logicpearl.com

What It Is

LogicPearl positions itself as a “deterministic decision primitive for AI agents.” It is not another chat model. Instead, it compiles business decision rules into versionable artifact bundles, WASM, or native binaries: the LLM is responsible for extracting structured features from natural language, while LogicPearl performs reproducible evaluation. In its refund example, Claude Sonnet 4.5 repeatedly approves the request under the same policy, whereas LogicPearl consistently returns DENY based on the rules days > 29 and changed_mind. This illustrates its core value: separating high-risk decisions from probabilistic generation.

Core Capabilities

Its MCP server provides three tools: evaluate for deterministic evaluation, returning verdict, fired_rules, counterfactual_hints, bitmask, and latency; describe_artifact for returning the feature schema, default action, and extraction prompts; and list_rules for enumerating rules so the LLM can explain results without hallucinating. As for rule generation, LogicPearl can discover rules from labeled traces via decision trees and sequential covering, then compile them into deployable artifacts. Each rule corresponds to one bit in a bitmask, making diffing, testing, signing, and auditing easier.

Pricing and Integration

The main documentation does not disclose commercial pricing or SaaS plans. The source code is marked MIT licensed, and it provides npx @logicpearl/try, @logicpearl/mcp, @logicpearl/browser, and a Rust engine. The installation experience is developer-oriented: a single npx command can write MCP configuration for Claude Desktop and Cursor, and manual configuration is also supported for any MCP-compatible host. The current transport is stdio only, which suits local Agent toolchains; SSE/HTTP and a hosted registry are still on the roadmap.

Pros, Cons, and Limitations

Its strengths are strong determinism, clear explainability, lightweight deployment, no runtime service dependency, and the ability to provide counterfactuals showing “what would make this pass.” It is suitable for auditable scenarios such as healthcare claims, compliance authorization, risk control, refund approvals, and legacy rule governance. The limitations are also clear: v1 has no server-side authentication, and stdio assumes the host is trusted; there is no multi-tenant hosting; artifact registry and signature verification have not yet been implemented; the observer layer is still relatively simple, and extracting features from complex text still requires an LLM, ML, or custom code. If the training traces are wrong, the resulting rules will also be systematically wrong.

Who It’s For and Access from China

LogicPearl is better suited to engineering teams building AI Agents or policy execution systems, rather than ordinary no-code users. The main documentation does not provide details on access from China. Since it can be used via npm/cargo/GitHub, local MCP, and WASM, real-world availability depends on access to GitHub/npm, the model provider, and the host tool; payment information is missing. Alternatives include OPA/Rego, traditional rule engines, Drools, or self-built rule services.