Proxima v5.0.0

[Zen4-bit]

What's New

• Bring Your Own Key (BYOK) Mode — Direct API key support for ChatGPT, Claude, Gemini, Perplexity, DeepSeek, and any OpenAI-compatible endpoint. Features native OS keychain encryption (SafeStorage), multi-model selection, and connection testing.
• Proxima Agent Autonomous System — A standalone local Python platform designed to automate any task, browser automation via Chrome DevTools Protocol (CDP), execute code in sandboxed workers, run desktop commands, and manage local environments.
• In-Process MCP Middleware Pipeline — Re-architected the MCP server to use the official SDK and Zod schema validations. Includes task-aware smart routing (EMA latency weighting), ONNX semantic memory vector lookup, Context7 live documentation helper, along with upgrades and optimizations across all core tools.
• Interactive Terminal CLI Shell — The proxima-agent terminal console allows running, monitoring, switching models, and checking logs directly.
• Token & Cost Auditor — SQLite-backed token and transaction cost auditor keeping a local ledger of inputs, outputs, and financial metrics across all models.
• New Orchestration Tools — run_workflow (sequential AI chaining), run_loop (iterative quality refining), and crew (role-based multi-agent workflows).

Bug Fixes & Optimizations

• Removed slow DOM scraping with session-API engines for unified provider response streams.
• Decoupled Electron main process into modular Settings, Core, and CLI IPC channels.
• Implemented strict SSRF protections on web scrapers (blocking private/loopback/link-local IP ranges).
• Added loopback-only CORS origin protections on local FastAPI and Electron servers.
• Gated the MCP/IPC TCP socket channel with dynamic, cryptographically generated token validation (ipc-token.json).
• Pinned native environment runtime verification (monitoring Chrome, Tesseract OCR, and window tools).
• Packaged offline Python virtual environment setup using pre-bundled local wheels.
• Consolidated search and content retrieval tools under unified interfaces.

Download

• Windows: Proxima-Setup-5.0.0.exe

Requires Windows 10/11 or macOS/Linux, Node.js >= 18, and Python >= 3.10.

---

This discussion was created from the release Proxima v5.0.0.

[richardchen874-sys]

Proxima's BYOK mode plus the SQLite token cost auditor is the right boundary to inspect, because local agents still need provider-level usage evidence.

I would define a small provider capability matrix and attach usage/latency/error metadata to every routed call. That gives you a clean path to cost-aware routing without hiding why a route was chosen.

This overlaps with something I am testing: an OpenAI-compatible multi-model API layer around official Chinese models, especially around cost-aware routing and normalized usage behavior. For Proxima, would you rather optimize routing first for cost, latency, or fewer failed agent steps?