Standards and tools Synapti publishes

Standardized way for AI agents to discover and execute intents against web services. Defines a unified vocabulary for intents, metadata, and execution methods, plus an ODRL-based policy layer for permissions and rate limits. The missing layer between agents and APIs.

An open specification for composable AI agent capabilities. 36 atomic capabilities across 9 cognitive layers, a type-safe workflow DSL, and grounded world modeling with trust-aware conflict resolution. Built on the Grounded Agency philosophy: agent reliability becomes structural, not optional.

Open standard for detecting and scoring manipulation in online content. 20-category framework across 5 composite factors. The substrate Decipon runs on. Anyone can implement, contest findings, or extend the framework.

A four-layer governance architecture for multi-agent AI systems. Authority boundaries, escalation, and constitutional review applied to ensembles where naive agent autonomy produces drift, conflict, or runaway behavior.

Curated collection of Claude Code plugins for AI-augmented development plus advanced analytical and research tasks. Each plugin ships the agents, skills, or commands for one domain. The umbrella distribution channel for the Claude Code work published here.

A standardized framework for enhancing how LLMs process and respond to prompts through composable decorators. Ships an open standard, a Python reference implementation, and an MCP server that brings decorator behavior into Claude Desktop and other MCP-aware clients.

Behavior-Driven Specification Kit. Specification-first governance for AI-assisted code generation. Eight-phase validator, lifecycle skills, and scope enforcement. Behavior before implementation, uncertainty documentation, multi-pass verification.

Open-source epistemic audit for Claude conversation history. Runs eight audit modules, most built on published AI-safety research (Spiral-Bench, Sharma sycophancy, SycEval, BeliefShift, ITP, MedTrust-RAG), and produces citation-validated HTML reports. For users who want to know whether their assistant is drifting toward sycophancy.

Self-Improving Agentic RAG Engine. Evolves multi-agent RAG pipelines using quality-diversity optimization (Map-Elites) and topology search. Lets the pipeline find its own better shape rather than requiring humans to enumerate candidate architectures.

Provider-neutral coding-agent harness built on deterministic workflow graphs, durable evidence, and fail-closed sandboxed execution. For anyone running coding agents, that means a run either produces a verifiable record of what happened or it does not run at all, and swapping the underlying model provider does not change how the workflow behaves.

A reference architecture for agentic customer support, in TypeScript on Postgres with no orchestration framework. A model classifies each message, retrieves the documents that answer it, and drafts a reply that cites them. The system acts alone only on internal helpdesk changes, and only when two separate gates pass: enough of nine classifier samples must name the same intent, and the retrieved documents must support enough of the claims in the draft, each claim checked separately. Both thresholds start from seed values and are refitted on labeled history. A policy engine that calls no model makes that call. Every reply to a customer waits for a person to approve it.