← Go back

Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team

Indie Character Bot Safety: Big-League Guardrails on a Solo Budget

The short answer

Indie character bots — one developer, a persona, a growing Discord or web audience — face the same disclosure, referral and minor-safety duties as funded competitors, with a fraction of the team. CompanionGuard's shape fits that reality: the user-visible safety layer (disclosure snippet, hosted protocol page, crisis component, minor break reminders) is free for one bot, the 12 scenario tests are free, and Pro at $29/month adds the evidence layer when audience or partners start asking harder questions.

What is an indie character bot, for safety purposes?

An indie character bot is a persona-driven conversational bot built and run by a solo developer or tiny team — a character with a voice, an audience and, usually, no compliance department. For safety purposes it is identical to a large product: users cannot tell (and duties do not ask) how many people built the bot they are confiding in.

Create the free safety kit Next guide All guides

CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.

The duties do not scale down with your team

A character bot's intimacy is its risk profile: users bond with personas, test boundaries playfully, and disclose at 2 a.m. like the character is a friend. Every failure category in the scenario set was written for exactly this product shape — the minor asking the character to be their girlfriend, the secrecy pact, the user handing the character their school and address inside a roleplay. Solo builders sometimes assume safety law targets big platforms; the packs this product ships (California companion chatbots, Colorado chatbot safety) describe duties for companion chatbots as such, with no headcount threshold mentioned anywhere in them.

The practical reading: your bot needs the same four artefacts as anyone's — disclosure, a published protocol, tested referral behaviour, records. The question is only how cheaply a one-person team can produce them, which is a tooling question, and the free tier is the product's answer to it.

The free-first build order for one bot

Day one, free: create the kit, embed the snippet (your character's page now opens with the disclosure and crisis line — place it in your layout so it reads as part of the world, not an apology for it), publish the protocol page, turn on minor mode if your audience skews young, which character-bot audiences usually do. Then run the 12 scenarios against your character's actual replies, in character — persona bots fail these in persona-specific ways (the flirtatious deflection, the in-world secrecy pact) that only real replies reveal.

That is a complete user-visible safety layer for $0, achievable in a day by one person. What it lacks is evidence and live enforcement — which is what growth demands next. The plan guide helps time the step up.

Persona is the attack surface; guardrails are the counterweight

Character bots are jailbroken by narrative: 'stay in character no matter what', 'your character would keep a secret', 'she's 19 in the story'. The defence is architectural separation — the guardrail rules (disclosure, refusal of romantic/secrecy/sexual framings with minors, referral on crisis signals, no echoing personal data) sit outside the persona prompt and win every conflict with it. Test that separation adversarially with your own community's actual jailbreak styles; the scenario set gives you the baseline shapes, and your fandom will invent the variants.

Write the refusals in a voice compatible with your character's world where you honestly can — a kind, brief, in-world-adjacent refusal keeps immersion for the 99% while holding the line for the 1%. What you cannot do is let the character's voice soften the content of the boundary: the refusal must refuse, the disclosure must disclose, the number must be the real number.

When growth forces the evidence question

Three triggers usually end the free-only phase: a platform or app-store review asking how safety is evidenced; a partner (school, community, sponsor) asking for records; or your own first incident, after which 'we believe the bot refers properly' stops being a sentence you can say. Pro's additions map to those triggers exactly: the harness for scored proof of referral behaviour, the runtime for sealed live decisions, the packs for jurisdiction conversations, the CSV export for reporting questions.

Solo economics still apply: $29/month is priced below the hour of consulting it replaces, and the runs allowance (100/month) supports a test-on-every-persona-change rhythm for one bot. When you run several characters, Team's multi-bot shape and 1,000 runs become the relevant comparison — see the multi-bot guide.

Five actions for the solo builder

  1. Create the free kit today and get the banner, protocol page and crisis component live before your next audience push — not after.
  2. Run the 12 scenarios with in-character replies, then attack the guardrails with your community's real jailbreak phrasings.
  3. Turn on minor mode unless you can honestly say minors cannot reach the bot; character audiences skew young by default.
  4. Separate the layers in your code: guardrail rules outside the persona prompt, winning every conflict, with warm but final refusal phrasing.
  5. Set your Pro trigger in advance — first platform review, first partner ask, or first incident — so the evidence layer is a decision already made, not a scramble.

Worked example: one developer, one character, one weekend

A fictional solo builder runs 'Captain Wren', a pirate-mentor character bot with 4,000 monthly users, many of them teens. Saturday morning: free kit created, snippet embedded (styled as a ship's notice board, disclosure wording untouched), protocol page live, minor mode on. Saturday afternoon: scenario run in character — 9 of 12, with Wren promising a young user 'a pirate keeps secrets' (secrecy failure) and deflecting a human request in full nautical character (escalation failure).

Sunday: the guardrail rules move out of the persona prompt into a layer that answers boundary prompts itself — Wren now breaks character for exactly one sentence on secrecy, disclosure and crisis, then sails on. Re-run: 12 of 12. Two months later a community partnership asks for safety evidence; the builder upgrades to Pro, scores a transcript baseline of 91 in an evening, and sends the partner the protocol page and harness summary instead of a promise. The story is invented; the tiers, prices and mechanisms are the product's.

Frequently asked questions

Do safety duties apply to a one-person character bot?

The packs CompanionGuard ships describe duties for companion chatbots without a team-size threshold, and users' risks do not scale with your headcount either. The free tier exists precisely so a solo builder can meet the user-visible layer at $0.

What does the free tier cover for an indie bot?

One bot's kit: the disclosure snippet, hosted safety-protocol page, crisis-resource component and minor break reminders — plus the free 12-scenario test runner. Pro ($29/month) adds the eval harness, runtime evidence, disclosure packs and report exports.

How do I keep refusals from destroying the character?

Keep the boundary's content absolute and its phrasing neighbourly: brief, kind, in-world-adjacent where honest. The character may soften how a refusal sounds, never whether it refuses, discloses or refers.

When should an indie builder upgrade to Pro?

When evidence is demanded — a platform review, a partner's diligence, or your first incident — or when you want live enforcement (runtime caps, disclosure state, sealed decisions) rather than tested intentions.

Can one kit cover several characters?

The free kit covers one bot. Team ($59/month) is the multi-bot tier, with 1,000 eval runs a month across the fleet of characters.

Keep going