📦 Installation

1. Download APIChat.dmg (3.8 MB).
2. Open the DMG and drag API Chat into Applications.
3. Launch it from Applications or Spotlight.

💬 Your first chat

When you first open API Chat you land on the welcome screen. Two ways to get talking in under a minute:

Option A — cloud provider

1. Pick a provider from the toolbar menu (Anthropic is selected by default).
2. Paste your API key in the inline field — a direct link to each provider's key console is provided.
3. Type and press ⏎. Answers stream in real time with live token & cost accounting.

Option B — 100% local, no key at all

1. Open Settings → Local MLX (⌘,).
2. Click Install Engine (one click — the app bootstraps mlx-lm via uv).
3. Download a model from the built-in library (start with Qwen3 4B, 2.3 GB).
4. Select the MLX Local 🔥 provider and chat — free, offline, private.

🔑 API keys & security

Keys never leave your Mac. You choose the storage backend in Settings → Providers:

Switching backends migrates every stored key automatically. There is no account, no cloud sync, and no telemetry — the only network calls the app makes are the API requests you trigger.

🌐 Providers & models

Eleven providers speak through two wire protocols (OpenAI-compatible SSE and Anthropic's native Messages API), all handled transparently:

Per-conversation controls

• Style: Precise / Balanced / Creative (mapped to temperature where the model accepts it).
• Custom temperature slider (0–2) with automatic detection of models that reject sampling params.
• Occupations: reusable AI personas injected as system prompts.
• Context mode: full history, last-N messages, or single-turn.
• Per-message token counts and cost, plus per-conversation and global totals.

🔥 Local MLX models

API Chat embeds a full local inference stack for Apple Silicon, built on Apple's MLX framework — no Terminal required.

How it works

1. The app installs the mlx-lm engine with one click (bootstrapped via uv, Python pinned automatically).
2. You download models from a curated library of 20 mlx-community builds — from Qwen3 0.6B (0.4 GB) to Llama 3.3 70B (40 GB) — with live progress and resume.
3. A local OpenAI-compatible server starts on demand the moment you send a message, and models load lazily.

Models incompatible with your RAM are flagged before you download. Reasoning models (Qwen3, DeepSeek R1) show their chain of thought live as a quote block.

Use the local server from your own code

While API Chat runs, the local server is a standard OpenAI endpoint:

curl http://127.0.0.1:10240/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "mlx-community/Qwen3-14B-4bit",
       "stream": true,
       "messages": [{"role": "user", "content": "Hello!"}]}'

🛠️ Tools in chat

Enable Settings → Tools and regular conversations gain the full tool belt: the model can search the web, read files and call MCP servers mid-answer, streaming its reasoning around the results. Each call appears as a collapsible row in the transcript.

The tool belt

• Web — web_search, SerpAPI (Google/News/Scholar), Exa neural search, Firecrawl scrape/crawl/map.
• Files — read, list, search, write, edit — scoped to folders you explicitly grant.
• System — shell, AppleScript, Apple Shortcuts.
• MCP — every tool from your connected Model Context Protocol servers.

Try the permission model

Every tool has its own permission. Click through the modes:

🖼️ Vision input

Click the 📎 paperclip and attach images — PNG, JPEG, HEIC, WebP, GIF, TIFF — alongside documents. They preview as thumbnails above the composer and are sent as native multimodal blocks.

• Images are auto-resized to 1568 px (the sweet spot for vision models) and re-encoded as JPEG to keep requests light; small PNGs (screenshots) pass through untouched for crisp text.
• Works with Claude, GPT (incl. the Responses API), Gemini, Grok and every vision-capable provider.
• Images persist in the conversation and are re-sent on follow-up turns, so you can keep asking about them.

🤖 Autonomous agents

Agents run a think → tool → observe loop for multi-step work — streamed live so you watch every word and tool call as it happens. Five ready-made agents ship with the app:

Build your own in Settings → Agents: instructions, model, max steps (1–30), a per-tool permission map, and MCP servers. Agents are exportable/importable as JSON. File tools obey the Allowed Folders you grant via the system picker.

🧠 Memory

API Chat quietly learns durable facts — with your approval. After an exchange, candidate memories appear in a banner; keep or dismiss them.

• Profile memories are always injected (who you are, how you like answers).
• General memories are retrieved semantically (on-device Apple NLEmbedding — nothing leaves the Mac) when relevant.
• Project memories are scoped to a conversation folder.

Say “remember that …” to store something instantly. The whole store is AES-GCM encrypted on disk, searchable, pinnable, exportable, and erasable in one click. Each conversation has an incognito toggle 🧠 to opt out.

🎨 Canvas artifacts

When a model produces something substantial — a code file, an HTML page, a long structured document — it's promoted to the Canvas, a side panel with:

• Full version history with notes, revert and branch-on-edit.
• Manual editing (a real text editor with undo) and AI editing — select a range, describe the change in plain language.
• Live preview for HTML and Markdown, one-click copy or export to file.

⑂ Branching & sources

Hover any message for its action bar:

Clickable sources

Web-grounded answers finish with a row of source chips — Perplexity citations and the URLs surfaced by web tool calls, deduplicated and clickable straight to the page.

⌥␣ Quick Chat panel

Press ⌥Space anywhere in macOS: a floating panel appears above whatever you're doing. Ask, get a streamed answer, then Copy or Paste into app — the reply is inserted right where your cursor was.

• Trigger words: /gpt raw prompt, /p name prompt-library template, /set opens Settings.
• Hotkey, modifier and excluded apps are configurable in Settings → Inline.

⌨️ Keyboard shortcuts

📤 Export & privacy

• Export any conversation as Markdown, PDF, or ShareGPT JSON.
• All data lives in ~/Library/Application Support/APIChat — plain JSON for conversations, encrypted blobs for keys and memory.
• No account. No analytics. No crash reporting. The app's only network traffic is the API calls you make.

❓ FAQ