AICoven: Local AI Chat

AICoven Local now mirrors the full AICoven layout, and your on-device models are front and center. • Redesigned workspace: a Discord-style icon rail switches between Strix and your covens, no more sidebar dropdown. • Knowledge Hub and per-coven Settings are now first-class tabs in the sidebar. • New split Settings workspace surfaces Provider Keys, Budgets & Usage, Connected Apps, and MCP Servers at a glance. • On-Device Models: manage and run local Apple Silicon models from a dedicated Local Agent settings section. Under the hood (recent groundwork) New self-hosted/provider clients: OpenClaw (OpenAI-compatible proxy) and Nous Hermes (Together AI or local LM Studio/llama.cpp). Reworked chat turn architecture and background actors for memory/summarization. Local security hardening (phase 1) for at-rest encryption. MCP server integration and Connected Apps. • An Activity bell replaces the old profile menu, with a badge for pending proposals and updates. Everything stays on your device — bring your own keys, or run fully offline with on-device models.

AICoven Local: Advanced Tool Use & Memory Integration This update is about making your local agents more reliable and your context more persistent. We’ve overhauled how AICoven handles background tasks and tool execution to ensure you aren't fighting the model for the right answer. Smart Memory Proposals Thread summaries are no longer hidden in your metadata. AICoven now automatically generates Memory Proposals from your conversations. 3-6 bullet points summarizing key decisions and facts from your threads. Review them in the Memory Proposals view and explicitly save them to your permanent context. Your background summarization now strictly honors your active Provider and Model settings—no more heuristic "guessing" for your thread history. Tool Pre-Execution for Local Models Small local models (MLX/Ollama) often struggle with generating complex JSON tool calls. We’ve fixed this with a new ToolChainResolver. If you ask for a web search or a file, AICoven executes the tool before the model even starts its response. By injecting the raw results directly into the context, we ensure your local models provide grounded, perfectly formatted answers without JSON errors. Performance & Reliability Zero-Latency Intent Gate: AICoven now instantly detects purely conversational queries (like "what do you think?" or "summarize this"). This bypasses the tool-resolution loop entirely, saving you token latency. Hallucination Guards: If you ask about your emails or calendar without an active MCP connection, the system now intercepts the query. Instead of the model hallucinating details, it will correctly admit it lacks access. Infinite Loop Detection: We’ve added canonical signature tracking to prevent "agent thrashing." If a model tries to call the same tool with the same arguments 3 times, the engine breaks the loop and forces a new strategy. The reality is: Context is a budget, and this update helps you spend it better. Questions or feedback? We’re comparing notes at support@aicoven.ai.

The agents are finally leaving the sandbox. Giving Agents Hands (MCP) You can now plug in MCP servers—like Zapier, Make or your own custom endpoints—to give your agents actual tools. They can now touch your email, Slack, or calendar instead of just talking about them. Discovery is automatic so you don't have to map every tool by hand. Local Models (MLX) Run models like Llama 3.2 or Gemma locally on your device with zero cloud dependency. We also added a filter that hides models your device can't actually handle. iOS Stability & Memory The iPhone is a bit of a stickler for memory usage, so we’ve implemented a more aggressive management system. We’re now proactively unloading models after every response to keep the app fast and responsive. It’s significantly more stable, even when you're pushing the hardware to its limit. Smarter Tool Selection The AI is now much more discerning. It won’t try to trigger a complex Zapier workflow just to tell you what time it is (built-in tools now take priority). We’ve also added support for SSE streaming for MCP servers, because waiting for a full response before seeing a single word is just cruel. Quality of Life Unified Settings: One place to manage all your providers and models. Markdown Fixes: Text selection on macOS finally behaves like it should. Mobile Keyboard: Fixed the issues where the keyboard would hide the text you were trying to type. If things go rogue or you hit a bug, we're comparing notes at support@aicoven.ai.

Smoother conversations — Fixed an issue where the app could freeze while your AI was responding. Chats now stream smoothly without hiccups. Provider Keys actually work now — The "Test" and "Edit" buttons on your API keys now do what they should. Test checks if your key is valid, and Edit lets you update it without deleting and re-adding. Better backgrounds everywhere — The signature nebula background now fills the entire screen properly on all tabs, including Covens and Profile. No more odd cutoffs at the top. Calmer default look — The animated background is now static by default for a cleaner experience. You can turn the animation back on in Settings if you miss it.