Concepts
AI Runtime

AI Runtime

How an agent's call actually reaches a model — and what's between you and the inference.

The adapter layer

Every AI call in Cabinet — a task, a routine, a heartbeat, a chat message — flows through a provider adapter. The adapter knows how to talk to one specific provider. Cabinet itself doesn't know which model it's calling, just that it asked an adapter to run a prompt and got a transcript back.

agent persona / task brief
        ↓
provider adapter  ──→  Claude Code CLI / Codex CLI / (future: API / local)
        ↓
persisted conversation + live transcript view

This split matters because it means Cabinet can support new providers without changing how agents are written. A new adapter ships, the existing personas just gain a new option in their model: field.

What ships today

Two adapters are in default builds:

  • claude_local — talks to the Claude Code CLI on your machine.
  • codex_local — talks to the Codex CLI on your machine.

Both are local CLI providers — you authenticate the CLI once with your own account, and Cabinet uses it as a subprocess. Your prompts don't go through a Cabinet-hosted backend. There's no Cabinet API key.

See Bring your own AI for the full provider list and what's coming next.

Conversations and transcripts

Every adapter run is persisted as a conversation — a stream of messages with timestamps, tool calls, and outputs. While a run is in progress, you watch the transcript live in the Approval Queue or in the agent's workspace. After it finishes, the conversation lands in the cabinet folder so you can scroll, search, or git log it.

This is the difference between "the agent ran" and "the agent did this, you can read every step".

Per-run overrides

A persona declares a default model. A routine, a task, or even a single dispatch can override it.

# In persona.md — the default
---
name: GTM Lead
model: claude-opus-4-7
---

# In a routine — Friday wraps go on Opus, daily pulses on Haiku
heartbeats:
  - cron: "0 9 * * 1-5"
    model: claude-haiku-4-5
    prompt: "Inbox triage."
  - cron: "0 17 * * 5"
    model: claude-opus-4-7
    prompt: "Friday wrap."

You can also pick model and reasoning effort right at the composer when you launch a one-off task:

┌─────────────────────────────────────────┐
│ → GTM Lead   ▾  gpt-4.1   ▾             │
│ ◯ low  ● medium  ◯ high                 │
└─────────────────────────────────────────┘

The terminal still exists

Cabinet ships with a web terminal — an interactive local AI CLI inside the browser. It's not the only way agents run anymore (the adapter layer is), but it stays as a first-class surface for:

  • Direct CLI sessions when you want to drive the agent yourself.
  • Debugging adapter behavior.
  • Future terminal-native features (Cabinet-managed tmux-style workspaces).

You can ignore it if you're working entirely through the agent dispatch UI.

Read on