There's a class of work that eats entire mornings: research a company, write up a summary, send it to someone, book a meeting to discuss it. Four steps, four different apps, maybe 2–3 hours of context-switching.
We wanted to see if Companion could do all of it from a single prompt.
The prompt was:
Analyze the AI assistant Companion from aioscompanion.com. Create a Google Docs report, email it to [recipient] with the app download link, and schedule a meeting tomorrow at 10 a.m.
Here's what happened.
What Companion did
Companion broke the prompt into tasks and executed them sequentially — no additional input from us:
- Opened the browser and analyzed the entire aioscompanion.com website
- Created a comprehensive Google Docs report with structured insights, feature breakdowns, and recommendations
- Composed and sent an email to the specified recipient with the report attached and the download link included
- Scheduled a calendar meeting for the next day at 10:00 a.m., with the recipient properly invited
Each step used the relevant MCP tool. The browser for research, Google Docs MCP for document creation, Gmail MCP for sending, Google Calendar MCP for booking.
The whole sequence ran automatically. We watched it happen.
Verifying the output
After it finished, we checked each result:
The email — professional subject line, well-written body, correct attachments. Exactly what you'd write yourself after 20 minutes of drafting.
The Google Docs report — a comprehensive analysis of the product with formatted sections, key feature highlights, and actionable recommendations. Clean enough to send to a client without editing.
The calendar meeting — scheduled correctly for 10:00 a.m. the following day, with the title "Discussion about Companion AI Assistant Report," the recipient properly invited, and a Google Meet link attached.
Everything was right. Nothing needed to be fixed.
What it cost
This is the part that surprised us.
The entire automation — website research, document creation, email composition and delivery, meeting scheduling — used 79,000 tokens and cost $0.0067.
Less than a cent. For work that would have taken hours.
This is possible because of two things:
BYOK via OpenRouter. Companion uses your own API key routed through OpenRouter. You're paying model rates directly — no platform markup, no hidden subscription fees. For this task, we used a mid-tier model well-suited for tool use.
Janitor. Companion's context compression system removes stale tool call history as the agent works. Without it, a multi-step task like this one would accumulate 40,000–60,000 tokens of tool call history alone. With Janitor, most of that gets pruned before the next LLM call. It's why complex tasks stay cheap.
The economics of this
Let's put it another way. A knowledge worker spending 3 hours on this task costs a company roughly $75–150 in labor. Companion did it for $0.0067.
That's not a marginal improvement. That's a different category of cost entirely.
The relevant comparison isn't "AI vs. other AI tools." It's "AI vs. the time you were going to spend on this anyway."
How to replicate this
You'll need Companion installed with three MCP servers active:
- Browser MCP — for website research
- Gmail MCP — for sending emails
- Google Calendar MCP — for scheduling meetings
- Google Drive / Docs MCP — for document creation
Once those are set up, the prompt above works as written. You can adapt it for any research target, any recipient, any meeting time.
The structure that makes it work:
- A research target (URL, company name, document)
- An output format (report, summary, bullet list)
- A delivery action (email to X with Y)
- A follow-up action (schedule meeting at Z)
Companion handles the sequencing. You handle the prompt.
One prompt, multiple apps
What makes this different from just using ChatGPT or Claude is the tool layer. A standalone LLM can write a report. It can't send an email or create a calendar event — because it has no connection to those systems.
Companion does, through MCP. And because those connections are active simultaneously, a single prompt can cross app boundaries that would otherwise require manual handoffs.
That's what "agentic" actually means in practice: not a smarter chatbot, but an assistant that acts across your real tools.
Download Companion and try the prompt yourself. The whole setup takes about 5 minutes.