Humalike
Behavioral infrastructure for AI agents. Seven composable HTTP APIs give an existing agent the social skills it lacks: turn-taking, group norms, persona, theory of mind, social memory, social signals and observability. Model-agnostic, built for one-to-one and for groups.
What is Humalike?
Humalike is a behavioural layer that sits above an existing AI agent rather than replacing it. Published by Soofte, Inc. — a Delaware corporation whose team works out of Spain — it packages seven social competences as composable HTTP APIs. Turn-Taking is the flagship: it decides when the agent speaks, waits or interrupts, and bundles the others underneath. Norms picks up a group's tone and unwritten rules. Persona turns a thin bot description into a character grounded in real community data. Theory of Mind tracks what each participant believes, wants and feels. Social Memory remembers people across sessions. Social Signals reads typing pauses, edits and removed reactions. Social Observability reports who was engaged, bored or annoyed.
The suite is deliberately model-, use-case- and stack-agnostic, and it treats multi-party rooms as a first-class case rather than an afterthought: the same endpoints serve one person, two or ten, in voice or in text. Everything shares one base URL, one bearer-token scheme, one error envelope and one credit balance, so the learning surface stays small. Official examples ship in cURL, Python and TypeScript.
Seven scenarios are put forward: game NPCs and teammates, AI coworkers that join internal channels, mental-health agents that catch distress signals, hardware companions, ed-tech study partners, domestic humanoids, and community bots for creators. A separate product, AI NPC for FiveM, drops persistent voice-enabled characters into a GTA roleplay server, metered in its own turns rather than credits. An official plugin covers the Hermes Agent framework, and the founders publish their own research on arXiv, including a benchmark for local social norm adaptation.
Maturity is the caveat. The homepage still badges the behavioural APIs Coming soon and use-case pages invite early-access requests, even though documentation, pricing and billing are live. SOC 2 Type II and ISO 27001 are both described as in progress, and the security page states plainly that nothing is certified yet.
What it does
- Decide when an agent should speak, when it should wait, and when interrupting is the right move
- Read a group's tone, unwritten rules and inside jokes, then answer the way that group accepts
- Track what each participant believes, wants and feels as the conversation moves
- Remember individual people, their opinions and their behaviour across separate conversations
- Build a persona grounded in real community data instead of a generic backstory
- Catch the signals most agents miss: a typing pause before sending, an edited message, a removed reaction
- Report how the agent actually landed — who was engaged, who was bored, who was annoyed
When to use Humalike / When not to
A quick filter to help you decide if Humalike is the right fit.
When to use Humalike
- Engineering teams that already run a conversational agent and need it to behave properly in a live group rather than answer faster
- Game studios building NPCs, teammates or opponents that remember a player from one session to the next
- Makers of hardware companions and domestic humanoids — pendants, speakers, glasses — that must read a room before speaking
- Digital-health teams shipping mental-health support agents that have to catch distress signals and know when to escalate to a human
- Creators and community operators running bots that keep a Discord or a stream chat alive instead of merely moderating it
When not to use Humalike
- Consumers and anyone with a non-commercial project: the privacy policy and the terms both restrict the service to B2B use and ask others not to register
- Anyone under 18, or under the age of majority in their jurisdiction
- Teams looking for a language model or a complete voice stack: Humalike sits above one and expects you to bring your own
- Non-technical buyers who want a finished agent: everything runs through HTTP calls and a bearer token
- Organisations whose procurement requires a completed SOC 2 or ISO 27001 audit today, since the security page states plainly that nothing is certified yet
How to use Humalike
A typical end-to-end flow, from setup to results.
