roboto-sai-sdk

Python SDK for the Roboto SAI platform — the canonical client and source of truth for public model names, runtime profiles, and the frozen runtime contract.

Install

pip install roboto-sai-sdk

Requires Python 3.10+.

What pip install roboto-sai-sdk pulls in (0.7.0+)

Since 0.7.0 the core install is the full roboto-sai.com backend runtime including its heavy ML stack and PyTorch, so the site needs nothing but pip install roboto-sai-sdk — there are no [site]/[server] extras. That is 48 direct requirements and ~190–210 installed distributions. Measured site-packages on Linux x86_64:

CPU-only install (servers and CI without a GPU) — install torch from the PyTorch CPU index first; pip then treats torch==2.14.1 as satisfied and skips the CUDA build:

pip install torch==2.14.1 --index-url https://download.pytorch.org/whl/cpu
pip install roboto-sai-sdk

Warning: a Render 512 MB Starter instance previously ran out of memory (OOM-killed) with this heavy stack installed. Use a larger instance, or keep heavy imports lazy.

• Web/ASGI: fastapi, uvicorn, uvloop (not on Windows), httptools, pydantic[email], pydantic-settings, python-multipart, python-dotenv, websockets
• HTTP: httpx (the only one the SDK code itself imports), requests, aiohttp, urllib3
• Storage/queues: sqlalchemy, aiosqlite, redis, aioredis, celery, kombu
• Security/content: cryptography, PyJWT, bcrypt, bleach, itsdangerous, defusedxml, mistune, pygments, pillow
• Ops: psutil, structlog, slowapi, prometheus-client, opentelemetry-distro, opentelemetry-instrumentation-fastapi
• Integrations (exact pins): mcp==2.2.0, supabase==2.31.0, plus boto3
• ML: torch==2.14.1, numpy (<3), transformers, sentence-transformers, accelerate, onnxruntime, chromadb, networkx, rank-bm25, langchain-community, matplotlib

On Python 3.10, sqlalchemy, onnxruntime, networkx and matplotlib resolve to their newest 3.10-capable lines (the site's floors need 3.11+). Not included: xai-sdk (opt in with pip install "roboto-sai-sdk[xai]"). The machine-readable list is roboto_sai_sdk/contracts/dependency_contract.json. Need only the lightweight client? Pin roboto-sai-sdk<0.7.

Quick Start

from roboto_sai_sdk import RobotoSAI

client = RobotoSAI()  # uses SAI_ACCESS_TOKEN from env
result = client.send_message("Review this repository and propose the next refactor.")

print(result["model"])      # "roboto-sai"
print(result["subtitle"])   # "Core runtime"
print(result["text"])       # assistant reply

Stream mode:

from roboto_sai_sdk import RobotoSAI, extract_text_delta

client = RobotoSAI(model="roboto-sai")
for event in client.stream("List five things to test in a new SDK release."):
    print(extract_text_delta(event), end="", flush=True)

Personal Memory

Retrieve the authenticated user's memories (Supabase-backed) with the same token:

from roboto_sai_sdk import RobotoSAI

client = RobotoSAI()
for mem in client.list_memories(limit=10):
    print(mem["category"], "—", mem["content"])

hits = client.search_memories("what did I say about the launch?", limit=5)

Public Model Surface

The SDK exposes four canonical models:

Use only canonical roboto-sai-* names. Legacy aliases (e.g. roboto-sai-lightning, -heavy, -flex) were removed in 0.2.0.

Inspect Runtime Profiles

from roboto_sai_sdk import (
    DEFAULT_PUBLIC_MODEL,
    PUBLIC_MODEL_NAMES,
    get_runtime_profile,
)

print(DEFAULT_PUBLIC_MODEL)             # "roboto-sai"
print(sorted(PUBLIC_MODEL_NAMES))       # all 4 names
print(get_runtime_profile("roboto-sai"))

Authentication

The Roboto SAI API authenticates with a Supabase JWT. The SDK resolves it in this order:

1. Explicit access_token=... argument
2. SAI_ACCESS_TOKEN environment variable

Get a token by logging in (e.g. roboto-sai auth login in the CLI). Set SAI_API_BASE_URL to override the default base URL (https://api.roboto-sai.com).

Public API

The stable surface is roboto_sai_sdk.__all__ (17 names; client, memory, keys, and MCP are lazy-loaded on first access):

RobotoSAI, SendMessageResult, StreamEvent, extract_text_delta, Memory, DEFAULT_PUBLIC_MODEL, PUBLIC_MODEL_NAMES, RuntimeProfile, get_runtime_profile, normalize_public_model, get_mcp_config, get_public_runtime_contract, CONTRACT_VERSION, CONTRACT_LAST_UPDATED, SAI_ACCESS_TOKEN, resolve_access_token, hash_key

import roboto_sai_sdk

assert roboto_sai_sdk.__all__ == [
    "RobotoSAI",
    "SendMessageResult",
    "StreamEvent",
    "extract_text_delta",
    "Memory",
    "DEFAULT_PUBLIC_MODEL",
    "PUBLIC_MODEL_NAMES",
    "RuntimeProfile",
    "get_runtime_profile",
    "normalize_public_model",
    "get_mcp_config",
    "get_public_runtime_contract",
    "CONTRACT_VERSION",
    "CONTRACT_LAST_UPDATED",
    "SAI_ACCESS_TOKEN",
    "resolve_access_token",
    "hash_key",
]

Voice/audio (roboto-sai-rovox) requires the realtime integration — send_message and stream raise NotImplementedError for that model. A future RobotoSAIVoice client will land in a separate release.

Repository

Source code, examples, and full release notes: https://github.com/Roboto-SAI-LLC/roboto-sai-sdk

License

Proprietary. See LICENSE in the repository.