Create a character and start talking
Choose a backstory, style, language and voice in the browser, then refine the result through text or voice conversations. Unlisted visibility is not the same as Private.
O03O04Convai
Start with a character idea, then bring a talking NPC into your game.
Evidence statusPublic-evidence reviewMedium confidenceLast checked Aug 31, 2026By MakeGameWithAI
PRODUCT OVERVIEW
What Convai is—and how it works.
Convai is an online character-creation and runtime service. Write a character's identity, background and speaking style, choose its language and voice, then talk to it in the browser. It generates responses to player input rather than simply helping you write a fixed script.
Unity and Unreal SDKs connect that character to a game, including input, speech and configured actions. You still supply the scene, avatar assets, rules and permitted behavior. A character can respond to game-provided state or request an action; it does not automatically know every quest, inventory item or save value.
This gives beginners a practical first step without building the speech and language stack themselves. Start with one guide, tavern character or quest helper, learn through conversations and play, then add the game logic you need. Coding and engine judgment improve control as the project grows; they are not qualifications for beginning in the Playground.
HOW YOU USE IT
Start with an account and a browser character. For game integration, install the matching SDK and use a small sample scene. Unity 4.5.0 requires Unity 6000.0.80f1 or newer; Unreal 4.0.0-beta.27 remains beta. AI-assisted configuration can help, but follow the current documentation rather than an older tutorial.
Give it knowledge, context and memory
Knowledge files describe the world; dynamic context carries game state and events; paid long-term memory links sessions to stable player and character IDs. These are different from an authoritative save system.
O02aO10O11Connect dialogue to bounded actions
Configure look, gesture, walk or follow behavior and guide narrative goals. Walking needs navigation setup, and quest changes still need game rules. Some controls currently point to Legacy Playground.
O03O09O15Use an engine route or a hosted experience
Use the SDK in your own game, or explore Convai's hosted Avatar Studio/Sim route. They share some service concepts but cloud-rendered minutes and cloud-experience concurrency are separate limits.
O02aO08O14O23INTERFACE & EXAMPLES
See how it works
The character editor separates the core description, speaking style, and embodiment settings.
Official source retrieved September 2026; interface version not specified.Source: Convai Documentation
REVIEW SCOPE
A first talking NPC, a small game prototype and understandable ongoing costs. Public cases show useful routes, not current-version guarantees for Chinese dialogue, latency, concurrency, narrative accuracy or a complete commercial game.
- Beginner friendlinessBeginner friendly
- Configure and talk to a character in the web playground without coding first; learn components, scenes and actions when moving into an engine.
- Agent integrationEasy to integrate
- The official character interaction API sends text and returns streamed replies; the verified beta route requires Professional or above.
Scope, integration requirements and sourcesChecked
Beginner scope: Create and test a first web character, not immediately ship a networked NPC game.
Agent scope: Send a conversation turn to a configured character and receive text responses.
Confidence: learning Medium · Agent Medium
- Convai Interaction APIOfficial · Beta
- Access
- Configure X-API-Key, character_id and an eligible plan; retain the session ID for continuity.
- Billing
- Budget against plan interaction allowances; a free playground does not establish free access to this API.
- Getting results
- Send text to the interaction endpoint and receive character replies over SSE.
- Limitations
- This assessment covers text interaction; voice, action execution, backend keys and full engine integration need their own interfaces.
These are independent editorial judgments, not an overall score or a claim of hands-on agent integration testing.
- Convai character creation · Official · Learning evidence · Integration evidence
- Convai Interaction API beta · Official · Integration evidence
KEY FINDINGS
What the evidence supports—and what it does not.
Editorial inferenceMedium confidence
You can begin before you can build the whole stack
The browser workflow lets you create and talk to a character before learning engine integration. SDK samples, configured actions and AI-assisted setup provide further entry points. The evidence supports trying a first interaction, not a promise that every beginner can ship without learning or help.
O03O04O09O18O23User reportsMedium confidence
Working projects and resolved problems both matter
A Unity/Quest creator reports a usable project with improved reliability. TUMSphere built a campus VR guide, while an iOS developer eventually resolved custom-build connection issues. These are different projects and conditions; the campus study used a PC-linked headset, not standalone Quest hardware.
U02U03U04Official factHigh confidence
One concurrent session does not mean one NPC
The limit applies to actively connected sessions sharing the deployed API key, not how many character definitions exist. Different players or characters can draw from the same pool. Creating extra NPCs does not create extra concurrency, and a second-character setup problem is not proof of a one-NPC product limit.
