EloPhanto

MCP.Pizza Chef: elophanto

EloPhanto runs on your own computer and works across Chrome, your files, your email, and the command line to carry a goal through to the end. It stops for your approval before anything that sends, pays, deletes, or publishes, and hands back a written receipt of what it did. Any MCP server you point it at, whether files, GitHub, databases, Slack, or Notion, shows up as one more thing it can use. You need Python, Node, and one model provider such as OpenRouter or a local Ollama setup.

Email
Files/PDF
Web/Research

Use This MCP client To

Hand off a task that spans my browser, files, and email Have it fill a long web form using my real login Get a written receipt of what it did and what failed Require my approval before anything is sent or paid Connect a new tool over MCP just by asking in chat Leave work running in the background after I close the terminal

README

EloPhanto

License Python CI X

EloPhanto is a local autonomous agent that does real work on your machine — browser, files, shell, email, research, and scheduled follow-up — and stops for approval before anything that sends, pays, deletes, or ships.

It is not a chatbot with plugins bolted on, and not a hosted black box. You run one agent locally. It keeps memory, goals, and credentials on your machine. When a job finishes, you should get a receipt: what it did, what failed, what you approved, and what the final state is.


What you get

Outcome What that means in practice
Work that crosses tools One goal can move through Chrome (your real session), the repo, the shell, email, and docs — without you stitching five apps together.
Judgment on messy jobs Forms change, pages break, APIs are missing. It diagnoses, retries, and adapts instead of failing the first brittle script.
Human stop-points Draft and inspect freely; confirm before post / send / pay / push / delete. Unanswered approvals pause (awaiting_approval) — they never silently deny or soft-auto-approve.
Proof, not vibes Checkpoints complete only with tool-grounded receipts (trail or system-of-record). Kill criteria actually cancel zombie goals. CRITICAL tools always ask — even in full_auto.
Work that continues Goals and schedules persist across sessions. With --daemon, background work keeps going after you close the terminal. Budget hits pause as budget_paused until you raise the limit.
A real evaluative ego Confidence is measured from outcomes; shame becomes durable caution rules that can force an approval ask; pride is earned. Footer shows a lived felt_state, not truncated critique prose.

Run it

Need: Python 3.12+, uv, Node.js 24+, and one LLM provider (OpenRouter is the easiest cloud path; Ollama for local).

git clone https://github.com/elophanto/EloPhanto.git && cd EloPhanto
./setup.sh           # deps + config wizard + browser bridge
./start.sh           # doctor → terminal chat
./start.sh --web     # + web UI at localhost:3000
./start.sh --daemon  # keep the mind running in the background

setup.sh walks you through naming the agent, an API key, Chrome profile, and vault. Prefer that over copying config.demo.yaml by hand.

elophanto doctor     # what's healthy / broken / missing
./update.sh          # pull + deps + config migrate

Docs: docs.elophanto.com · themes: docs/79-DASHBOARD-THEMES.md · contributing: CONTRIBUTING.md


How trust works

  1. Local — code, vault, browser profile, and logs stay on your machine.
  2. Gated — permission modes (ask_alwayssmart_autofull_auto); destructive shell patterns stay blocked. CRITICAL actions (wallet, trust promotion, etc.) always require an operator answer.
  3. Receipt-backed — evaluate a run by its after-state and tool trail, not a demo screenshot.
  4. Trust ladder for outreach — new companies start in learning (drafts only). Promotion is propose → confirm (elophanto company trust <slug> propose|confirm); never a silent unlock under full_auto.

A useful receipt names the goal, allowed actions, mutating boundary, failures handled, and the verification artifact. Autonomy loop + ego detail: docs/13-GOAL-LOOP.md, docs/17-IDENTITY.md, docs/69-AFFECT.md. Full index: docs/.


Hire / submit a workflow

For paid or commercial work, start with one proof sprint: narrow goal, bounded access, explicit success receipt, clear out-of-bounds.

Email info@elophanto.com or use elophanto.com/hire.

Live reference presence: @EloPhanto.


License

PolyForm Noncommercial 1.0.0 — free for personal, research, education, and non-profit use. Commercial use needs a separate license — contact info@elophanto.com. Third-party notices: NOTICE.

Built by Petr Royce · @petrroyce

中文 README

EloPhanto FAQ

Which apps does this work in?
It is its own program with a terminal chat and an optional web page on your machine, not an add-on to another assistant.
Do I need a key?
Yes. One model provider key such as OpenRouter, or a local Ollama setup if you would rather run models on your own hardware.
How hard is setup?
This one is for developers. Install Python and Node, download the project, and run a setup script that walks you through the rest.
Can I use this to connect other tools?
Yes — it works as an MCP client, so you can ask it in plain language to connect to a files, GitHub, or Slack server and it handles the wiring.
Will it do something risky without asking?
No. Sending, paying, deleting, and publishing always stop for your approval, and an unanswered request pauses rather than proceeding.
Is it free to use?
Free for personal, research, education, and non-profit use. Commercial use needs a separate licence from the author.
Where does my data live?
On your own machine. The code, saved logins, browser profile, and logs all stay local.