RWKV Agent Projects

Cross-platform RWKV AI assistant and Agent desktop client

A Tauri 2 desktop client built with Rust and TypeScript. It combines local RWKV inference with remote AI APIs and provides a personal assistant, coding agent, knowledge-work agent, and Markdown-defined custom agents. The application also integrates long-term memory, tool-enabled workflows, Skills, MCP, and remote access through phone browsers and messaging platforms.

Local RWKV CLI and agent runtime

A local RWKV command-line application for Apple Silicon macOS. The Go layer handles conversation transactions, session persistence, and terminal interaction, while a separate dynamic library handles tokenization, sampling, and MLX inference. The runtime does not depend on Python, PyTorch, an HTTP service, or external processes.

Auditable RWKV code-diagnosis Agent runtime

A code-diagnosis Agent runtime with replayable traces, evidence validation, structured tool actions, and verified task checkpoints. It supports direct RWKV adapters, OpenAI-compatible RWKV APIs, and deterministic replay.

Persistent and verifiable RWKV long-horizon agent runtime

A Python runtime for persistent, verifiable long-horizon execution with RWKV. It stores immutable goals, task graphs, attempts, evidence, validation results, checkpoints, events, and per-request sampling decisions in SQLite. The runtime combines RWKV planning and action selection with a scoped file, command, and evidence harness, deterministic postcondition checks, bounded working memory, retry and replan handling, and resumable execution after interruption.

Typed RWKV decision infrastructure

An experimental Jev-style decision model built on RWKV-7 G1k 3B. It handles typed Choice, Noul, and Score questions and returns constrained probability distributions without generating answer tokens. It includes state tuning, recurrent branching, a local SDK, HTTP service, WebUI, and reproducible evaluations.

RWKV-powered web and document research assistant

A parallel web and document research assistant powered by RWKV.

RWKV structured decision engine

A Jev-class RWKV-7 decision model compatible with TypeSafe's /v1/systemone contract. It reads typed Choice, Noul, and Score decisions from a constrained softmax over an exact label set and provides local CUDA and CPU backends.

RWKV persona-state training and hot-swap runtime

An RWKV-7 initial-state tuning and serving project. It freezes the base-model weights, trains compact S0 states for personas or task modes, and hot-swaps them per session through an OpenAI-compatible API and WebUI. It also provides state mixing, low-rank and int8 compression, O(1) recurrent multi-turn sessions, evaluation tools, and reproducible state-behavior research.

Zero-generation structured decision runtime for RWKV-7

A TypeSafe Jev-compatible structured decision runtime for RWKV-7. Instead of autoregressive text generation, it scores a fixed candidate set at decision slots and returns normalized probabilities for Choice, Score, and Noul primitives. It includes a Python SDK, FastAPI service, WebUI, model downloader, evaluation cases, and CPU/CUDA execution strategies without requiring fine-tuning or distillation.

Persistent RWKV cognitive-runtime research

A research project for an always-running RWKV cognitive runtime that ingests system events, organizes persistent memory, routes tools, and acts on the local environment through extensions. The repository includes runtime, memory, policy, DSL, training, and deployment foundations.

Self-evolving local-first RWKV model router and Agent gateway

A production-oriented smart-routing component that classifies LLM requests into R0-R3 tiers using rules and a resident RWKV 0.1B classifier, then routes them to local RWKV models or cloud APIs. It ships as a Rust library, C ABI library, Python and Node packages, and a standalone OpenAI/Anthropic-compatible HTTP and MCP gateway.

Local RWKV7 Agent runtime and tool-calling experiment

An experimental extension of rwkv-agent-kit that runs RWKV7-G1j-13.3B locally through ai00-core and web-rwkv. It combines local inference, multi-turn conversation, SQLite memory, tool registration, and a minimal runnable ReAct-style Agent loop.

RWKV-7 semantic embedding and Agent model-routing research

A reproducible research project for extracting semantic embeddings from RWKV-7 hidden and WKV states with supervised task-specific projectors. It evaluates semantic similarity, topic clustering, and task classification, and documents deployment of the classification pipeline as the smart model router in the Ai00-X desktop Agent.

RWKV tool-use and Agent capability benchmark

An RWKV-oriented fork of Primitive Bench that evaluates tool selection, file operations, multi-turn tool feedback, truthful reporting, prompt-injection resistance, and final submission across structured and raw-completion protocols.

RWKV tool-calling benchmark and training research

A reproducible research repository for RWKV tool selection, argument filling, abstention, and multi-step capability. It includes an 82-case judge, raw generations, scoring code, state-tuning and LoRA artifacts, GRPO experiments, checkpoint surveys, probes, schemas, and training recipes.

Local-first RWKV agent and web research system

Provides multi-turn RWKV conversations, tool calling, web search, knowledge-base search, long-form question answering, evidence selection, citation generation, and reproducible evaluation. Its main interfaces are a Rust terminal client and an RWKV Agent HTTP backend. RWKV Search is its internal retrieval subsystem; the former rwkv-search repository now redirects to this repository with HTTP 301.

Auditable RWKV web-retrieval agent

A runnable public-beta retrieval agent with RWKV as the decision core. RWKV decomposes tasks, selects tools, queries and sources, reviews evidence, decides whether to replan or finish, and writes the final answer. The engineering layer executes tools, fetches and cleans pages, maintains per-task evidence ledgers, records traces, and controls resources without replacing the model's retrieval route.

RWKV browsing-Agent evaluation package

A reproducible BrowseComp-Plus package covering the official 830 questions. Its evaluation Agent calls search, get_document, and final_answer in a bounded loop, aggregates parallel candidates, stores results in PostgreSQL, and exports the official judge format.

Primitive Agent and tool-use benchmark

A dependency-light benchmark for models between text prediction and full Agent behavior. It provides task-scoped tools, emulated environments, bounded multi-turn execution, exact submission scoring, open probes, isolation tests, JSON reports, and HTML traces.

Conditional web-search agent workflow

An HTTP API workflow in which a small model decides SEARCH or NO_SEARCH, optionally rewrites the query and calls Exa, and a larger model answers from the retrieved material.

RWKV agent harness with persistent state and tool use

A TypeScript agent harness for local RWKV inference. It combines fixed-size RWKV state and checkpoint persistence with a bounded AgentLoop, structured tool-call grammar, file and shell tools, configurable agent packages, CLI agent mode, and evaluation coverage. The repository currently emphasizes storytelling but also includes default, coder, envoy, and storyteller agents.

Lightweight MVP for validating RWKV agent capabilities

A minimal validation build extracted from EvoOne. It retains the agent loop, tool calling, and critical execution path, and currently provides the exec and web_fetch tools.