Your own AI agent.
It works while you sleep
and improves itself.
ChatGPT and Claude wait for your next prompt. EloPhanto doesn't. Give it a goal and it codes, browses, researches, and runs the long jobs on its own – then keeps working while you sleep. It learns from every task and builds the tools it's missing, so every week it does more and gets harder to replace. Run it on your own machine or in the cloud; the agent is yours either way.
$ git clone https://github.com/elophanto/EloPhanto.git && cd EloPhanto && ./setup.sh$ ./start.sh$ ./start.sh --web$ ./start.sh --daemonA few minutes to set up: one API key, or your existing ChatGPT login. Then you're talking to it.
Three layers, each mechanically wired.
The LLM never writes its own numbers. Confidence moves on real failure signals; emotion runs on a substrate that decays over time. This is what makes the third week of running feel different from the first – a self-image that has been hurt, recovered, and revised.
who it claims to be
Values, beliefs, and capabilities discovered through reflection, written to a nature.md the agent edits over time. You name it in the setup wizard; the rename propagates end-to-end across DB, knowledge, dashboard, and LLM context.
→ nature.md · docs/17-IDENTITYhow reality graded that claim
Per-capability confidence moved by three real failure channels: tool outcomes, verification PASS/FAIL, and a 13-rule user-correction detector. Failures hit harder than successes; unused capabilities decay. Grounded in Higgins' Self-Discrepancy Theory.
→ core/ego.py · docs/17what it's feeling right now
State-level emotion on a PAD substrate with OCC appraisal labels. Three channels decay over minutes-to-hours. Corrections fire frustration, checkpoints fire pride. Biases router temperature, prompt tone, and risk appetite.
→ core/affect.py · docs/69-AFFECTEgo is who the agent has become; affect is who it is right now. Built on Higgins (1987), Mehrabian's PAD, and OCC appraisal. To our knowledge, no other open-source autonomous agent ships all of this.
Steady and a little guarded. Recent work landed; I'm moving with intent but double-checking before I commit.
router temperature bias +0.02 (looser, more exploratory)
→ PAD substrate · OCC appraisal · core/affect.py
“I'm most sure of myself in the browser and against a knowledge base I've built; I've been wrong enough about trading to stay humble there. I finish what I verify, not what I assume.”
the agent's words, not marketing→ core/ego.py · knowledge/self/ego.md
This is the actual software.
A terminal chat and a real-time web dashboard, shipping today. Watch the mind think, browse the knowledge it built, audit every tool call, and inspect the ego and affect for yourself.
Useful in minutes.
Irreplaceable in weeks.
From the first message it has 200+ tools and every channel, and it does real work right away. What changes over the weeks is how well it knows your work, how much it does on its own, and how much you let it.
Working immediately
Clone it, name it, go. 200+ tools and every channel are live from the first message. You approve each action while you get the measure of it.
It learns your way
You correct mistakes and it keeps them. Every “no” or “not like that” is caught and saved as a lesson it won't repeat.
It stops asking
Lessons load automatically on similar tasks, skills kick in, and the safe work starts auto-approving. You stop repeating yourself.
It runs itself
Goals run across sessions, the background mind works between your messages, and specialist clones take whole workstreams. You're the operator, not the driver.
200+ tools, 177+ skills, seven model providers (including your own ChatGPT subscription), its own wallet and inbox, and every major channel. A few of the things that means in practice:
It learns how you work
Correct it once and it keeps the lesson. Findings from every task load automatically on the next similar one, so you stop explaining the same thing twice.
It builds its own tools
Hit a task it can't do? It researches, writes the tool, tests it, and uses it, then has it ready forever. The agent grows; it isn't a fixed script.
It works while you sleep
Daemon mode keeps the mind thinking between your messages and across reboots. Close the laptop, come back to finished work.
It becomes a team
When work goes parallel it spawns specialist clones, each a full agent with its own memory. Risky plans get reviewed by three models before anything ships.
Your real browser, your accounts
Real Chrome with your own sessions and cookies, 49 tools, plus pixel-level control of any desktop app. Already logged in everywhere you are.
Dangerous work, contained
It runs rm -rf, untrusted installs, anything risky inside a locked-down throwaway container that can't touch your machine. Blast radius: zero.
One entity.
Not a persona stable.
Other “agent platforms” host N character personas behind one engine – swap the SOUL.md, swap the bot. EloPhanto is structurally different: this installation is one agent. One identity, one wallet, one self-model grown over weeks. When you want more, you spawn another full EloPhanto – separate vault, separate wallet, separate self-model. Peers, not personas.
It gets harder to replace the longer it runs. That's the architecture, and everything else stands on it.
accumulates per-capability confidence – for whom, if the persona is swappable per request?
carries state between calls – whose frustration, whose pride?
builds on-chain reputation – five personas sharing one wallet is dilution.
tracks a Brier score for this predictor – meaningless for a rotating set of facades.
pins known-hosts per PeerID – multiple personas behind one key would break it.
Ship a SaaS overnight
“Build me an invoice SaaS.” It validates the market, plans the MVP, spawns coding agents in isolated worktrees, deploys to Vercel + Supabase. You approve at each gate.
Development team
Give it a feature spec. Get back reviewed, tested code from multiple agents – Claude Code, Codex, Gemini – that challenged each other's work.
Self-building
Ask it to do something it can't. Come back to find it built the tool, tested it, and already used it. Next time it just knows how.
Cross-platform intel
Ask a question that needs email, web, and documents. Get one answer with sources, gathered across channels.
Persistent specialists
“I need ongoing marketing and research.” It spawns clones with their own mind, vault, and schedule. Teach through feedback; trust-scoring auto-approves the good ones.
Background mind
Close your laptop. The daemon keeps the mind thinking, executing, improving. Open it tomorrow to results.
You stay in control.
Run it on your own machine or in the cloud. One WebSocket gateway, one set of sessions across every channel, and a permission gate on every risky action. Nothing happens on its own until you say so.