O02aO02bU07Editorial inferenceMedium confidence
There is no defensible universal latency number
One older Unreal report describes variable waits and a short outage that later recovered; a different Unity/Quest project in the same discussion reports improvement with a modified audio setup. The study uses another measurement definition. Version, voice, model, context and network differ, so these reports are not a common benchmark.
U01U02U04Editorial inferenceMedium confidence
Knowledge and narrative controls are not game-state authority
Knowledge, memory and narrative goals help give a character context, but the campus study still reports invented information. Treat generated actions as requests that the game validates. Inventory, rewards, payments and saved progress should not change merely because an NPC says they should.
O09O10O11O15O16U04EDITORIAL VERDICT
Our take
Medium confidenceConvai is worth considering when you want a character that responds to players before you know how to build an entire voice-and-language system. Its browser entry point and engine components make that idea approachable, and public projects show that it can go beyond a standalone chat window. Begin with one character and one useful interaction. Let game code control consequential actions, and budget for real players and simultaneous sessions before widening access.
Conclusion scopeOwner-approved, medium-confidence public-evidence review. Seven independent cases across five platforms support bounded workflow conclusions, not a success rate, current latency promise or fixed cost per accepted conversation.
BETTER FIT
Worth auditioning when
- Creators, including beginners, who want to start with a character description and refine it by talking and playing.
- Small narrative prototypes, guides, quest hints and VR interactions with a clear character role.
- Projects that can keep action permissions and quest state under game logic while allowing flexible dialogue.
- A limited audience whose monthly player count, active sessions and service budget can be planned.
POORER FIT
Do not depend on it yet when
- Requiring a completely offline NPC or no ongoing service dependency for the reviewed workflow.
- Needing every line and story branch to be deterministic without a scripted fallback.
- Expecting one prompt to deliver a complete game, secure networking and maintenance automatically.
- Opening unlimited public access before checking player, concurrency, data and spending conditions.
WORKFLOW FIT
Where it fits in a production workflow.
Shape one character through conversation
Give a guide or quest helper a short role and a few facts it should know. Try ordinary questions and questions outside that role; use the answers to refine the character before building a larger scene.
GuardrailBegin with non-sensitive material and verify visibility. File knowledge and long-term memory are not included in Free.
Add one visible action to a small scene
Use the matching SDK sample, connect one NPC and configure a gaze or gesture. Add walking, game events and narrative conditions only when the first loop is understandable and useful.
GuardrailA walking action needs navigation setup. Keep permissions and important state changes in game logic.
Open a bounded playtest, then estimate a launch
Choose a small audience and track distinct players, visits, active connections and voice/model use. Evaluate the target build, not only the editor, and use the measured workload to revise the budget.
GuardrailCap access and provide a fallback for unavailable dialogue. These are suggested checks, not tests completed by this site.
RECOMMENDED WORKFLOW
Use this tool inside a complete production workflow.
Start from a testable brief, then move through prototyping, assets, audio, testing, and release with an explicit handoff and human check at every step.
SHORTEST RESPONSIBLE PATH
Start with one character, one action and a small audience
This is a low-risk starting workflow synthesized from the evidence.
- 01
Create a non-sensitive Playground character, start with text, then try voice. Prepare a small set of questions including ones it should decline or admit it cannot answer.
- 02
Install the SDK for your engine version in a disposable project copy. Connect the sample character before adding custom scene components.
- 03
Add a look or gesture action and one game-state input. Validate quest completion and rewards in game logic, with a scripted fallback.
- 04
Build for the intended device and invite a limited group sequentially. Record actual quota changes and active sessions before proposing a larger player budget.
- 05
Before public access, check short-lived credentials, player notices, visibility, attribution and account caps. Do not turn a successful editor conversation into a performance guarantee.
PRICING & RIGHTS
Budget for conversations, players and active sessions
Pricing and terms last checked: Aug 31, 2026
- Monthly USD prices and interaction quotas: Free $0 / 100; Indie Dev $29 / 3,000; Professional $99 / 10,000; Scale $499 / 50,000; Business $1,199 / 125,000. Prices exclude tax.
- In the same plan order, monthly active end users are 1 / 75 / 250 / 1,250 / 2,500, and character-session concurrency is 1 / 1 / 3 / 15 / 30. MAU counts different people in the billing cycle; concurrency counts active connections, not NPC definitions.
