General

General is where your workspace-wide defaults and safety posture live: whether agents may control your computer, which model they reach for, how long they’re allowed to keep working, which service vendors they prefer, and the privacy toggles. It also holds the gentle “re-show me the intro” buttons and the destructive Reset workspace data.

askTheodor is local-first: your conversations, memories, files, and credentials stay on this machine. Nothing is uploaded anywhere unless you connect a cloud LLM provider — and even then, only the prompts go out, nothing else.

🖱 Allow computer control (POWERFUL)

The leading, most powerful setting. When enabled, agents can take screenshots of your desktop, move the mouse, click anywhere on screen, type into any focused window, and press keys (including cmd+Q, cmd+W, and other system shortcuts). OFF by default — you must consciously turn it on, and turning it on asks you to confirm.

Its safety story:

• Approval-gated — every destructive action requires your approval by default.
• Kill-corner — slam your physical mouse into the top-left pixel of the screen to abort agent actions for 5 seconds.
• Audit — every action is logged in the Activity tab alongside the screenshot the agent saw before acting.
• macOS — the first agent action prompts for Accessibility and Screen Recording permission; both are required, granted in System Settings.

This hands an agent your mouse and keyboard

Computer control is the most powerful — and most dangerous — surface in the app. Leave it off unless you specifically need an agent to operate other applications, keep it approval-gated, and remember the kill-corner. When in doubt, prefer the worker’s normal tools (browser, files) over full desktop control.

Workspace default LLM + fallback chain

Sets the model your workers use when they don’t specify their own, plus an ordered fallback chain — if the first model is unavailable or errors, askTheodor tries the next. Companies inherit this default and can override it. This is the knob new users hit first; set it to whatever you connected in Settings → Providers.

🔁 Max agent iterations

How many tool-call rounds an agent may take inside a single run before the loop stops with a “max iterations reached” warning.

• Default: 12. Range: 1–100.
• The agent’s accumulated context is preserved when it stops — to resume, send any follow-up (e.g. “continue”) in the same conversation and it picks up where it left off.
• Raise it for multi-step research that chains many fetch / browser calls; lower it to bound spend on chatty models.

The value saves when you click out of the field.

Preferred service providers

Workspace-wide vendor defaults for non-LLM services — domains, hosting, payments, and the like — that your Companies inherit and can override. (This is not about LLM providers; that’s the default-LLM section above.) It steers which vendor a worker reaches for when a task calls for one.

Send anonymous usage metrics

A privacy toggle, OFF by default. When on, it would share anonymous, aggregate usage — app version and OS, simple counts (how many workers / routines / companies), which features get clicked, and crash reports without payloads. It never contains your prompts, files, credentials, persona names, memory, or approval contents.

Plumbing only, for now

No telemetry endpoint exists yet — this toggle just records your preference so it can be honored if and when telemetry is ever added. Nothing is being sent today regardless of the setting.

✨ AI Field Coach

ON by default. Adds a small ✨ pill to text fields across the app (worker prompts, missions, goal and routine descriptions). Click it for AI-drafted starter text, a refinement of what you wrote, hand-written examples, or a plain-English explanation of what the field is for. It defaults to your local model (Ollama / built-in) and only routes to a cloud provider if you’ve set one as preferred. Every call is audited and rate-limited to 30 per hour to prevent runaway loops. Turn it off to hide every ✨ pill if you’d rather have no AI assistance.

Re-show the intro

Two non-destructive helpers that change nothing else about your data:

• 🎩 Re-open Theodor’s welcome — re-shows the first-launch greeting where Theodor introduces himself and asks which model to run on. Handy if you dismissed it earlier to go add an API key.
• 🔁 Re-run the onboarding wizard — clears the onboarding-completed flag so the 4-step setup wizard fires again on next launch. Your workers, conversations, Library, and audit log all stay put.

🔥 Reset workspace data (DESTRUCTIVE)

Soft-deletes every Company, worker, project, goal, task, Library item, and conversation — useful for testing a fresh start without uninstalling. To run it, you’ll be asked to type RESET. Items remain restorable from Settings → Trash for 30 days, then auto-purge.

This one is recoverable

Unlike the Factory reset in Settings → Demos (which deletes for good), Reset workspace data here is reversible — everything lands in Trash for 30 days. If you reset by mistake, open Trash and restore.

In practice

Leave computer control off unless you have a specific job for it. It’s the most powerful setting in the app — an agent operating your actual mouse and keyboard — and most work doesn’t need it. When you do enable it, learn the kill-corner first (mouse to the top-left pixel aborts immediately) and watch the first few runs rather than walking away.

Max iterations is a runaway brake, not a quality dial. Raising it doesn’t make a worker smarter; it lets a confused one loop longer at your expense. If a worker keeps hitting the cap, the brief is usually ambiguous or its toolset is wrong — investigate that before raising the ceiling.

A fallback chain earns its keep on unattended work. For anything running while you’re asleep, a second model behind the first turns “the provider had a blip and nothing ran” into a slightly different result. For interactive chat it matters less — you’re there to retry.

The Activity log is the answer to “why did it do that?” Every agent action is recorded, including the screenshot an agent saw before each computer-control step. When output surprises you, read the log before re-writing prompts — the cause is usually visible there.

Terms in this page

• Local-first — your data lives on your own machine by default; nothing leaves it unless you connect a cloud provider, and then only the prompts do.
• Computer control / computer-use — letting an agent take screenshots and operate your mouse, keyboard, and screen. Powerful, off by default, approval-gated.
• Approval-gated — an action that pauses for your explicit OK before it runs.
• Kill-corner — moving your mouse to the screen’s top-left pixel to instantly abort agent actions.
• Audit / Activity tab — the log of what agents did, including the screenshot an agent saw before each computer-control action.
• LLM — a large language model, the “brain” a worker uses to think.
• Default LLM — the model workers use when they don’t pick their own.
• Fallback chain — an ordered list of models to try in turn if the preferred one is unavailable or errors.
• Agent iteration — one round of the agent calling a tool and reading the result; max iterations caps how many rounds a single run may take.
• Run / context — one execution of a worker and the accumulated history it builds up; the context is kept when a run hits its iteration cap, so you can continue.
• Preferred service providers — your default non-LLM vendors (domains, hosting, payments) that Companies inherit.
• Telemetry / usage metrics — anonymous, aggregate data about how the app is used; off by default, and never your content.
• AI Field Coach — the optional ✨ helper on text fields that drafts, refines, or explains what to write.
• Rate limit — a cap on how often something may run (here, 30 Field Coach calls per hour) to prevent runaway loops.
• Onboarding wizard — the short first-run setup flow; this tab can make it appear again.
• Soft-delete vs factory reset — Reset workspace data moves items to Trash for 30 days (recoverable); the Factory reset in Demos removes them permanently.