Search

Interactive CLI to scaffold AI agent infrastructure for new projects

Authoring agent: RecordDriver over local Playwright + Claude. The only package allowed to depend on @anthropic-ai/sdk and playwright (both lazy-imported).

Run a prompt through a LangChain (system + human) chain over Gemini on Vertex AI; optional LangSmith tracing.

Agent-focused semantic mapping and language services for mixed-language codebases.

Modular Python Agent Framework

Pluggable memory system for LLM agents

Deterministic spatial multi-agent environment with an LLM-agent framework, verifiable rewards, and a connect-out benchmark platform

Per-API-call token cost decomposition for Claude Code, Claude desktop agents, and Cursor.

AI model: mankadevansh/promptflow-agents

Soul Compiler by Koji Limited (Hong Kong AI venture studio) — compiles character DNA for AI agents, not prompts.

A passive policy-enforcement layer for agent tool calls

A Claude-Code-first, external control-plane template for regulated small teams that need auditable, human-gated, containerized AI-assisted development without embedding governance machinery inside target repositories.

Practical loop engineering for AI coding agents: scoped task records, orchestrator/maintainer/engineer roles, fresh verification evidence, and review gates — so complex work runs for hours without drifting scope or grading its own exam. Markdown-first; works with OpenCode, Claude Code, Codex, Copilot, and Cursor.

Verdict policy language (VPL) lexer, parser, and compiler to VIR. Also compiles the JSON policy form.

Verdict embedded evaluator: VIR interpreter, decision combining, tracing, caches. Zero network calls; runs on Workers, Lambda, Bun, Deno, Node.

Verdict storage abstraction and adapters: in-memory, SQL (SQLite/D1/Postgres via minimal drivers), and object storage (filesystem, R2/S3-compatible).

Verdict async audit pipeline: batching sink with bounded buffering, retry, drop accounting, and queue adapters. Decisions never block on audit I/O.

Verdict client SDK for JavaScript/TypeScript: one interface (check, batchCheck, authorize, explain, simulate) over embedded engines, hybrid SyncedVerdict, or a remote PDP.

Verdict control-plane services: policy lifecycle (draft/validate/simulate/publish/rollback), tenant bundle builder with Ed25519 signing, API key management.

Verdict hybrid-mode sync client: local evaluation over bundles synced from a control plane. ETag polling, SSE, WebSocket, and push transports; signature verification, replay protection, staleness policies, rollback ring.