OpenPRX
openprx/prx
Framework IA self-hosted com roteamento multi-provedor e auto-evolução
Ver no GitHubFramework de assistente IA auto-evoluível construído em Rust, com suporte multi-provedor (Anthropic, OpenAI, Google Gemini, Ollama, AWS Bedrock e outros) e multi-canal (Signal, WhatsApp, Telegram, Discord, Slack). Inclui roteador inteligente de modelos, motor de decisão consultivo, agendador de tarefas autônomo (Xin) e console web para gerenciamento.
Instalação
npx convez-ai openprx/prx1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir assistentes IA personalizados e auto-aprimoráveis
- Integrar múltiplos provedores LLM com roteamento inteligente
- Automatizar tarefas via múltiplos canais de mensageria
Veredito

Solução robusta e production-ready para quem precisa de orquestração IA complexa, com governança e self-evolution. Exige experiência em Rust e DevOps; ideal para equipes que querem controle total e extensibilidade.
README
A tradução para o português deste README ainda está na fila.
OpenPRX
Self-hosted AI assistant framework built in Rust. Multi-channel, multi-provider, with built-in self-evolution.
Forked from ZeroClaw and extended with production reliability, governance-aware AI, and a self-evolution system.
Highlights
- Broad LLM provider catalog — Anthropic (Console API key), OpenAI, Google Gemini (API key), Ollama, AWS Bedrock, GLM, OpenAI Codex via Sign in with ChatGPT, local runtimes, and compatible endpoints. GitHub Copilot login is disabled. Provider requests send
User-Agent: prx/<version>. - LLM Router — heuristic routing (capability + Elo + cost + latency), KNN semantic routing (cold-start guard + 100ms timeout fallback)
- Advisory decision engine — bounded intent analysis and turn-outcome records; off by default, never selects tools or grants approval
- Multi-channel messaging — Signal, WhatsApp, Telegram, Discord, Slack, and more
- Built-in tools and integrations — shell, MCP, memory, scheduling, remote nodes, and an integration catalog
- Xin, from the Chinese word for heart, task engine — autonomous heartbeat scheduler with 3 execution modes (Rust/LLM/Shell), 3 built-in maintenance tasks, durable Goal/Step execution, and SQLite persistence
- Web Console — browser-based management interface (
console/) - Remote Nodes — control macOS/Linux/Pi devices via
prx-nodeagent - Self-Evolution — autonomous prompt improvement staged through the skills lifecycle
- Subagent Spawning — ordinary fan-out without a global concurrency cap, config inheritance, and explicit bounded recursive grants; live runs listed and ended with
prx tasks - Extensive automated test suite — unit, integration, PTY, gateway, migration, and security coverage
LLM Router Flags
- heuristic model routing is always available
- semantic KNN scoring activates when embeddings and history are available
Advisory Decision Engine
The advisory engine ([decision_engine]) is off by default
(enabled = false). Off means no analysis brief, telemetry store, or outcome
writes. Tool selection, approval, and admission stay with the tool surface
owner and the run's approval strategy. Enable it only with an explicit
[decision_engine] enabled = true.
Three turn limits live under [decision_engine] and apply whether or not the
engine is enabled:
latency_budget_ms— extra latency one turn may spend on advisory work (default:500, range50–5000)token_ratio_limit— extra tokens as a ratio of the turn's baseline (default:0.35, finite and greater than0)max_candidates— most candidates a turn considers (default:12, range1–32)
The optional [decision_engine.backend] table selects the analysis backend;
without it the built-in backend serves and nothing leaves the machine. A
[causal_tree] table is no longer accepted and is refused at load. See
advisory decision engine and the
configuration reference.
Quick Start
Linux and macOS:
curl -fsSL https://github.com/openprx/prx/releases/latest/download/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"
prx init --interactive
prx daemon
prx chat --message 'What is 2+2?'
Windows PowerShell:
irm https://github.com/openprx/prx/releases/latest/download/install.ps1 | iex
prx init --interactive
prx daemon
Foreground commands (prx chat, prx go, prx doctor, prx status, and
other commands whose stdout is the product) default to warn-level tracing
on stderr, so a scripted prx chat --message 'What is 2+2?' prints only the
reply. Those commands previously wrote tracing to stdout.
