Model adapters
Keep your model and your keys. Humane works around them.
Humane talks to models through four adapters: OpenAI-compatible, Anthropic, Amazon Bedrock and Replicate. Change providers without touching the behavioural layer.
Model adapters
Humane
state · gates · timing · memory
OpenAI-compatible
Any OpenAI-style endpoint
Anthropic
Messages API
Amazon Bedrock
Converse API
Replicate
Hosted open models
Your keys · your model · base URLs SSRF-checked
Four adapters, many providers
OpenAI-compatible
Any provider with an OpenAI-style API. The catalogue has presets for services such as Azure OpenAI, Groq, Together and Mistral, and for a local Ollama server.
Anthropic
A native adapter for Anthropic's API, with streaming.
Amazon Bedrock
Uses the Converse API, so one request shape covers the models Bedrock hosts. Credentials from your environment or an instance role.
Replicate
Run hosted open models through Replicate with the same Humane layer on top.
Wrap your client, or let Humane call the model
wrap() puts Humane in front of the OpenAI-style client you already use, and on BLOCK it never calls the model. process() goes further: Humane builds the prompt, calls your configured provider and checks the reply before you see it.
- Per-tenant base URL and key, checked against SSRF rules
- Provider errors cleaned before they reach your users
- Automatic retry on 429 and 503 responses
- Streaming on providers that support it
Playground
Raw model
Great question! There are several things you can try. First, let’s go over the options in detail…
With Humane context
I’m sorry you’ve had to ask three times. Here’s exactly where it stands.
tone warm_grounded · length short
Wrap the client you have
import os
from openai import OpenAI
from humane_ai import wrap
# One line: every chat.completions.create() now goes through Humane first
client = wrap(OpenAI(), humane_key=os.environ["HUMANE_API_KEY"])
resp = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "I'm really struggling with this"}],
user="user_42", # a stable id for this end user
)
print(resp.choices[0].message.content)
print(resp.humane.safety.action) # "PROCEED" | "HOLD" | "BLOCK"
print(resp.humane.user.mood) # this user's mood, 0.0 to 1.0
print(resp.humane.analysis.reasons) # the phrases behind the readingModel questions
Something else? Ask us.
Works well with
Streaming
stream() returns server-sent events: a metadata event with the user's state, safety decision and timing before the first token, then the tokens, then done. If a gate blocks, the stream says so and ends.
ExploreTone and context
Before your model runs, Humane assembles its prompt from the user's state, relationship, recent conversation, relevant memories and open tasks, scoring and trimming each section to fit a budget. It also sets the tone, empathy, formality and length the reply should take.
ExplorePlayground
The playground runs Humane's analysis on a message and shows what it detected and how it would shape the reply. Point it at a model and compare the raw reply with the Humane-shaped one side by side.
ExploreEvery AI should know who it's talking to.
The safety and human-oversight layer for emotional AI: user state across sessions, risk alerts, safety gates and human takeover, with every decision explained.
