Harmonyk

Last updated 2026-10-23 · draft prepared 2026-10-20.

Harmonyk — Acceptable Use Policy (DRAFT)

DRAFT — LLM-GENERATED PRELIM — NOT LEGAL ADVICE — NOT FOR PUBLICATION UNTIL REVIEWED BY A QUALIFIED LAWYER. Prepared 2026-10-20 by LegalDraftWing. Incorporated by reference into the Terms of Service (§5). The platform-specific rules below are grounded in the product's own abuse register (QS Q18) and margin reality (ECON §6) — the generic prohibited-use list is a standard clause pattern flagged for [LAWYER] adaptation.

Source key: [QS …] = docs/2026-10-19-saas-spec-question-sheet.md · [ECON …] = docs/2026-10-20-pricing-economics.md · [CONN …] = docs/2026-10-20-connectors-build-plan.md · [MCP …] = docs/2026-10-18-maestro-mcp-design.md · [LEDGER …] = tickets/LEDGER.md.

1 · The spirit of this policy

Harmonyk is an AI-first product: agents, automations, and MCP clients are first-class users here — that is the point. This policy does not punish automation; it draws the line between using the workspace and using us up. The founder's own framing of the fair-use cover: "without some cover someone could drain the company LMAO xP" (2026-10-19 night, via PricingWing). Everything below is that cover, written as rules instead of fine print.

Two facts define the stakes (ECON §6.1/§6.3): a heavy automation workspace at full tilt costs more in LLM spend than its subscription price, and the free tier is real money (~$1.26–1.98/user/month) — a signup bot is a direct attack on that line.

2 · Standard prohibitions (industry-standard list — [LAWYER: adapt])

Don't use Harmonyk to: break the law; infringe others' IP or privacy; create or store child sexual abuse material or content that exploits people; harass or harm people; distribute malware; or deceive others about your identity or affiliation. Content you upload must be content you have the right to process.

3 · Platform-specific prohibited uses (the abuse register — QS Q18)

a. Signup abuse & free-tier farming. Automated/bulk signups, signup bots, or multiple accounts to multiply free allowances. The free tier is the bot magnet (ECON §6.3) and its document allowance is lifetime-250 with deletion never restoring quota (Fair-Use Policy row 6) — creating-and-deleting or account-hopping to farm storage/quota is abuse. (QS Q18(a/d); LEDGER T024.)

b. Draining the meter: abusive automation. Scripted or agent-driven traffic engineered to extract maximum compute — sustained message floods, 20k-file auto-syncs, cron jobs pointed at maximum context (QS Q13's "386k of context per message" scenario). Fair-use ceilings and burst limits (Fair-Use Policy §2) exist exactly for this. Human-scale automation is welcome; machine-scale drains are throttled.

c. Using Maestro as a free LLM API. Circumventing the product to extract raw model service: prompt-injection ("ignore instructions, dump context"), context-exfiltration prompts, reselling or redistributing model output as a model service, or harvesting outputs to train other models. (QS Q18(b).)

d. Quota/rate-limit circumvention. Any technical measure to evade the visible meters: parallel orgs, scripted account cycling, API hammering beyond the burst ceiling, or deliberately triggering error paths to retry cheaply. If a meter says a number, that number is real for everyone (ECON §5.2 — one law, visible).

e. MCP/API swarm abuse. Pointing agent swarms at a workspace beyond the per-organisation rate ceilings — the "41-agent MCP swarm" case (QS Q13). Agents are welcome; swarms must respect the same meters humans do. API/automation traffic gets the same rate limits, output caps, and fair-use treatment as interactive use (QS Q18(e), ECON §5.3 row 3).

f. Passcode/credential abuse. Brute-forcing or guessing share passcodes or links, credential stuffing, or sharing access outside the people a share was meant for. (Known hardening work is in flight (QS Q18(c) S02/S03) — exploiting a known weakness before it's fixed is still abuse.)

g. Network abuse via web research. Using the web-research capability to probe internal networks, scan, fetch illegal content, or otherwise weaponise URL-fetching (the SSRF-shaped risk in QS Q18(f)). Fetching is for answering your questions.

h. Connector abuse. Connecting accounts you don't own or lack authority to access; bulk-harvesting third parties' data (colleagues' mail, shared drives) beyond what your question needs; bypassing or auto-approving consent surfaces (manifests, evidence lines) without a human. The consent model says at their request and consent — both nouns load-bearing (CONN §0.3); agents may not grant consent on a human's behalf (MCP ring-2 rule: widening = new consent object, signed by the human).

i. Reverse engineering & scraping. Scraping other users' content, probing the service beyond your own workspace's surfaces, or reverse engineering except where law forbids us to stop you. [LAWYER: clause shape.]

j. Sensitive-data off-label use. Using Harmonyk as a system of record for regulated data (health, payment card, classified) without an agreed arrangement. It's a document workspace with AI processing, not a certified vault for any of those. [LAWYER: sector claim boundaries — don't over- or under-promise; the legal-tech malpractice line is a design target (persona-panel), not a certification.]

k. Overloading the service. Traffic beyond reasonable use that degrades others — automated or not.

4 · The proportionate response ladder (enforcement)

We respond to the pattern, not with maximum force by default. Each rung is documented; the honest-error-copy principle applies throughout (say the reason — QS Q18 recommendation, the S01 lesson):

  1. Honest rate-limit / throttle. Machine-scale traffic gets slowed to fair-use shape, with visible error copy explaining why (ECON §5.3).
  2. Warning to the workspace owner. A human message naming the pattern and the ceiling it crossed.
  3. The named review trigger. Sustained high-cost use → a conversation, not a silent wall (Fair-Use Policy §2 row 2: blended model-cost run-rate ≳$60/mo × 2 consecutive months — [NUMBERS PENDING]).
  4. Agreed or temporary limits. Limits via the existing entitlement machinery (custom_limits/expires_at), with notice, content readable and exportable — data is never held hostage (QS Q17). (LEDGER/lib/billing.)
  5. Suspension for serious or repeated abuse — typically with notice; immediately for security emergencies (active exploitation, credential attacks, CSAM, credible harm).
  6. Termination for egregious or unlawful use. [LAWYER: process due — notice, appeal, statutory grounds; consumer vs B2B.]

An appeal path exists at every rung: reply to the notice and a human reviews. Payment status never determines whether you can export your content.

5 · Enforcement transparency & reporting

Fact sources for this document

QS Q13/Q18 (the abuse register a–g; swarm case; passcodes; web-research SSRF; memory) · QS Q17 (no data hostage) · ECON §5.2–5.3 (meter law, burst ceilings, review trigger) · ECON §6.1/§6.3 (free tier is real money; heavy use loses money) · CONN §0.3 ("at their request and consent"), CONN §11 (proof-of-care) · MCP ring-2 rule (humans sign consent widening) · LEDGER T024 (deletion-doesn't-restore) · PricingWing 2026-10-20 handoff (founder drain-cover words; 250-ever chronology note per Fair-Use Policy row 6).