RUST_LOG can raise or lower the level (RUST_LOG=info, RUST_LOG=error).
Interactive prx chat writes the same filter to chat.log (WARN+ by
default). prx daemon, prx gateway, and prx channel start keep info
on their current destination.
The installer downloads a platform-specific binary from GitHub Releases, verifies its SHA-256 checksum, and installs it without requiring Rust or Git. See the installation guide for exact versions, source builds, supported platforms, upgrades, and removal.
To build from a checkout with the pinned Rust 1.99.0 toolchain:
git clone https://github.com/openprx/prx.git
cd prx
cargo build --release --locked --bin prx
./target/release/prx init --interactive
Commands
| Command | What it does |
|---|---|
prx chat | Fullscreen interactive session: streaming, tool cards, and saved sessions. --plain or PRX_TUI=0 uses the reedline fallback. --message <text> (-q) runs one turn, prints only the reply on stdout and exits 0, 2 on a policy refusal or 1 on failure; it is not saved unless --save or --session is given. |
prx go | Zero-config start in a temporary workspace. Writes no config files. --message runs one chat turn like prx chat --message; otherwise it opens chat. |
prx tasks | List or kill live subagent runs on a running gateway (list, kill, send, pools). |
prx cost | Read the workspace cost ledger without initializing configuration. --json for machine-readable output. |
prx config | Inspect and change configuration through a transaction (show, plan, apply, history, rollback, status, schema). |
Binaries
| Binary | Description |
|---|---|
prx | Main AI daemon and CLI — providers, channels, tools, evolution |
prx-node | Lightweight remote node agent — runs on managed devices |
Architecture
Channels Tools Remote Nodes
Signal · WA · TG · ... Shell · MCP · ... macOS · Pi · ...
│ │ │
▼ ▼ ▼
┌─────────────────────────────────────────────────────┐
│ prx daemon │
│ Agent Loop · Gateway · CTE · Xin · Memory · Evo │
└──────────────────────┬──────────────────────────────┘
│
Providers
Anthropic · OpenAI · Google · ...
Documentation
| Topic | Description |
|---|---|
| Installation | Binary install, source build, upgrades, rollback |
| Providers | Provider catalog, fallback chains, token refresh |
| Channels | Messaging platforms, DM/group policies |
| Tools | Built-in tools, hooks system, webhooks |
| Scheduled work | Cron, Xin, and HEARTBEAT.md: what runs, leases, ordering |
| Remote Nodes | prx-node agent, device pairing, JSON-RPC |
| Web Console | Browser-based management interface |
| Skill Lifecycle | New instruction skills, reviewed executable bundles, evidence gates, publication and recovery (opt-in); controlled V01 studies |
| Evolution | Self-improvement pipeline |
| Configuration | Config reference, workspace files, security |
| Router | LLM Router config, flow, safety boundaries |
| WASM Plugins | Plugin developer guide (Rust/Python/JS/Go) |
| Host Function Reference | WASM plugin host API reference |
| Plugin Runtime Lifecycle | Atomic generations, subscriber pumps, trust and hook bounds |
| Release Gate | Exact-SHA CI/security gate that guards release publication |
Links
- Documentation — Full PRX documentation
- Community — OpenPRX community forum
- OpenPRX — Project homepage
Related Projects
| Repository | Description |
|---|---|
| openprx/prx | AI assistant framework (this repo) |
| openprx/prx-memory | Standalone memory MCP server |
| openprx/openpr | Project management platform |
| openprx/openpr-webhook | Webhook receiver for OpenPR |
| openprx/wacli | WhatsApp CLI with JSON-RPC daemon |
Contact
info@sylvode.org — general and partnership enquiries. Report security issues privately as described in SECURITY.md.
Origin & License
Forked from zeroclaw-labs/zeroclaw (MIT / Apache-2.0). "ZeroClaw" is a trademark of ZeroClaw Labs. This project is OpenPRX, an independent fork.
Dual-licensed under MIT and Apache-2.0.