A startup inside Midcore
Sixty-three capabilities. Five modes. One chat.
A single operator can build a robot policy, deploy it to a fleet, ingest 47 adverse-event reports, run causality, draft a PSUR, export a capsule — and never leave the conversation. Every action under an approval gate. Every gate written to a ledger.
Five modes
Pick the safety posture before you pick the task.
A mode is a contract between you and the agent. The agent reads it. The gate engine enforces it. The ledger writes it down. You change modes from a single toggle.
Vibe
Interactive. Destructive actions prompt you once each. The right mode when you are exploring an idea and want the agent close.
Auto
Autonomous. The agent shows you the plan, you approve once, and it runs the whole sequence without further prompts. The right mode when you trust the plan and want to step away.
Planner
Read-only. The agent proposes actions but refuses to mutate. The right mode for "what would you do if you owned this codebase?"
Researcher
Read-only except research tools. Lets you explore Research Studio workflows without touching files or fleets.
Designer
Read-only except robotics-designer tools. Iterate on robot configurations without touching live fleets.
Providers
Bring the model you like. Keep the agent you trust.
Anthropic Claude, OpenAI, Google Gemini, xAI Grok, DeepSeek — behind one picker. Every provider runs through the same approval gates, the same ledger, the same DLP layer. Swap models without rewriting your workflows.
Approval gates
Six gates. One ledger. No surprises.
File-system writes, git mutations, network writes, secret access, robot control, agent spawn. For each: Allow once, Allow for session, Always allow, or Deny. Decisions live in an approval ledger. Rules expire after thirty days so consent does not silently fossilise.
Parallel execution
Six tools at once, under hard caps.
Up to six tool calls run concurrently — grep across six directories, six causality analyses, six policy probes. Hard budgets: 200 tool calls per run, 8 MB aggregate results, a per-turn token cap. When the agent runs out of budget, it stops and shows you what it learned.