Friday — A 24/7 AI Assistant Built Entirely on Claude Code

An always-on personal AI system using only Claude Code CLI ($100/month) and Telegram — no custom AI, no cloud VMs, no fine-tuning.

Live page: missingus3r.github.io/friday-showcase

What Is This?

A self-evolving AI assistant that runs 24/7 on a standard Windows, Linux, or macOS machine. It talks over Telegram, runs scheduled tasks, manages files and projects, and keeps persistent memory across sessions — and it learns from its own behavior: acquiring skills, reflecting nightly, inferring preferences from repeated corrections, and proposing its own improvements. All powered by Claude Code (Anthropic's CLI agent) on the $100/month Max Plan — no custom AI backend, no fine-tuned models, no orchestration framework. The only custom code is a lightweight Flask server for memory and the self-evolving subsystems.

How It Works

Claude Code sits at the center, reaching external services through MCP (Model Context Protocol) plugins and shell tools. A CLAUDE.md file is the system prompt — behavior, tools, and cron schedules. Self-evolving subsystems (skills, reflection, preference learning, world modeling, proposals) run as scheduled jobs on the same memory server, feeding insights back into behavior with no fine-tuning.

User (Telegram) ---> Claude Code (with MCP plugins)
    |
    |--> Memory API        ---> SQLite (conversations, memories, embeddings)
    |--> Self-Evolving     ---> Skills, reflections, preferences, world model, proposals
    |--> Knowledge Base    ---> Notes, wiki, structured data (Notion MCP)
    |--> GitHub            ---> Repos (push, commit, PR)
    |--> Voice API         ---> Text-to-speech / Speech-to-text (ElevenLabs)
    |--> Email (REST API)  ---> Send, receive, forward (AgentMail)
    |--> Web Search/Fetch  ---> News, research, data
    |--> Cron system       ---> Recurring autonomous jobs
    |--> Local tools       ---> Shell, scripts, system utilities

Why Not Use an Agent Framework?

Agent frameworks (OpenClaw, NemoClaw, and others) are impressive but add layers: custom runtimes, orchestration code, deployment pipelines, often their own API costs.

Claude Code already is the runtime — native tool use, MCP plugins, cron scheduling, sub-agent spawning, file I/O, git, and shell, all built in. No glue code between the LLM and the tools.

One plan, one CLI, one model: stay within a single subscription, respect the provider's ToS, and let the model do what it was designed to do.

Star History

The Stack

Screenshots

Telegram Examples

How Friday looks in everyday use — the same Telegram chat, across different tasks.

Memory Graph

D3.js force-directed view of conversations, memories, and entities. Drag, zoom, and click to explore.

Architecture Diagram

Map of all components and their connections. Nodes are draggable; positions persist server-side.

The memory server renders all logs, memories, and entities as interactive graph nodes. This implementation is called Friday, but the name is just a variable in the CLAUDE.md configuration.

Capabilities

• Scheduled briefings — weather (Open-Meteo), forex/crypto (DolarAPI, ExchangeRate), AI news (WebSearch), movies (YTS)
• Autonomous monitoring — scans 60+ HuggingFace orgs, blogs, and aggregators for new AI model releases
• Notes & knowledge — saves links, text, and structured data to Notion (MCP) and local markdown
• Voice messages — transcribes audio in (ElevenLabs Scribe STT), replies with synthesized speech (ElevenLabs TTS)
• Video analysis — downloads, transcribes (YouTube subs / ElevenLabs), and reports via LLM (Groq / OpenRouter)
• Email — check, draft, send, forward (AgentMail REST API)
• Git — commit, push, PRs, repo management (GitHub API + git CLI)
• Web research — search, fetch, summarize, report back (WebSearch + WebFetch)
• Self-healing crons — monitors its own jobs and recreates any that expire
• Proactive messaging — reaches out first: check-ins, reminders, follow-ups based on memory and context. An optional rule-based outbox paces them like a person would: non-urgent items wait for a good moment (not while you sleep, work, commute or sit in a meeting) and go out together in one short message instead of on the clock
• Skill acquisition — extracts reusable patterns from solved tasks, stored with trigger patterns
• Daily self-reflection — nightly review of logs for mistakes, wins, and patterns → insights
• Preference learning — infers rules from repeated corrections and applies them automatically
• World modeling — a behavioral model of the user (activity, topics, correlations) with confidence + expiry
• Self-improvement proposals — formal proposals with diffs; never applied without approval
• Memory API health — periodic checks with auto-restart and failure notification
• Live camera & object detection (optional extra) — turns a security camera (IP, USB, etc.) into a local detector: YOLOv8n (Ultralytics + OpenCV) detects people on-device, captures a snapshot with a bounding box + a ~10s clip, and alerts via Telegram — nothing leaves the machine. Enabled by the --with-camera flag in the setup prompt

Scheduled Jobs

A handful of autonomous cron jobs keep the system alive and learning. Over time the assistant consolidated its own schedule — collapsing 25 sprawling jobs into 7 grouped ones (25 → 8 → 7, retiring jobs whose purpose expired). Exact times are arbitrary, so adapt them to your routine; what matters is the split between the essential jobs (the framework itself) and the project-specific ones you swap for your own.

Essential — the framework:

Project-specific (examples — swap for your own): AI-model release monitor, dataset scrapers, monthly usage reports. These aren't part of the core framework — they're whatever you want running 24/7.

The heartbeat and briefing crons act as watchdogs — they verify every job is active, recreate missing ones, and refresh them before the 7-day lifetime expires. The dashboard's Crons tab shows a two-column diff: runtime jobs with live countdowns vs disk-persisted prompts with sync badges.

Disaster recovery

/backup/export and /backup/import move a whole-DB snapshot over HTTP. Schedule a nightly off-site rsync of the export; if the machine dies, import the snapshot into a fresh install, restart, and Friday picks up where it stopped. A 💾 Backup button on the dashboard does the same from the browser.

Memory Server

Reference snapshot: github.com/missingus3r/memory-graph. ⚠️ Not a template to clone — that code is generated autonomously by Claude Code following the SETUP.md in this repo, so every user's version differs. The repo is committed only as the source for the screenshots and a