Opinionated field notes on agent architecture, BYOK orchestration, local LLMs, and prompt economics — from building AICoven on iPhone, iPad, and Mac.
My agent opened feature branches in repos that had nothing to do with the feature. With several clients, that's one mistake in three places: whose AI account pays, whose apps get touched, and whose repos get written to. Here's how one coven per client keeps them apart.
The app has been redone around eight plain jobs, and each coven now has an Asset Library: the files, datasets, and images your agents and you make, reviewed, approved, and sent where they need to go.
AICoven can now connect to Azure OpenAI with Microsoft Entra ID instead of an API key, find your subscriptions, resources and deployments for you, and, if you tell it the deployment type, estimate what the connection costs.
AICoven now connects to any server that speaks the OpenAI chat API, on your own hardware or hosted, and finds the common local ones on your Mac. OpenRouter gets a row of its own.
Connect Notion and a coven's agents can search and read the pages you allow, and you can save a page into the coven's memory. Comments that mention you in Notion, Google Drive, and GitHub now wait in On me, where you decide whether they're yours.
Sign in with Figma, or use a personal access token to reach projects too. Choose the files and projects each coven can see, and let agents read them, look at frames, and work through the comments, without wandering into the rest of your Figma.
Connect Slack and your covens can read the channels you choose, answer when you ask them in Slack, draft messages you approve word for word, and turn a mention into a task.
The AICoven Helper lets the Claude you already pay for answer in your covens, with no API key. Claude signs in with Anthropic on your Mac, and AICoven never sees that sign-in.
AICoven gathers the issues assigned to you and the pull requests waiting on your review into a coven's Tasks. Tick one off and, if you allow it, GitHub or Linear hears about it too.
Hermes and OpenClaw are agent runtimes you host. AICoven routes cloud APIs, native runtime controls, compatibility HTTP, and Mac-local tools without pretending they are one pipe.
Anthropic's advanced tool use research showed us the path. We adapted it for every provider — and for local models that can't even call tools natively.
Dumping 200 MCP tool schemas into a local model's context is a great way to ensure it never answers your question. Here is how we fixed the 'amnesiac expert' problem.
AI doesn't return clean JSON — it streams chaotic fragments of text, tool calls, and errors in real time. Here's what it took to render that smoothly, and why it still breaks.
When you build a local-first AI app, you hit a brutal hardware constraint almost immediately: the context window. Every token is precious. Here's how we're managing it.
Building native cloud APIs is easy. Getting a local Llama-3 model to execute shell commands requires the dark arts of prompt engineering and a lot of regex.
Everyone wants an infinite context window. The math says no. Here's how we compress history to keep the system prompt from getting lost in the middle.
We love JSON mode because it makes agents reliable. But forcing a model to think in JSON burns 3x more tokens than letting it reason in plain text.
If you aren't using your own API key, you're trusting a middleman with your codebase. Here's how the big three handle your data differently depending on how you access them.
