Personal Agent

developer preview

AllPath · Personal Agent

A personal agent that
teaches itself to you.

Most agents hand you fifty tools and a manual. AllPath starts as a plain chat and introduces one capability at a time — when your own messages show you're ready for it, and only until you've actually used it.

View on GitHub First briefing by tomorrow

Runs: On your machine
Models: Claude · OpenAI · Gemini · Grok
Stack: Python stdlib · SQLite

allpath-agent

YOU

❯ send me a news briefing every morning at 7

● create_automation · guided flow

ALLPATH · SETUP

Please confirm this automation:
• Morning brief — every day, 7:00 America/Los_Angeles
• Delivered to: telegram · this conversation

YOU

❯ confirm

ALLPATH · SETUP

Saved. Next run: tomorrow 07:00.

NEXT · workspace_files

Start Allpath in a project folder and ask it to explain the codebase.

the actual terminal vocabulary — panels, activity lines, one NEXT card

Why it's different

It keeps a curriculum of itself — and tracks what you've actually learned.

Sixteen capabilities, each with prerequisites and a lifecycle. Progress moves on deterministic evidence — a tool that really executed, a channel that really verified — so the agent knows the difference between “I showed you a tip” and “you can do this now.”

unseenyou haven't met it yet
offeredsuggested once, after your answer
triedyou expressed the intent
succeededverified by a real execution
habitualused three times — never taught again

The launch card

Every start shows exactly one next step, chosen from what you haven't learned yet. Finish it — or dismiss it — and the card moves on. It never repeats a lesson you've completed.

The composer hint

A quiet one-liner beside the input: the next capability worth trying, phrased as something you can type right now. It advances the moment you connect a channel mid-session.

The NEXT card

After an answer — never instead of one — at most one suggestion per session, with cross-session cooldowns. Decline it and it stays declined.

Dismiss any lesson once — it stays dismissed. Nothing curriculum-related ever enters model context.

The first hour

Four conversations. Tomorrow morning it's just there.

chat
connect a model
in conversation — API key hidden, verified before saving
chat
connect Telegram
a four-step guided flow; the bot answers you minutes later
chat
create automation
say it in Telegram — delivery preselects this conversation
you
briefing arrives
next morning, in your pocket, from your own machine

Every step happens inside the conversation — no config files, no dashboard scavenger hunt. The whole path is covered by a deterministic end-to-end test against fake transports, and it's the same sequence the launch card walks you through.

Boundaries, in code

The safety model is part of the product, not the disclaimer.

Evidence, not vibes

The curriculum advances only on verified execution records — a lesson counts as learned when the tool actually ran, never because a tip was shown.

Every side effect waits

Writing files, running commands, clicking pages: side-effecting tools ask first, with the exact bounded arguments in the approval panel.

Secrets never meet the model

API keys and bot tokens go through hidden input into a local secret store. They never enter conversation history, model context, or logs.

Public internet only

The browser and web lookup refuse local and private addresses — on the first request and on every redirect hop.

Unattended runs can't go rogue

Scheduled jobs keep side-effect tools denied. A denied request marks the run “needs attention” instead of failing silently.

Yours, inspectably

Standard-library Python, SQLite, and plain TOML in ~/.allpath-agent. Nothing leaves your machine but the model calls you configured.

Quick start

From nothing to a morning briefing, in five steps.

step 1 / 5
Install
One line. The installer manages Python and a private virtualenv, then opens your first conversation — no API key required to look around.
```
curl -fsSL https://raw.githubusercontent.com/dukesky/allpath-agent/main/scripts/install.sh | sh
```
step 2 / 5
Connect a model — in the chat
Say it; don't configure it. Use a Claude Code or Codex account you already have, or an API key from Anthropic, OpenAI, Gemini, or Grok. Keys are entered hidden and verified before anything is saved.
```
You> connect a model
# guided picker → hidden key entry → verified → live
```
step 3 / 5
Connect Telegram
A four-step guided flow around the official BotFather. Message your bot once afterwards so it becomes a delivery destination.
```
You> connect Telegram
# BotFather → username → hidden token → verified
allpath-agent gateway   # start answering from your phone
```
step 4 / 5
Schedule your briefing
From the terminal or straight from Telegram. The agent collects what's missing, echoes the full plan, and saves only after you confirm.
```
You> create automation
# name → task → schedule (0 7 * * *) → timezone → confirm
# delivery: telegram · this conversation
```
step 5 / 5
Let it run
Install the gateway as a background service and the briefing shows up every morning — no terminal window required.
```
allpath-agent gateway install
allpath-agent gateway status
```

Developer preview — expect sharp edges. 300 automated tests and an honest changelog travel with it.

A personal agent thatteaches itself to you.