How Do I Deploy a Polymarket Bot? VPS Setup, Key Safety, and Operations (2026)
Deploying a Polymarket bot is an operations problem before it is a trading problem: the code that finds edges matters less than the setup that keeps it running through reconnects, rate limits, and 3 a.m. failures. The production baseline: a VPS with stable connectivity, a proxy key in environment variables (never your main key), process supervision that restarts cleanly, tight clock sync, and a kill switch you have actually tested.
What does the baseline stack look like?
| Layer | Choice | Why |
|---|---|---|
| Host | VPS (any major provider) | Always-on; stable IP for rate limits |
| Runtime | Python 3.11+, venv or container | Reproducible dependencies |
| Process | systemd unit or Docker restart policy | Auto-restart on crash |
| Config | Environment variables | No keys on disk in plain text |
| Logging | Structured logs to file + rotation | Post-mortems need history |
| Monitoring | Dead-man’s-switch ping or uptime check | Silence must alert someone |
How do I handle keys safely?
Keep the main EOA key in cold storage, put only the proxy signing key on the VPS in a root-readable env file, keep a rotation proxy ready for compromise, and never log or commit key material.
The proxy pattern does the heavy lifting:
- Main EOA key: cold storage, never on the server
- Proxy key: the only signing key on the VPS, funded with working collateral
- Rotation path: a second proxy ready to take over if the first is compromised
# /etc/systemd/system/pm-bot.env (root-readable only)
PM_PROXY_KEY=0x...
PM_PROXY_ADDRESS=0x...
PM_CLOB_HOST=https://clob.polymarket.com
Rules: chmod 600 the env file, never log key material, never commit configs, and
keep dev and prod on separate proxies so a debugging mistake can’t touch real
capital.
What does the systemd unit look like?
A minimal unit loads the env file, runs the venv python entrypoint from a fixed working directory, and uses Restart=always with a 5-second delay so crashes self-heal while your code handles logic-level recovery.
[Unit]
Description=Polymarket bot
After=network-online.target
[Service]
EnvironmentFile=/etc/systemd/system/pm-bot.env
WorkingDirectory=/opt/pm-bot
ExecStart=/opt/pm-bot/.venv/bin/python main.py
Restart=always
RestartSec=5
[Install]
WantedBy=multi-user.target
Restart=always with a short delay covers crashes; your code must still handle
graceful shutdown (SIGTERM → cancel open orders → flatten or persist state) so a
redeploy doesn’t strand orders on the book.
What breaks deployments in practice?
- Clock drift — signed requests need fresh timestamps; enable NTP sync
(
systemd-timesyncdor chrony) or face random signature rejections - Silent rate-limit death — a 429 loop without backoff burns your bucket; log
Poly-RateLimit-Remainingand alert on sustained decline - Unhandled reconnect state — after a WebSocket drop, reconcile books and positions before resuming; a bot that trusts pre-crash state doubles positions
- Momentum-less restarts — the bot forgets its open orders after restart; persist an order ledger locally and reconcile against the API on boot
- Alert-less silence — the most expensive failure is the quiet one; ping a monitoring endpoint every cycle so absence of signal is itself a signal
What does the kill switch need to do?
One flag must cancel all open orders, halt new signal generation, optionally flatten positions (your documented choice), and leave the process alive for post-mortem — tested monthly.
One flag — a file, an API endpoint, a Telegram command — that:
- Cancels all open orders (
DELETE /cancel-all) - Halts new signal generation
- Optionally flattens positions (your choice; document it)
- Leaves the process alive for post-mortem
Test it monthly. A kill switch that has never fired is a hypothesis, not a control.
FAQ
Q: Do I need a VPS, or can I run the bot locally? A: Local works for development; production wants a VPS — uptime, stable IP (your rate-limit identity), and no sleep-mode surprises. Home IPs also change, which resets Cloudflare’s per-IP budgeting.
Q: Which key lives on the server? A: Only the proxy wallet’s key. Your main EOA stays offline; a leaked proxy key risks that account’s collateral, not your identity.
Q: How do I update the bot without stranding orders? A: Graceful shutdown on SIGTERM: cancel resting orders, persist state, exit. systemd restarts the new version, which reconciles against the API before trading.
Q: Should I containerize?
A: Optional but recommended — it pins dependencies and makes the systemd unit a
one-liner (docker run --restart always). The operational rules (env keys, clock
sync, monitoring) are identical either way.
Q: What monitoring is enough? A: At minimum: process liveness (ping/uptime check), error-rate log alerts, and a daily balance reconciliation against the Data API. Any divergence between your ledger and the chain is an incident.
Sources
Disclaimer
Not financial advice. Educational purposes only. Updated October 2026.