Agents and runtimes
Last updated 10 September 2026
Behind one conversation, Ghillie can call on Claude, Codex and local models. It decides who does what; you only ever talk to Ghillie.
The runtimes
| Runtime | Role |
|---|---|
| Claude | An agent runtime. Scopes, researches and reasons over a goal and its context. |
| Codex | An agent runtime. Inspects and changes real repositories with normal search, edit, build and test tools. |
| Ollama | A local model host used by Ghillie's own runtime. Extracts private facts and classifies results without anything leaving the Mac. |
All three emit the same event vocabulary into Ghillie's audit trail while keeping their native session identifiers, so a run can be resumed inside the runtime that started it.
Fan-out
One request can fan out across runtimes and come back as one answer: Claude scopes and researches, Codex inspects or changes the code, Ollama extracts anything private locally, and Ghillie verifies, synthesises and records the run. You can turn fan-out off or pick the lead runtime from the Powers screen.
What stays free, what gets structured
Ghillie constrains effects, not thought. The agents receive the real goal, the real workspace and the relevant context. They can investigate surprises and propose changes to the plan.
- Ghillie owns the conversation, the task graph, runtime session IDs and hand-offs.
- Tools are named Powers with typed inputs, granted through expiring leases with call budgets.
- Writes outside the workspace, outbound communication and destructive actions wait for approval.
- Runtime events, tool calls and policy changes enter the audit chain.
Private work
Work you mark private can be forced onto local models. The extraction and classification steps then run on your Mac through Ollama, and no hosted model sees the content.
Named agents
Ghillie keeps a roster of named agent roles for different kinds of work, each with its own portrait in the Agents pane. Cards on a project board carry the runtime logo of whichever agent last worked them.