- Annual upfront totals are $264 / $828 / $3,588 / $5,988 for Indie through Business, equivalent to $22 / $69 / $299 / $499 per month. Those are annual commitments, not monthly purchase prices. Quotas are monthly; unused interactions do not roll over.
- Indie includes 1,500 flagship-model interactions and 500 HQ third-party-voice interactions within its 3,000 total, not on top. Free has no file knowledge bank, long-term memory or included HQ-voice quota. Exhausting a subcap can restrict the selected model or voice while total interactions remain.
- Connecting your own ElevenLabs key changes which voice quota applies, but Convai's integration guide requires ElevenLabs Pro or above for compatibility. External voice costs remain separate; a key is not a free quota expansion.
- Avatar Studio/Sim cloud-rendered minutes are 60 / 600 / 2,400 / 10,200 / 21,000 per month, with cloud-experience concurrency of 1 / 1 / 1 / 5 / 10. Creation activity also uses these minutes. This is not a per-minute charge for rendering in your own Unity or Unreal game.
- The public v1.4 table gives Scale 1,250 MAU, while an older FAQ example says 200. The calculations use the table and flag the conflict; confirm the applicable limits before buying.
- Character Crafting API controls character configuration, not ordinary NPC conversations. Its absence on Indie does not mean Indie cannot connect to an engine. Hosted Experience embedding has separate domain restrictions.
COST TO OUTPUT
What the monthly fee can support
Illustrative monthly USD allocation, not observed billing: assume one ten-minute visit and two generated replies per minute, so 20 interactions per visit. Except the cloud example, use your own engine renderer with regular models/voices. Exclude retries, other use, tax, external services and labor. An interaction is an input plus generated response, not a minute or an accepted gameplay task.
O02O02aO02bO213,000 interactions, 75 MAU and one active character session. $29 ÷ 3,000 ≈ $0.0096667 per interaction as nominal allocation. You still pay the full $29; you cannot purchase only the fraction you use.
20 × $29 ÷ 3,000 ≈ $0.19. The total quota suggests 150 visits, not 150 different players: MAU is 75 and sessions run one at a time. All-flagship use covers at most 75 visits; included HQ voices cover 25, before other usage.
Assume 100 preparation interactions + 10 × 20 for sequential visitors. The actual monthly payment is still $29. Quota alone fits 10 such workflows, but one creator plus new groups of ten people fits only 7 complete groups within 75 MAU. The 100-interaction preparation budget is an assumption.
100 × 10 × 2 = 2,000 interactions. Indie has enough interactions but not enough MAU. Professional is the lowest published monthly plan meeting these conditions: $19.80 nominal allocation, but $99 cash paid, or $0.99 per person if these are the only 100 users.
The workload remains 2,000 interactions, but Professional permits only three active character sessions. Scale meets the stated cap: $19.96 nominal interaction allocation versus $499 paid, or $4.99 per person. This is plan matching, not a latency or throughput guarantee.
For the same 100 ten-minute visits, Professional's 2,400 cloud minutes cover visits alone; editing adds usage. Its cloud-experience concurrency is only one, not three. Three simultaneous cloud experiences require a plan such as Scale under the same assumptions.
Theory is not accepted output. At one, two or four replies per minute, the ten-minute Indie allocation is about $0.09667 / $0.19333 / $0.38667, while its MAU cap still applies. None of the listed monthly plans covers 100 active character sessions at once; confirm a custom arrangement rather than extrapolating a per-interaction price. Actual accepted-conversation cost must include real subscription spend, external services, hosting, iteration and operations, divided by conversations that meet your criteria; that result is unknown.
Commercial-use condition
The terms leave ownership of customer and generated content with the customer, but input rights, third-party voices/avatars and content restrictions still apply. Making the service available to others requires attribution unless a written or white-label agreement waives it; the table lists custom branding on yearly Scale/Business. Creator data-use terms and Enterprise agreements differ. Approval of this review does not certify every game, asset or data practice as permitted.
Unused interactions do not roll over, and payments are generally non-refundable under the terms. No universal free-retry, failed-request refund, interrupted-response metering or automatic overage price was established. Confirm checkout, applicable caps and upgrade terms; spending more does not by itself establish faster responses.
PRODUCTION RISKS
Resolve these before production use.
Engine support is specific to the release and target
Unity 4.5.0 starts at Unity 6000.0.80f1. The Unreal beta.27 matrix lists UE 5.0–5.8, with Win64/Android targets; its editor window needs 5.2+ and MCP tools need 5.8. Unity WebGL uses browser audio without Unity spatial audio, has documented lip-sync limitations, and needs microphone permissions. Older tutorials or issue reports do not establish current compatibility.