- Create an account from the sign-in page; the customer console authenticates through Clerk
- Obtain a bearer token from the Authentication page and export it as an environment variable
- Point every request at the single base URL, https://api.humalike.com
- Make one call end to end — the quickstart uses POST /v1/social-learning/actions/extract with a transcript of at least one message
- Read 'When to call each API' to work out which component belongs where in your loop
- Wire Turn-Taking as the spine: open a thread, submit messages, ask for a respond decision, record events back
- Add Social Memory through its three endpoints — ingest, recall, ask — to carry people across sessions
- Layer Persona, Theory of Mind, Norms, Social Signals and Social Observability as your use case requires
- Track spend with POST /v1/credits/projections/usage-summary, which returns 30-day totals and a 7-day daily series
- Alternatively skip the integration work: install the Hermes Agent plugin, or drop the humalike resource onto a FiveM server
Pros & Cons
Pros
- Composable primitives rather than a monolithic stack — take the two components you need and leave the rest
- Genuinely stack-agnostic: it layers over whatever model and voice provider you already run, and the comparison pages suggest plugging a rival in underneath
- Multi-party conversation is a first-class case, and relational memory is per person across sessions rather than mere chat resumption
- One base URL, one authentication scheme, one error envelope and one credit balance keep the learning curve short
- Billing is legible: the price is computed and capped before any work starts, failed calls are free, and a 402 is a hard gate rather than a surprise invoice
- Unusually candid security disclosure for a young vendor — thirteen named sub-processors with purpose, data and transfer mechanism, a dated changelog, and a written statement that no certification has been obtained
- Platform data hosted on AWS in the EU, with published research and a public benchmark behind the product claims
Cons
- Maturity is ambiguous: the homepage still badges the behavioural APIs 'Coming soon' and use-case pages ask you to request early access, while docs and billing are live
- The pricing page is not linked from any menu or footer — it exists only if you guess the URL — and two navigation links, /app and /ai-npc, return 404
- No postal address is published anywhere, in the United States or in Europe, and no support email exists: support runs through Discord and a form
- Three different billing units appear across the site — interactions on the pricing page, credits in the documentation, turns for the NPC product
- Nothing is said about whether customer content trains models, in either direction, and no opt-out is documented
- Neither SOC 2 Type II nor ISO 27001 has been obtained, and even the GDPR badge on the homepage is qualified 'In progress'
- The terms are dated January 2025, still call the platform 'humalike.com', reserve the right to change without notice, and disclaim any availability guarantee
Pricing & Plans
A permanent free plan is available. The Developer tier costs USD 0.00 per month and includes 10,000 interactions per month, the basic social engine and community support. The lowest paid entry point is the Pro tier at USD 0.002 per interaction, with unlimited volume and no monthly commitment; an Enterprise tier is quoted on request. In practice billing is metered in credits against a single account balance, each request being priced from its inputs and capped at that price before it runs. Every new account receives USD 20 in free API credits without a payment card. Payments are handled by Stripe.
- 10
- 000 interactions per month
- Basic Social Engine
- Community Support
- unlimited interactions
- Advanced Timing & Interruptions
- Persistent Memory
- Priority Email Support
- volume discounts
- custom SLA
- dedicated solutions engineer
- on-premise options
Data, GDPR & hosting
A consolidated view of how Humalike handles your data.
GDPR overview
Implementation is documented rather than merely claimed. The privacy policy names Soofte, Inc. as controller, states alignment with the GDPR, and lists the full set of data subject rights, including erasure, portability and complaint to a national authority. Transfers outside the EEA rely on Standard Contractual Clauses under Article 46, supplemented by the EU-US Data Privacy Framework for several vendors, with transfer impact assessments claimed. A public sub-processor list is maintained under Article 28(2), with ten days' notice before any addition and a right to object. Two gaps stand out: no Article 27 EU representative and no data protection officer is named anywhere, and the homepage badge itself reads GDPR — In progress. Privacy contact is privacy@humalike.ai, while the policy routes rights requests to mtk@soofte.com.
Who owns the data?
Customer Data — everything you submit, including the output the service generates from your inputs — stays yours. The terms carve out a non-exclusive, royalty-free, worldwide licence for Humalike to use and display it, but only as far as delivering the service requires, and Humalike reserves a right to review content in order to curb abuse or breaches of its ethics policy. Soofte, Inc. acts as processor or sub-processor of that data, and treats it as confidential, disclosing it only to staff, affiliates and vendors bound by confidentiality. Two clauses cut the other way: feedback you send is assigned outright to Humalike, and your name may be used as a customer reference.
Reuse rights
You keep the right to use your own data as you see fit; nothing in the terms limits what you do with your Customer Data or with the output the APIs return. Humalike's own use is bounded by the service: conversation and prompt content travels to third-party inference providers (Google Gemini, Anthropic), voice audio to Deepgram, reply text to ElevenLabs, persona briefs to Parallel.ai. Aggregate usage and performance metrics may be processed to improve the platform, and the terms state these exclude the content you submit. Website data is processed on named legal bases — contract, legitimate interests, consent, legal obligation — with no automated decision-making or profiling. Whether customer content ever trains a model is not addressed anywhere on the site.
