts each source may attest.

Worker Tools

Versioned capabilities implemented locally through bounded Adapters.

Constitution

Independent prohibitions and points that require human approval.

See the system work

Start with a workflow you can verify end to end

5-minute local run

Verified JSON calculation

A local REPL and Worker read input.json, let the board derive an exact result, write output.json, and finish only after an independent read-back matches.

Run it locally → · Inspect the result →

Design-partner workflow

Software release control

Collect test, review, and approval evidence; derive release readiness; block deployment when required evidence or authorization is missing.

Discuss a pilot →

Public beta

Use Rulith Cloud free during public beta

Usage limits apply. Pricing will be introduced with advance notice. You will never be charged without opting in.

Create an agent Explore the docs

Bring one agent workflow

Connect it, run real work, and tell us where the evidence, policy, or execution model breaks. Early feedback directly shapes the public API.

Start building free Send feedback

Developer questions and design-partner conversations: contact@rulith.com