O06O08O12O13O14U05U06U07Player data and client keys need deliberate handling
Creator terms allow customer and generated content to be stored and used to improve services unless agreed otherwise in writing; customer content includes player speech, text and scene information. Unlisted is not private, and Business is not automatically an Enterprise contract. For Unity, use the documented backend/short-lived-token path rather than shipping an obfuscated account key; handle player notices and consent.
O03O17O19O20Memory and guardrails do not eliminate narrative mistakes
Stable identity helps memory connect sessions, and game-supplied context helps the character track events. Neither makes generated statements authoritative. Custom guardrails API is paid, and the case study still reports invented information despite a knowledge base. Verify actions and keep fallback dialogue for important paths.
O10O11O15O16U04A playable prototype still needs a service plan
Limits, outages, SDK changes and external voices can affect access. The 2024 Convai Connect announcement describes a player-account partnership route, not a verified default way to make developer costs disappear. Its current availability and terms remain unknown. Limit exposure and keep the game understandable when its character cannot respond.
O02bO06O12O21O22U01U03NOT VERIFIED
Claims this page does not make
- Current Unity 4.5.0 / Unreal beta.27 dialogue quality, Chinese performance, narrative compliance and action acceptance rates.
- Comparable end-to-end latency, failure rates, peak throughput or performance for a specific network, engine and device combination.
- Actual cost per accepted conversation, long-term player reception, iteration time and commercial-game operating bills.
- The final Scale MAU conflict, account-specific checkout/overage rules, current Connect availability and project-specific data retention or Enterprise terms.
RESEARCH METHOD
Research-reviewed from public evidence
Seven independent cases across five platforms: Convai Forum, Epic Developer Community, GitHub, Reddit and an original Frontiers project study. The three evidence types are community reports, a technical issue and an academic project. Convai Forum contributes 3/7 (42.9%) and Reddit 1/7 (14.3%). Two authors describe different projects in the same forum thread; they count as separate cases, not separate platforms. The study counts once, not once per participant. Four records have explicit dates within the preferred year, one shows only '3mo ago', and two are historical. Searches also covered GameDev.net, Unity Discussions, note.com, Zenn and G2; weak, incomplete and derivative material was excluded. Staff source S01 is grouped with official material. We retained positive counterexamples and resolved outcomes, did not average stars or incompatible latency figures, and separated official facts, user reports and budget assumptions. Current exact-version evidence remains limited; workflow advice is editorial synthesis.
Research windowPrimarily August 31, 2025–August 31, 2026. Earlier engine reports are historical; publication dates do not establish the SDK or time of use.
OFFICIAL SOURCES24
- O02Convai pricingChecked Aug 31, 2026
Current subscription entry point; the public table and FAQ are recorded separately below.
- O02aPublic pricing table — v1.4Checked Aug 31, 2026
Public pricing-page data, not a checkout: USD prices, monthly quotas, MAU, two concurrency limits and feature subcaps.
- O02bConvai billing FAQChecked Aug 31, 2026
Interaction and active-session definitions, no interaction rollover and flagship/HQ-voice limits; the Scale MAU example conflicts with the table.
- O03Creating a New CharacterChecked Aug 31, 2026
Browser character setup and visibility; some actions/narrative controls still point to Legacy Playground. Unlisted is not Private.
- O04Testing a CharacterChecked Aug 31, 2026
Documented text, voice and debugging entry points. Reading this guide is not a first-party product test.
- O05Install the Convai Unity SDKChecked Aug 31, 2026
Package Manager and Asset Store routes, dependencies and configuration prerequisites.
- O06Convai Unity SDK release notesChecked Aug 31, 2026
4.5.0 dated August 14, 2026, with Unity 6000.0.80f1 minimum and migration changes. Older 3.x reports are not tests of this release.
- O08Unity platform support matrixChecked Aug 31, 2026
Platform-specific paths and WebGL audio/microphone limits; documented support is not a compatibility test by this site.
- O09Character actions quick startChecked Aug 31, 2026
Configured look, gesture, walk and follow actions; locomotion needs targets and navigation setup.
- O10How long-term memory worksChecked Aug 31, 2026
Paid cross-session memory depends on stable user and character identity; it is not a complete save system.
