Chief of Staff Agent - Mira
An AI agent that automates daily operations of a company to increase efficiency.
1. The agent decides what to do
The agent chooses what actions to take based on the input, whether from you or a customer. It takes plain words and decides what it needs to run, and in what order.
2. Nothing sends without confirmation
All external communication (emails, campaigns, replies) and spending requires your confirmation, so the agent can't act on your behalf.
3. It learns with continued usage
The agent is built to update its memory on both the company and the individual daily, so you only have to explain something once.
Click any process to see what Mira can do with it.
What should Mira do?
Pick a task, click a process on the map, or type a request.
CONFIRMS ARE HANDLED BY CODE, NEVER BY THE MODEL
Live demo, invented data
Mira, a chief of staff for a small agency
You run the agency. Mira runs its tools from one chat, and asks you before anything is sent. Take one morning with it: three clicks, about a minute.
What Mira Does
Mira was built for an agency to assist in managing and automating everyday operations. This involved building lead lists, connecting to Instantly through their API, running client scans, generating reports for prospects, creating social media posts, and managing client data. When run manually, each of these business processes involved repeated work and a lot of the team's time.
Once built, Mira was able to handle all of it within one conversation. Anyone on the team asks in plain words, by messages or voice note, and Mira picks the tools, runs them in order and reports back. Anything that would leave the agency stops and waits for confirmation.
How It Is Built
Mira is a Python program built on top of Anthropic's Claude Agent SDK. The Claude models run the conversation and decide which tool to call next. The tools are built as Python functions, grouped into MCP servers that run inside the same program, so each one has a name, a description and typed inputs the model can read and understand.
Some of the smaller jobs (morning briefs, nightly memory updates, etc.) are single model calls with no additional tools. Helper agents with web searching capabilities are used for research and lead investigation.
- Agent
- Claude Agent SDK, Python 3.12
- Models
- Haiku for small jobs; Sonnet, Opus, and Fable for more complex and important tasks
- Runs on
- A Linux server, with jobs scheduled by systemd
What It Connects To
Each outside service is called straight through its API, with its own login. None of them are reached through a third party connector, which keeps every call in one place where it can be logged, capped and tested.
What it reads
- TelegramYour messages and voice notes
- GmailNew mail
- Google CalendarThe day's events
- InstantlyCampaigns and their inbox
- AI enginesAnswers from OpenAI, Gemini, Perplexity and Claude, for the visibility scans
Mira
- Claude Agent SDK, Python
- Haiku, Sonnet, Opus and Fable
- Tools grouped into MCP servers
What it does
- TelegramReplies, reports and confirm buttons
- GmailStages replies for you to confirm
- Google CalendarStages events for you to confirm
- InstantlyBuilds campaigns, which start after you confirm
- Social mediaPosts to LinkedIn, X, Instagram and Reddit after you confirm