How Krater's Agent, Auto Router, Memory, and Projects Work Together

Krater's four core systems form a single loop: Projects hold the context, Memory carries your preferences across sessions, Auto Router picks the right model for each step, and the agent executes multi-step work on your behalf. The result is that you set up a task once — goal, files, style — and subsequent work in that workspace doesn't require re-explaining yourself or manually switching between models. This suits anyone doing recurring or multi-step knowledge work (research, content, visuals, scheduled tasks) more than someone who just wants a single chat window.

The four pieces and what each one does

How they chain together in practice

The interaction is sequential, and each stage feeds the next:

1. You define the work in a Project. Drop in the source material — the site notes you can add PDFs, URLs, or videos — so the workspace has something to reason over.
2. Memory supplies your standing preferences. Because it learns how you work, the agent doesn't start from a blank style each time.
3. You give the agent a goal, not a prompt. Per the site, you "set a goal" and Krater "plans it, researches, browses" — the planning step is the agent's, not yours.
4. Auto Router assigns models per step. Rather than you choosing one model for the whole job, routing happens at the task level, so a research step and an image step can use different models.
5. Output lands as an editable artifact. Describe a visual, slide deck, or website once and you get an editable result with a live preview, which you can then revise inside the same workspace.

The compounding effect is the point: steps 1 and 2 are one-time setup, so steps 3–5 run without you re-establishing context.

A concrete example

Say you maintain a recurring competitor digest. In a Project, you add the relevant URLs and any PDFs. Memory already knows you want a short bulleted format with sources cited. You tell the agent the goal — "summarize this week's changes across these sources." The agent plans the sweep, browses for current information, and Auto Router sends the retrieval and summarization steps to appropriate models. The digest comes back as an editable artifact. Next week you repeat only the goal; the Project and Memory are still there.

Where the surrounding features fit

These systems don't operate alone — the site lists adjacent capabilities that plug into the same loop:

• Deep research stays connected to the web, with findings saved into Keep, so research output persists rather than disappearing into a chat log.
• Tasks and 1,400+ app connections let Krater run work on a schedule, which is where the agent moves from on-demand to recurring.
• Speech, sound, and voice cover text-to-speech, sound effects, voice changing, and transcription for audio-oriented steps.

What to check before relying on it

• Model access is bundled, not individually subscribed. The site states ChatGPT, Claude, Gemini, and 500+ models come under one subscription, so the value case rests on using several of them rather than one.
• Pricing and plan tiers aren't specified in the material available here. Check the pricing page directly before assuming what a subscription includes.
• The agent's autonomy has boundaries. It plans, researches, and browses toward a goal, but the output is presented as an editable artifact — expect to review and revise rather than publish untouched.

If your work is repetitive and context-heavy, the Project + Memory + Auto Router + agent combination is designed to collapse the setup overhead. If you only need occasional one-off answers, the multi-system structure adds more than it saves.