Data retention & training
Hosting summary
The platform runs on Amazon Web Services hosted in Europe, using managed services such as RDS, S3 and VPC network isolation, and the security page states that platform data is hosted in the EU. Traffic is encrypted in transit with TLS and data is encrypted at rest with AWS-managed AES-256. Thirteen sub-processors are published with their location: AWS and Google span EU and US, while Anthropic, Deepgram, ElevenLabs, Parallel.ai, Clerk, Stripe, Resend, PostHog, Vercel, Better Stack and Discord process in the United States. Transfers rely on Standard Contractual Clauses, with the EU-US Data Privacy Framework additionally covering AWS, Google, Clerk and Stripe. Customer support is delivered from Spain. Production access follows least privilege, identities sit in Google Workspace with mandatory multi-factor authentication, and secrets live in dedicated tooling rather than in code. Infrastructure is monitored continuously and availability is published on a public status page.
Where Humalike works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Humalike.
- Nothing on the site states whether customer content is used to train models, and no opt-out is documented — a blank worth clarifying in writing before sending sensitive conversations
- SOC 2 Type II, ISO 27001 and even the GDPR badge are all shown as in progress; the security page confirms no certification has been obtained
- Conversation and prompt content leaves for third-party inference providers, voice audio for a transcription vendor, and reply text for a synthesis vendor, several of them in the United States
- The terms can be amended at any time without notice, and continued use counts as acceptance; there is no availability guarantee and liability is capped at twelve months of fees
- By default you authorise Humalike to name you as a customer and publish your use case, and any feedback you send is assigned outright without attribution or payment
- Delaware law and exclusive Delaware jurisdiction apply even to a European customer, and no Article 27 EU representative is designated despite the EEA-facing product
- The product exists to make agents pass as socially competent — the founders' own paper studies a deliberately human-seeming facilitator — so disclosure to end users deserves deliberate thought rather than default silence
Setup & Integrations
Technical difficulty
Moderate, and squarely for developers. Integration is pure API work: obtain a bearer token, send JSON over HTTPS to a single base URL, and read one error envelope. The quickstart completes a first call end to end, and official examples cover cURL, Python and TypeScript. No official SDK is advertised on the pages reviewed. Two shortcuts exist for teams that would rather not write the integration: the Hermes Agent plugin, installed from GitHub, and the FiveM resource, which runs standalone. The real prerequisite is upstream — you must already have an agent and a model, because Humalike supplies neither.
Deployment
Integrations
Behind Humalike
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Humalike.
Frequently asked questions
What exactly does Humalike sell?
Do I have to change my model or my voice provider to use it?
What does it cost?
Can I try it without a payment card?
How is a call priced, and what happens if I run out?
Where is my data hosted?
Is customer data used to train models?
Is Humalike certified SOC 2 or ISO 27001?
Who can use the service, and from what age?
How do I reach support?
Should you pick Humalike?
Humalike stakes out a narrow and genuinely unusual position: it treats an agent's social behaviour as separate infrastructure, the way authentication or payments became separate infrastructure, and sells it as seven primitives you compose rather than a platform you adopt. The execution matches the ambition. The APIs share one base URL, one token scheme and one credit balance; billing is priced and capped before a call runs and costs nothing when a call fails; the documentation is complete; the founders publish their own research and a public benchmark. The security disclosure is better than most of the segment — thirteen sub-processors named with purpose, data and transfer mechanism, a dated changelog, and a plain admission that no certification has been obtained.
Against that, this is a very young company. The domain dates from August 2025, the first web archive capture from March 2026, no funding round is public, and the site still carries the marks of a launch in progress: a Coming soon badge on the behavioural APIs, early-access invitations on the use-case pages, a pricing page linked from nowhere, and two dead navigation links. The gaps that matter for a purchase decision are informational rather than technical — no postal address, no support email, no stated position on training models with customer content, no completed audit, and terms still dated January 2025 that name a different domain.
The free tier and the twenty dollars of credits make a prototype essentially costless, and that is the right way to test this: build one agent, watch how turn-taking and social memory change the feel of a group conversation, and judge from there. For a regulated production deployment, the open questions deserve written answers first.
- Choosing a selection results in a full page refresh.
- Opens in a new window.