AI Ops Agent

A private operations control center for notes, tasks, calendar context, and scheduled briefings. Run it with any MCP-capable agent runtime, keep the data in your own vault, and watch the loop from a simple dashboard.

Architecture · Workflows · Tool catalog · Runbook

At a glance

This repo is the reusable core of a self-hosted operations agent. It gives an agent runtime deterministic tools for reading a private vault, managing tasks, building scheduled briefs, archiving voice notes, and exposing operational state through a dashboard.

Related MCP server: Flint Note

Common workflows

Repository map

Why

Most "AI assistants" are a chat box that forgets everything and does nothing when you stop typing. This is the opposite: an agent with an operating loop. It wakes on a schedule, keeps durable memory in a markdown vault and SQLite, does real work through typed tools, and only ever writes inside safe boundaries. You run it, you own the data, and the dashboard shows you exactly what it is doing.

The hard part was never "call an LLM." It was the operating system around it: what the agent should know before it speaks, which actions are deterministic tools instead of model guesses, and how state survives across days. This repo is that operating system, generalised so you can point it at your own vault, models, and schedule.

What it does

24 MCP tools in total: vault (vault_read, vault_append, vault_tree, semantic_search), tasks and state (task_add, task_close, task_list, tasks_render_md, mood_log, note_quick, activity_log, activity_update, workout_archive, voicenote_archive), calendar (calendar_list, calendar_add, calendar_delete), multimodal (vision_analyze, video_analyze, ocr_extract, image_generate), and data (query_db, web_search, web_fetch).

How it works

Three layers you change independently: the tool server (this repo), the agent runtime that drives it (any MCP-capable brain), and the schedule that wakes it.

flowchart TD
    Cron["Scheduler: morning brief, evening digest, weekly review, sweeps"] --> Runtime["Agent runtime (any MCP brain)"]
    Chat["You (chat or operator UI)"] --> Runtime
    Runtime --> MCP["FastMCP tool server (this repo): 24 tools"]
    MCP --> Vault[("Markdown vault")]
    MCP --> DB[("SQLite: tasks, habits, activity, notes")]
    MCP --> Search["Semantic vault search (embeddings)"]
    MCP --> Multi["Vision / OCR / image gen / calendar"]
    Dash["Ops dashboard (FastAPI)"] --> DB
    Dash --> Metrics["System + gateway uptime, jobs, token cost"]

Full detail in ARCHITECTURE.md.

Quick start

python3 -m venv .venv && . .venv/bin/activate
make install
make test          # 22 tests, no live secrets needed
make db-init        # create the SQLite schema

Run the tool server or the dashboard locally:

# MCP tool server (stdio)
AGENT_VAULT_DIR="$PWD/.vault" python3 scripts/agent_mcp.py

# ops dashboard at http://localhost:7474
AGENT_VAULT_DIR="$PWD/.vault" python3 scripts/dashboard_main.py

Configure it for your setup

Copy config.example.json to config.json and set your vault path, the model for each tier (scripts/models.py), and the schedule. Point the paths at your own vault and runtime with the AGENT_* variables in .env.example. Wire the tool server to an MCP runtime and a scheduler, and run it locally or on a VPS for always-on operation. Step by step in RUNBOOK.md.

Defaults to OpenAI (gpt-4o and friends); switch to any OpenAI-compatible provider (z.ai, DeepSeek, a local server) with a one-line edit in models.py.

Built with

Python · MCP (FastMCP) · SQLite · FastAPI · OpenAI-compatible model providers. No framework lock-in; the tool layer is plain Python behind a typed MCP surface.

Safety and privacy

Credentials never live in the repo (.env locally, or a server env file via AGENT_ENV_FILE). Vault paths are resolved under the vault root and reject escapes; free-form SQL is read-only. No hostnames, IPs, vault contents, or personal data are committed. The included config is a template, not a dump from a real deployment.

License

MIT, see LICENSE. Use it, fork it, adapt it for your own setup.

Contact

Built and operated by Mira Solutions, an AI engineering and automation studio. mira.solutions06@gmail.com