Run local AI, agents, knowledge, tools, channels, and automation from one macOS workstation.
Ai Keeper is a native macOS control surface for Apple Silicon local models and agent workflows. It can run local serving engines, expose OpenAI-compatible APIs, coordinate multi-agent workspaces, index private knowledge, connect messaging channels, and manage advanced security and system diagnostics.
Use this manual
Tip: hit ⌘K from any page to search across the manual. Click the Glossary button bottom-right to look up a term without leaving the page.
What Ai Keeper can do
Local model serving
Run AI models on your own Mac instead of paying a cloud company. There are 13 "engines" to pick from — but you do not have to choose. Leave the engine on Automatic and the app picks the right one for the model you downloaded.
Turns local model assets into running instances. Each instance bundles engine, model path, port, context budget, output limit, sampling profile, tool policy, and access policy.
- Recommended defaults:
omlxfor MLX models — SSD-tiered KV cache, multi-model serving, native tool calling, native MTP speculative decoding, continuous batching, OpenAI + Anthropic APIs.llama.cppfor GGUF — Jinja templates, embeddings/rerank, prompt cache. - Specialists:
mlx-serve(native Zig, no Python),mlx-optiq(mixed-precision KV quant + MTP + LoRA),ds4(DeepSeek V4 / GLM 5.2 GGUF, ~96 GB+ RAM),mtplx(multi-token prediction),vmlx(adaptive quantization, TTS/STT),llama.cpp-turboquant(rotated KV cache, Gemma 4 MTP),llama-diffusion(diffusion-LM GGUF). - Fallbacks and escape hatches:
vllm-mlx,mlx-flash(low-RAM weight streaming),mlx-vlm(vision/audio/OCR/video),rapid-mlx(broad format compatibility). - Not a local engine:
external-openaipoints Ai Keeper at any existing OpenAI-compatible server — Ollama, LM Studio, or a remote box — by base URL.
Unified API proxy
A single web address other apps (VS Code, Cursor, your scripts) can use to talk to your AI. Behind the scenes, Ai Keeper picks the right model.
Single OpenAI/Anthropic/Ollama-compatible endpoint that fans out to local instances or routed providers. System > API Access shows the base URL, ready models, management URL, and copy-paste examples.
Agents and workspaces
Build AI helpers with their own job, tools, and personality. Group several into a "team" that works together.
Agents are named AI roles with their own model, tool access, skills, memory scope, standing orders, hooks, and personality. Workspaces group agents for repeatable runs — sequential when order matters, mirrored parallel when independent specialists work concurrently.
Knowledge and memory
Give the AI access to your files, notes, and saved facts about you — so it actually knows your work and remembers what you told it.
The current Knowledge hub has 10 tabs: AI Context, Skills, Memory, Audit, Documents, Wiki, Ingestion, Dreaming, Contacts, and Obsidian. Wiki tracks timeless facts, dated snapshots, current-truth pointers, revisions, and freshness health. Ingestion applies reviewed reconciliation proposals; Obsidian supports watched reconciliation and optional conflict-checked write-back.
Automation and channels
Schedule the AI to run jobs by itself, and connect it to messaging apps so people can chat with it on Slack, Telegram, iMessage, etc.
Scheduled jobs, triggers, workflows, heartbeat turns, background tasks, hooks, standing orders, webhooks, and shared-inbox channel connectors. 23 channel families: Telegram, Slack, Discord, Twitch, IRC, Mattermost, WhatsApp, Matrix, WebChat, iMessage, Teams, Signal, and more.
Security and operator tools
Tools for keeping things safe and figuring out what went wrong — logs, audit trails, password vault, backups, and lots of advanced controls.
Request logs, inspectors, runtime/crash logs, diagnostics, sessions, usage trends, audit trails, failover chains, remote access, security scans, secrets vault, backups, tool groups, server lanes, ACP, node mesh, device pairing, media capture, exec sessions, personality editing.
First-run operating order
Follow this order on a fresh install. Each step assumes the previous one succeeded — if something fails, fix that layer before moving on.
- Open Engines and run Fix Required. This installs the serving engines and supporting tools the app needs. Nothing else works until this finishes.
- Open System > Health. Run Diagnostics and Doctor. Fix anything red before changing runtime settings.
- Pick Server or Client mode. Server mode runs models on this Mac. Client mode connects to another Ai Keeper server. If you are not sure, you want Server.
- Set the model directory. System > Settings > Storage. Choose a disk with room for the downloads — models are several GB each.
- Get a model. Download > Cookbook is the shortcut: turn on Fits my machine, pick a recommendation, click Download & Set Up. Use Download > HuggingFace only when you want something specific.
- Create a Runtime instance. Leave the engine on Automatic and run Optimize before touching anything manually.
- Test in Chat. Confirm streaming, attachments, tool approvals, slash commands, and the context meter.
- Expose the API only when you need it. Enable the proxy and copy the base URL from System > API Access.
- Add knowledge. Turn on memory or index documents after the plain chat path works.
- Create agents and automation. Add multi-agent workspaces, scheduled jobs, channels, and webhooks last — after the model and tools are reliable.
Each of these steps has a full click-by-click walkthrough in the step-by-step guides. If this is your first hour with the app, start with First 30 minutes.
Quick concept dictionary
Just the most-used terms. For the full glossary, open the bottom-right Glossary button or visit the Glossary page.
The "motor" that runs an AI model.
Backend process family. Examples: omlx, llama.cpp, mlx-serve, mlx-optiq, vmlx. The former direct mlx-lm serving backend is retired; persisted instances fall back to omlx.
One running copy of a model with its settings.
Runnable configuration: backend + port + context + output limits + sampling + tools + access + custom args.
A single address other apps point at to talk to your AI.
Unified API layer. OpenAI-compatible clients call local instances or routed providers through one base URL.
A word-piece. Models read and write in tokens, not whole words. Roughly 1 token ≈ 0.75 English words.
Smallest unit consumed and produced by the model. Used to measure context, output, and cost.
How much the AI can "see" at once. Bigger window = better memory but more RAM.
Total token budget per request. Consumed by system prompt, history, tools, retrieved documents, and output.
The AI searches your files for relevant info, then answers using that info.
Retrieval-Augmented Generation. Indexes local files; relevant chunks are inserted into the prompt at answer time.
A standard for AI agents to borrow tools from external programs.
Model Context Protocol. Ai Keeper can consume MCP tools/resources and expose its own tools to other agents.
An auto-action that fires when something happens — like "log every Slack message".
Event-triggered action. Can inject prompts, execute skills, notify webhooks, log memory, or run commands.
Source areas: AppShellNavigation, SettingsView, InstanceDetailSections, ModelTypes, CloudProvider, channel connector services, RAGService, MemoryView, AgentsView, AutomationView, BrowserView, ClawHubView, HuggingFaceDownloadView, MCPRegistryBrowser, slash command definitions.