ZeroTrace Companion
AI Assistant
Six providers — local or cloud — with full Model Context Protocol support, both as a client and as a server.
Companion's AI assistant is a flexible LLM workspace built around three principles: provider choice, full tool integration, and explicit privacy posture per provider. You pick a model from any of six supported providers — local or cloud — and the assistant can use Companion's built-in tools alongside any external Model Context Protocol (MCP) server you connect.
The assistant also exposes Companion's own AirLeak data as an MCP server, so other agents (Claude Desktop, Cursor, the claude CLI, custom scripts) can read your live capture data with your authorisation.
What's new in this release
- Six providers — Ollama, LM Studio, OpenAI, OpenRouter, Anthropic, Custom. Switch in seconds.
- MCP client — connect to any external Model Context Protocol server (stdio or HTTP transport). The model gets every tool that server exposes.
- MCP server — Companion publishes its AirLeak tools as an MCP endpoint. Any MCP-aware agent can read your live data.
- Streaming responses — token-by-token rendering for cloud and local providers that support it.
- Cost tracking — for cloud providers, Companion estimates token cost per request based on the model's published pricing.
- Model preload — warm a model into memory before chat to skip the cold-start penalty.
- Capability detection — Companion shows whether the active model supports tool calling, vision, and streaming.
Three modes of operation
Run a model locally via Ollama or LM Studio. Nothing leaves your machine. Best for sensitive investigations and air-gapped operation. Free to run; bounded by your hardware.
What the assistant can do
| Capability | Example |
|---|---|
| Conversational Q&A | "What is an OUI lookup?" |
| Explain device data | "Summarise what this AirLeak session captured." |
| Suggest next actions | "I see an unknown device probing for our corporate SSID. What should I check?" |
| Drive Companion via tool calls | "Show me every Apple device in this session." |
| Help write scripts | "Generate a HID script that types my email address." |
| Cross-reference device data | "Has this MAC appeared in my library before?" |
| Use external MCP tools | Once connected, the assistant can call any tool an external MCP server exposes |
What the assistant does not do
- Send your data to any provider you didn't pick. Provider choice is explicit.
- Take destructive actions silently. Reset, wipe, factory-restore — these always require your explicit confirmation.
- Replace expert judgement. The assistant is a copilot. For high-stakes decisions, verify what it tells you before acting.
Quick start
- Choose a provider — see providers for the comparison.
- Set it up — install the runner (local) or paste an API key (cloud).
- Open the assistant drawer in Companion (
Ctrl+Shift+A). - Ask a question.
Local setup: a few minutes. Cloud setup: under a minute (paste API key, save).
Where the assistant lives in the UI
The assistant is always available via the drawer (Ctrl+Shift+A). It opens on the right side of any Companion view without taking you away from what you're doing.
Within the AirLeak workspace, the assistant has additional context-aware variants:
- AirLeak AI drawer — the assistant pre-loads with knowledge of the active session's devices, alerts, and library.
- Per-device assistant — open the assistant from a device's detail view and the assistant knows the device.
MCP — Companion as both client and server
Companion implements both halves of the Model Context Protocol:
| Direction | What it does |
|---|---|
| MCP client | Companion connects to external MCP servers you configure. Their tools merge into the assistant's tool catalog. |
| MCP server | Companion exposes 16 of its own AirLeak tools as an MCP endpoint. Any MCP-aware agent (Claude Desktop, Cursor, claude CLI, custom code) can connect. |
This means you can:
- Add filesystem, browser, or git tools to Companion's assistant via existing MCP servers.
- Hand your live AirLeak capture data to a more capable cloud agent (Claude Desktop, etc.) for higher-quality reasoning while keeping the capture on your local hardware.
See MCP client and MCP server for the full setup.
For complex queries, the assistant chains multiple tool calls in one turn — pull data from one tool, filter with another, cross-reference with an MCP-provided third-party tool. This is where the AI workspace starts to feel genuinely powerful rather than gimmicky.
In this section
- Providers — the six supported providers and how to choose.
- Setup — per-provider setup walkthrough.
- Chat — drawer, history, streaming, cost tracking.
- Tools — Companion's built-in tool catalog.
- MCP client — connect external MCP servers.
- MCP server — expose AirLeak data to external agents.
- Models — recommendations across local and cloud.
- Privacy — the per-provider privacy posture.