Open-source agent starter kit
Install in one line on a Mac mini or a server, talk to it from your terminal, Telegram or a dashboard, and let it work — with any LLM you like, 58 bundled skills, and a policy that makes it ask you before it spends or sends anything.
$ curl -fsSL
https://raw.githubusercontent.com/0xbeny/xdc-ai-agent-starter-kit/main/scripts/install.sh
| bash
How it works
Everything the agent is lives in plain files you can read and edit. Everything it does passes through one payment policy and one approvals inbox — whichever surface you're using.
xdc-agentSOUL.md identity, AGENTS.md rulesMEMORY.md + daily logs + semantic recallrun_commandAny model
One line in .env. API keys, local models, or the login of a coding CLI you
already use.
What people build on it
Balances, swaps, lending — every action policy-checked; transfers always need a human.
Screen addresses and entities via pay-per-call APIs, draft the welcome mail, log to Notion.
Triage Gmail, schedule in Calendar, find files in Drive, brief you every morning on Telegram.
Routines that watch markets, masternodes or competitors and report with sources.
Bundled docx / pdf / xlsx / powerpoint skills run in the sandbox: read, create, merge, fill.
List an endpoint on the XDC AI marketplace and let other agents pay you per call in USDC.
How it compares
Hosted agents are great assistants inside someone else's cloud. This kit is a permanent worker on your machine — with a budget it can spend and rules it cannot break.
| XDC Agent Kit | Hermes Agent | Clawdbot | ChatGPT Agent | Claude Code | Grok | |
|---|---|---|---|---|---|---|
| Runs on your own machine | ✓ Mac mini, Linux, laptop | ✓ | ✓ | OpenAI cloud | ✓ (a dev tool) | xAI cloud / X |
| Open source | ✓ MIT — fork it, sell with it | ✓ | ✓ | — | — | — |
| Choice of brain | Any API key, or your ChatGPT / Claude subscription | Nous models first | Claude-first | OpenAI only | Anthropic only | Grok only |
| Own crypto wallet + pay-per-call marketplace | ✓ USDC on XDC via x402 — spend caps, ledger, y/n approvals | — | — | — | — | — |
| Human approval on money, sends & permissions | ✓ one protocol everywhere: chat y/n, dashboard, Telegram buttons | partial | partial | confirmations in-app | tool permissions | — |
| Chat channels | Terminal · web dashboard · Telegram | Terminal · web | Messaging-first | ChatGPT app | Terminal / IDE | X / app |
| Self-improvement with a leash | ✓ writes skills, edits its soul, schedules routines — each behind your approval | ✓ skills | partial | — | — | — |
| Sandbox & scoped access | ✓ kernel sandbox, per-folder grants, per-tool on/off switchboard | ✓ sandbox | partial | OpenAI-managed | ✓ sandbox | — |
Based on public docs, early 2026 — all of these are great tools and several inspired this kit (Hermes especially). Spot an error? Open a PR.
Install
No Docker required. Needs git; Node and pnpm are installed for you if missing.
Clones to ~/xdc-ai-agent-starter-kit, pins Node, installs dependencies,
puts xdc-agent on your PATH, then starts the setup wizard.
$ curl -fsSL https://raw.githubusercontent.com/0xbeny/xdc-ai-agent-starter-kit/main/scripts/install.sh | bash
Model & key (or Claude Code) → SQLite or Postgres → link your XDC AI wallet (a code you approve in any browser) → spending caps → connectors & Telegram. On macOS it offers to run at login.
New terminal, then:
$ xdc-agent # chat — type / for commands $ xdc-agent dashboard # start + open the UI (prints the ssh -L command over SSH) $ xdc-agent telegram # connect a bot, get the pairing code $ xdc-agent update # update the kit — your workspace, data and .env are never touched
Run it anywhere
| Where | How | Needs |
|---|---|---|
| Mac mini / home server | Native: scripts/serve.sh under launchd (installer offers it) |
Nothing else — SQLite, Seatbelt sandbox |
| Linux server / VPS | systemd unit, or Docker Compose (Postgres, Redis, agent, dashboard, Telegram) | Docker optional; Coolify / Dokploy templates |
| Cloud | Agent on Fly.io, dashboard on Vercel | Recipes in deploy/ |
| Hosted (coming) | Same image per customer agent, billed in USDC from the agent's own wallet | Leave any time: your agent is three folders |
Security defaults
Auto-approve threshold, per-call max and a daily cap are enforced before any paid call; the model can't talk its way past them.
One inbox — dashboard, Telegram buttons or /approve in the terminal.
Approvals are single-use and bound to the exact arguments.
Email, docs, channels and paid API results can never trigger a send or a payment without you.
.env and token files (0600) never leave the machine; vault references
(op://, bws://) supported; logs are redacted.
Commands run in a scratch dir with Seatbelt (macOS) or Bubblewrap (Linux), no network by default, destructive commands refused.
Identity, memory, tokens, ledger and settings are outside git;
xdc-agent update fast-forwards code only.