Local-first · daily assistant
personal agent is a daily self-assistant that runs on your own machine — turning everyday activity into a living profile, a self-updating website, and a triaged morning digest. Everything stays local; only what you point it at ever leaves.
About five minutes from clone to your first run.
Clone the repo and install it into a virtual environment. Python 3.11 or newer — that's the whole dependency footprint.
git clone https://github.com/tzhouam/personal-agent && cd personal-agent python -m venv .venv && source .venv/bin/activate pip install -e .
Run the guided wizard. It walks every section — LLM, GitHub, email — writes your .env, and validates each one live, so a wrong key fails now, not at 7 a.m. tomorrow.
assistant init # guided setup, validated live assistant init --check # re-check anytime: ✅ / ⚠️ / ❌
MiMo mimo-v2.5 on the standard plan — natively multimodal, so photos and receipts just work. Paste this into your .env:
ANTHROPIC_BASE_URL=https://token-plan-cn.xiaomimimo.com/anthropic ANTHROPIC_MODEL=mimo-v2.5 ANTHROPIC_DEFAULT_HAIKU_MODEL=mimo-v2.5 LLM_SUPPORTS_IMAGES=true
Dry-run first — it writes the digest to disk without sending anything. Happy? Drop the flag for the real thing, then let cron run it each morning.
assistant run --dry-run # full pipeline, nothing sent assistant run # the real thing # once a day, early morning: 0 7 * * * assistant run || assistant run --resume
The wizard writes all of this and validates it live — but if you'd rather edit .env yourself, here's every group. Run assistant init --check when you're done.
A fine-grained personal access token — read-only access to your repos and notifications is enough. It powers the activity collector.
Create a fine-grained token ↗GITHUB_TOKEN=github_pat_… GITHUB_USER=your-username
The easiest delivery path — a free API key from resend.com sends your daily digest.
Get a Resend key ↗RESEND_API_KEY=re_… RESEND_FROM=onboarding@resend.dev DIGEST_TO=you@example.com
Instead of — or alongside — Resend. With Gmail it also unlocks the inbox collector and email chat; use an app password, not your login.
Create a Gmail app password ↗SMTP_HOST=smtp.gmail.com SMTP_PORT=587 SMTP_USER=you@gmail.com SMTP_PASSWORD=app-password DIGEST_TO=you@gmail.com
Point it at your Overleaf git-bridge (or any git remote holding your .tex), then run assistant resume-init. Every edit is gated on your explicit okay.
Overleaf Git setup ↗RESUME_REMOTE_URL=https://git:TOKEN@git.overleaf.com/PROJECT_ID
Publish your profile to a GitHub Pages repo. The password encrypts the private todos / reading / routines pages in your browser.
Create the Pages repo ↗WEBSITE_REPO=you/you.github.io WEBSITE_PASSWORD=a-strong-passphrase
Sharpens chat answers and task research. Keyless DuckDuckGo works out of the box; add a key for better results.
Get a Gemini key ↗GEMINI_API_KEY=… # recommended TAVILY_API_KEY=… # or this
Route each role to its own model (LLM_ROLES), or run Mixture-of-Agents — several models propose in parallel and one aggregates — on the offline roles (LLM_MIXTURE). A dead or empty member is dropped, so it keeps working as long as one proposal survives.
Multi-model routing docs ↗LLM_ROLES={"chat":{"model":"mimo-v2.5"}} # one line needs no quotes # each member is {model, base_url?, api_key?}; span lines only if 'single-quoted' LLM_MIXTURE='{"members":[ {"model":"mimo-v2.5","base_url":"https://token-plan-cn.xiaomimimo.com/anthropic","api_key":"tp-…"}, {"model":"qwen3.6-plus","base_url":"https://dashscope.aliyuncs.com/apps/anthropic","api_key":"sk-…"}], "aggregator":{"model":"deepseek-v4-pro","base_url":"https://api.deepseek.com/anthropic","api_key":"sk-…"}, "roles":["pipeline"]}'
Every knob — website marks sync, WeChat announce, research sources — is documented inline in .env.template.
One quiet run each morning turns yesterday into four things worth having.
Morning digest
GitHub notifications sorted red / yellow / white against your profile, new papers worth reading, and your open todos — one email.
Living profile
A picture of you that keeps itself current — every skill cited to something you actually shipped. Nothing invented.
Personal site
Publishes a website straight from your profile, rendered deterministically — no model ever writes a public page.
Résumé sync
Keeps your résumé matched to your profile, gated on a compile and your explicit okay — never pushed without you.
Optional · chat channel
Email chat works the moment setup finishes. For WeChat — plus proactive reminders and a gateway that can run your daily schedule — connect it through OpenClaw. The agent stays the only brain; OpenClaw is just transport.
# the gateway needs Node ≥ 22.19 npm i -g openclaw @tencent-weixin/openclaw-weixin
# from the repo — --link tracks it, so updates are a git pull openclaw plugins install --link ./openclaw-plugin
openclaw channels login --channel openclaw-weixinThe gateway binds to the account that scans, so the sender is always you — no pairing to approve. It can also run your daily digest and weekly consolidation as cron jobs, so a WeChat setup needs no separate scheduler.
Day to day, you text it — over WeChat or email. Log spending, track workouts, send a receipt photo, or hand it a whole multi-step task.
Everything runs and stays on your machine. The only things that leave are the LLM calls, your digest email, and the sites you point it at.