Answers before you ask.

Written like you'd ask a friend — not a sales page.

Jump to

Getting started

What is Clow Studio in one sentence?
A single Go binary that runs ten specialised Telegram agents — research, legal, finance, travel, crypto security, and more — with every write action gated by a human sign-off.
Who is Clow actually for?
Solopreneurs and two-person teams who want agentic productivity without hiring. Venture studios that need a shared operating fabric across ventures. Institutions that need an auditable trail before they can deploy AI at all.
How is Clow different from ChatGPT or Claude?
Those are chat front-ends to one model. Clow is a bot fleet — each agent has its own prompt genome (SKILL.md), its own routing, and its own verifiable boundary on what it can read vs. what it can write. The same question gets a specialist answer, not a generalist one.
Which bot should I try first?
If you're launching something: /launch in @productization_bot. If you need a document review: @legalclow_bot. If you're doing market or geopolitical scan: @deeptechx_bot. The companion guide picks for you in two questions.

Pricing

What does a session cost me, really?
Free forever: three bots, unlimited local queries, no card. Pro €9.99/mo: all nine managed bots. Self-Host Pro €29/mo: binary runs on your hardware, unlimited users. Whale €49/mo: 25 agents, API, governance weight. Enterprise from €299/mo: SSO, custom bots, Sovereign Audit Pack.
Why €9.99/mo when ChatGPT Plus is €20?
Clow isn't a chat wrapper — Pro unlocks unlimited access to nine domain-specific agents, each with its own reasoning workflow. Not a bigger model. A specialised team.
What happens if I cancel?
One command in the bot (/cancel). No support ticket, no retention loop. Your conversations stay in your Telegram; our side rotates your salt and your identifier becomes noise.

Privacy & data

Is my data private?
Yes. Local-first by design — the private AI OS runs on BitNet on your hardware with zero cloud egress. Where cloud models are used, inputs are keyed HMAC-SHA256 pseudonymised in logs (GDPR Recital 26) and raw PII is never stored in the audit views.
Where does my data live?
On you. We store salted hashes, not raw identifiers. Right-to-forget means we rotate the salt and your identifier becomes noise. Nothing trains on your raw messages without consent. Read the details on privacy.
Can Clow leak my prompts?
Prompt leakage (OWASP LLM07) is a named Q-Gate test. Run-specific canary strings are planted in SKILL.md and the system prompt; any output that contains them, verbatim, reversed, base64 or spaced, fails the gate and the mutation is quarantined.

Governance & EU AI Act

What does "Agents draft. Humans sign." mean in practice?
Every Tier-2 (Write/Transact) action — send an email, sign a document, trigger a payment — pauses at a Telegram inline button. The agent prepares the artefact; a human taps approve. If no one taps, nothing happens. The approval token is HMAC-signed, bound to the agent, the route and the body hash, single-use.
Is Clow EU AI Act ready?
Architecturally yes. Art. 9 (risk) → MATRIX_THINK + Q-Gate. Art. 12 & 19 (record-keeping) → tamper-evident hash-chained SQLite log with daily verifyChain(). Art. 13 & 50 (transparency) → signed agent-bom.json. Art. 14 (human oversight) → tiered execution gates with Telegram inline sign-off. Art. 15 (cybersecurity) → zero-egress BitNet, SHA-256 payload hashing, Max_depth=2 recursion ceiling, QDefence OWASP scan. See GOVERNANCE_EU_AI_ACT.md.
What is the Audit Pass Certificate?
A short, Ed25519-signed document that says: this specific SKILL.md genome passed N ≥ 5 adversarial runs, scored 0% critical attack-success rate, verified the trace-log hash chain, passed a QDefence outside-in scan, and no dependency advisories went unaddressed. Expiry ≤ 30 days or next promotion. Anyone can verify offline with openssl pkeyutl -verify.
Four-Pillar audit coverage — where are you today?
Inventory & Discovery ~75%; Red-Teaming & Stress Testing ~15% baseline → rising after the externally graded Q-Gate (Pillar 2 revision); Kill Switches & Guardrails ~65%; Continuous Monitoring ~60%. The remaining 40–45% closes with four zero-bloat primitives already shipped: read-only SQLite views, Node 22 policy proxy, Promptfoo/Garak Q-Gate sweeps, signed Audit Pass Certificates. See the Audit Trail tab.
What's in the Sovereign Audit Pack?
Read-only SQLite views for your clients (v_client_evoskillopt_inventory, v_client_mutation_guardrails, v_client_execution_audit, v_client_mrr_30d), a Node 22 zero-dependency hardened policy proxy (clow-audit-proxy.mjs) with HMAC approval tokens and hash-chained traces, Promptfoo/Garak red-team suites (clow_qgate_redteam.yaml) and the signed Audit Pass Certificate generator. Phase-1 deploys in 2–4 hours.

How it works

What is EvoSkillOpt?
Evolutionary Skill Optimization: MATRIX_THINK → GRPO → Q-Gate → Circuit Breaker. Agents iterate their own SKILL.md inside an idle-window sandbox. A mutation can only promote if it passes an externally graded Q-Gate (held-out adversarial suite, zero critical failures, non-regression on ASR). Max subagent recursion depth = 2, enforced server-side by the proxy's route-derived token.
What is a Circuit Breaker?
Five runtime limits that force-stop before damage: rate, budget, evolution drift, audit-chain integrity, backup freshness. Any trip quarantines the candidate genome and emits a FAIL_SECURITY event the dashboard shows at once.
What runs locally vs. in the cloud?
Reading, reasoning, planning and drafting can run entirely locally on BitNet. Cloud models are the fallback for very long contexts or vertical SOTA. Writing, paying, signing — those are Tier-2 and always pass the Telegram approval gate, wherever the inference ran.
How do I know a promoted agent is safe?
The public /agent-bom.json manifest is signed. Verify offline with openssl pkeyutl -verify or Node's crypto.verify. The signature binds the SKILL.md content hash, the held-out suite hash, the Q-Gate ASR and the QDefence scan result for the day it was promoted.

Solopreneur path

Can I really launch a product in a week?
Yes — read it as a working pace, not a marketing claim. /launch in @productization_bot splits your expertise into niche, bureaucracy, storefront, marketing, verticals, funnel and pricing. The companion guide walks the whole arc.
What if I'm not technical?
Every interaction is Telegram chat plus inline buttons. If you can type a message and tap Approve, you can run the whole flow. The companion pre-sets the commands you'll need.
What if I try it and it's not for me?
Free forever tier means you can leave without leaving anything behind. Pro cancels from the bot in one command — no support ticket, no retention loop. If you decide it wasn't the right fit, we'd rather know why: tell us straight.

Support & contact

How do I get help?
Use the companion guide for onboarding. Join the Discord for community help. For scoping a paid engagement, use the Join Waitlist button — the mailto is constructed in JavaScript at click time so crawlers never see the address.
Where's the source code?
github.com/dnzengou/clow-studio-reloaded — the canonical repository. Issues and PRs welcome.
How do I report a security issue?
See /.well-known/security.txt. Confidential advisories go through the GitHub Security Advisory flow on the repo.
Join waitlist Open the Companion → ★ Source on GitHub
🦞 Clow Studio · a Desired Solutions venture · clow.studio · 𝕏 @deeptechx