- O11How dynamic context worksChecked Aug 31, 2026
Game-provided state and events inform the character. A context acknowledgement is not proof of a gameplay action completing.
- O12Convai Unreal plugin release notesChecked Aug 31, 2026
4.0.0-beta.27 dated August 11, 2026; beta and Blueprint changes remain material to upgrades.
- O13Supported Unreal Engine versionsChecked Aug 31, 2026
UE 5.0–5.8 runtime table; editor window requires 5.2+, while MCP editor tools require 5.8.
- O14Unreal platform support matrixChecked Aug 31, 2026
Win64 and Android are supported targets in this release; macOS, Linux and iOS are not in its supported set.
- O15Narrative DesignChecked Aug 31, 2026
Goals, sections, decision points and game-triggered events; not a guarantee of perfect story compliance.
- O16GuardrailsChecked Aug 31, 2026
Custom guardrails API is labeled paid; configuration does not establish zero hallucinations or universal safety.
- O17Ship a secure Unity buildChecked Aug 31, 2026
Developer-backend and short-lived AuthToken route; obfuscating a shipped account key is not secrecy.
- O18AI coding assistant quick startChecked Aug 31, 2026
SDK-aware configuration tools can assist scene setup; this is not evidence of a beginner success rate.
- O19Convai Terms of ServiceChecked Aug 31, 2026
Updated September 1, 2025: ownership, attribution, Creator data use, Enterprise agreements and payment conditions.
- O20Convai Privacy PolicyChecked Aug 31, 2026
Page dated February 21, 2023; data categories, service providers and US processing. No blanket runtime-data deletion guarantee.
- O21ElevenLabs API integrationChecked Aug 31, 2026
All Convai tiers may connect a key, but the current guide requires ElevenLabs Pro or above for 44.1 kHz compatibility.
- O22Convai Connect announcementChecked Aug 31, 2026
March 18, 2024 announcement of a player-account partnership route. Current availability, SDK support and revenue terms were not established.
- O23Convai Unity SDK getting startedChecked Aug 31, 2026
Character, player, room, input and sample-scene setup; a supported route rather than an audited game build.
- S01UI customization and Convai branding — staff clarificationChecked Aug 31, 2026
April 2026 staff guidance about interface changes and retaining the logo; treated as vendor evidence, not an independent user vote.
INDEPENDENT EVIDENCE RECORDS7
- U01Ham1ton — latency troubleshooting across charactersChecked Aug 31, 2026
March–April 2026, Unreal 3.6.8-beta.3, Business/Free comparisons. Reports variable waits and an outage that recovered that day; not a current-version benchmark.
- U02Tyke — Unity/Quest project counterexampleChecked Aug 31, 2026
A separate project in U01's thread reports improved reliability with a modified audio component and specific voice/model. SDK version and commercial ties are unstated.
- U03iOS connection and Cloud Build — resolved follow-upChecked Aug 31, 2026
Unity 4.0 to 4.1.0: samples worked but a custom TestFlight build failed. The author reports resolution on page 2 after preservation/audio-session changes; one case, not several votes.
- U04Next-Gen orientation: supporting international students with generative AI NPCs in VRChecked Aug 31, 2026
TUMSphere's 24-participant campus guide used UE5, Convai and a Quest 3 linked to a high-end PC. One project, not 24 independent records; SDK/date of collection are unclear and hallucinations remain a limitation.
- U05MetaHuman VR gaze conditions — Epic developer discussionChecked Aug 31, 2026
January 2025 historical report about location, gaze and button-dependent behavior; no confirmed resolution or exact SDK. Not proof of a current gaze defect.
- U06Quest 3 audio-capture crash — Unreal SDK issue #28Checked Aug 31, 2026
March 2025, UE 5.5.4 and Quest 3 in the legacy repository. The team reports a workaround; an open issue is not proof that beta.27 still fails.
- U07Adding a second NPC — Unity3D discussionChecked Aug 31, 2026
One character worked; adding another raised configuration problems and an AI-suggested setting could not be found. Exact date/SDK/plan and resolution are unknown; page showed '3mo ago' when checked.
DISCLOSURE
How this research was supported
Owner Review passed on August 31, 2026, accepting this bounded public-evidence review. This work introduced no sponsorship, affiliate arrangement or vendor-provided credits, and involved no paid generations or engine tests. The owner's previous use and commercial relationships remain unstated; approval does not resolve the listed pricing conflict or unknown runtime performance.
This public research review will be revisited when pricing, terms, product versions, or material new evidence changes.