Fewer tokens, same answers: built-in context compression, right-sized execution, and process reuse keep cost proportional to the work.
Context compression (headroom)
Conversation context is compressed before it reaches the model through the built-in headroom proxy — up to 95% fewer tokens, same answers. Compression is on by default and can be toggled in settings.
Right-sized execution
Model and reasoning effort are selected per task: a routine question doesn't pay for a deep-architecture configuration. Selection is automatic, with per-message override when the human wants to force a level.
Parallel workers
How many agent workers run at the same time is a single setting (default 2, up to 16). One worker is one live process for one active thread — whatever the runtime: Claude, Codex, or local models all count against the same pool, and the runtime priority order only decides which one is used, not how many run. When the pool is full, the least-recently-used idle worker is recycled; its warm context is rebuilt on the next message, which costs latency and tokens.
Sizing rule: match the setting to the number of conversations you expect to be active at once — a team of several companions working in parallel deserves more than the default — within what your machine's RAM and CPU allow.
Fast follow-ups
Within a thread, the agent process is reused between messages — in-flight context stays warm, so follow-up turns start immediately instead of paying a fresh start every time. Long conversations survive context-window compaction transparently, with key findings preserved.
Visibility
Live token usage and rate-limit state are shown in the app's status bar, per provider — consumption is observable, not a surprise on the invoice.