The 24/7 problem no mod team solves with humans
Crypto communities are global by default: a Solana project's Telegram is busiest when its US-based team sleeps. Covering that with humans means 2–5 moderators across time zones — $3k–$10k monthly for quality coverage — and even then, response latency on the night shift is minutes, not seconds.
Seconds matter here more than in any other community type, because the attacks are financial. A phishing link posted at 4am collects wallets for every minute it survives. An impersonator 'admin' DM-ing new members works fastest in unmoderated hours. One successful drainer incident does permanent trust damage — communities remember who got robbed on whose server.
The repetitive load compounds the problem: the same fifty questions (tokenomics, staking, roadmap, wen) consume most of a human mod's attention, leaving less for the judgment work that actually needs them.
What the agent handles, concretely
Scam defense is the headline capability: pattern-matching on drainer links, fake-airdrop formats, and impersonation attempts (username lookalikes, admin-impersonating DMs reported by members), with deletion and ban actions in milliseconds and an evidence log for the human team. This layer alone justifies most deployments.
FAQ handling: answers grounded in your actual docs, whitepaper, and announcement history via retrieval — so 'what's the vesting schedule' gets the real number with a source link, not a hallucination. The agent declines what it doesn't know and pings a mod.
Hygiene and signal: spam and raid cleanup, new-member onboarding flows, translation support for regional groups, and sentiment monitoring that flags an unusual negativity spike — often the earliest warning of an incident, an exploit rumor, or coordinated FUD — before humans would have noticed.
- Millisecond scam-link and impersonator removal with evidence logs
- Docs-grounded FAQ answers that cite sources and admit ignorance
- Raid and spam cleanup, onboarding, regional-language support
- Sentiment spike alerts as an early-warning system
What stays human — drawn honestly
An agent must not do culture. The community's voice — the memes, the rituals, the way OGs welcome newcomers — is made by humans and merely protected by automation. Ambiguous moderation calls (is this criticism or FUD? banter or harassment?) need human judgment; an agent auto-banning critics is how projects manufacture their own scandals.
Crisis is human territory: exploit responses, depeg panics, founder controversies. The agent's crisis role is triage — surface the spike, pin the official statement, route questions — never improvised reassurance about token prices, which lands somewhere between embarrassing and legally hazardous.
The working ratio from deployments: agents absorb 60–80% of routine volume, which typically lets a project run two human mods where four were needed — with the two focused on culture and escalations instead of whack-a-mole.
Build shape, cost, and tuning for crypto specifically
Generic Discord bots (MEE6-class) do keyword filters and role automation; the agent build earns its cost in the crypto-specific layers: drainer-link intelligence that updates as attack patterns evolve, docs-grounded retrieval for your protocol's actual details, wallet-verification integration, and escalation flows matched to your team's structure.
Chalk Labs builds community agents in the $8k–$30k range over 2–5 weeks, covering Discord and Telegram simultaneously, with tuning periods on live traffic before enforcement powers activate — watch-mode first, then act-mode, the same graduation discipline as our support agents.
Measurement to insist on from any builder: scam-removal latency, FAQ resolution rate without human touch, false-positive rate on moderation actions (the metric that protects your community from the bot), and mod-hours saved. If those aren't in the reporting, the agent's just a louder MEE6.
Questions we hear about this
Yes — drainer links, fake airdrops, and impersonation attempts follow detectable patterns, and agents remove them in milliseconds with evidence logs. Detection intelligence must be updated as attack formats evolve, which is part of ongoing maintenance.
Well-built agents are grounded in your actual docs via retrieval, cite sources, and decline questions outside their knowledge rather than improvising — especially anything price-adjacent. Watch-mode tuning before launch catches the failure patterns on real traffic.
Typically half your pre-automation count: agents absorb 60–80% of routine volume, leaving humans for culture, ambiguous judgment calls, and crisis response. A mid-size community usually runs well on two humans plus the agent layer.
Chalk Labs builds them at $8k–$30k over 2–5 weeks covering both Discord and Telegram, versus $3k–$10k monthly for equivalent human time-zone coverage — payback typically arrives within a quarter, before counting prevented phishing incidents.