Find the buried thing
Ask it to find the message, contact, task, or old decision you vaguely remember.
A side-benefit of the architecture
The same source-grounded context, in a chat box.
The Assistant runs against the same MCP tool catalog that external agents see — get_context_pack, get_thread_context, get_person_context, search, drafts, mutations. Every answer it gives links to the message it came from, because the underlying data layer makes ungrounded answers structurally hard.
AI assistant + tools
Ask for help, then let the tools do the legwork.
Not just a chat box — and not a blank one. The assistant works over your real messages across every channel you own, on your own machine: it searches, drafts, inspects approved systems, uses connectors, prepares actions, and stops for approval when the request crosses a line. Every answer links back to the message it came from.
Searches
Drafts
Uses tools
Asks before acting
Draft the reply
Give it the thread and a goal; it prepares a response you can edit and approve.
Use approved tools
Let it check connected systems, files, or MCP tools instead of guessing.
Pause at the line
When the action matters, it asks. You see the request before anything goes out.
Explain what broke
If setup or sending fails, it can translate technical errors into next steps.
Build a workspace
Ask for a tracker or workflow and let Vibes create the first version.
Connected tools stay visible.
Each connector shows what tools it exposes, so the assistant's reach is understandable.
Model choices are part of the controls.
Private, cheap, and high-quality jobs can be routed differently instead of using one setting for everything.
Technospeak, translated
Powerful tools are useful only when you can see the leash.
The assistant can use specific approved tools, like search, file access, connectors, or app actions.
Risky actions can be limited by paths, permissions, and approval gates.
Different jobs can use different AI models, balancing cost, privacy, and quality.