ainfera-sdk

Official Python SDK for Ainfera — the neutral infrastructure exchange for AI agents. Deploy, monitor, and trust-score agents from any framework (LangChain, CrewAI, OpenClaw).

Status: Alpha (v0.1.0). API surface may change before v1.0.

Two packages

Ainfera ships two Python packages. Pick based on your use case:

Package	Install	What it is
ainfera	pip install ainfera	CLI — deploy, trust-score, kill-switch
ainfera-sdk	pip install ainfera-sdk	This package. Typed async-capable Python client

The CLI includes an embedded sync SDK at from ainfera.sdk import ... — use ainfera-sdk if you want async clients and a standalone package with no CLI dependencies.

Install

pip install ainfera-sdk

Requires Python 3.10+. Stripe Connect is a core dependency — no extra needed.

Framework extras (optional)

Install the extras matching your stack:

Agent frameworks

pip install "ainfera-sdk[langchain]"   # LangChain + LangGraph
pip install "ainfera-sdk[crewai]"      # Multi-agent crews
pip install "ainfera-sdk[openclaw]"    # Ainfera-native framework

Model providers

pip install "ainfera-sdk[anthropic]"   # Claude models + MCP
pip install "ainfera-sdk[openai]"      # GPT + Assistants
pip install "ainfera-sdk[huggingface]" # Open-weight models
pip install "ainfera-sdk[nvidia]"      # NIM inference + guardrails

Or install everything:

pip install "ainfera-sdk[all]"

Quickstart

from ainfera_sdk import Ainfera

client = Ainfera(api_key="ainf_sk_live_...")

# List your agents
for agent in client.agents.list():
    print(agent.id, agent.name, agent.status)

# Deploy a new agent
agent = client.agents.create(
    name="research-assistant",
    framework="langchain",
    compute_tier="standard",
)

# Fetch the composite trust score
score = client.trust.get_score(agent.id)
print(f"{score.overall_score} ({score.grade})")
# e.g. 847 (AA)

The SDK reads AINFERA_API_KEY from the environment if no key is passed. Base URL can be overridden with AINFERA_API_URL or the base_url= constructor argument.

Async

import asyncio
from ainfera_sdk import AsyncAinfera

async def main() -> None:
    async with AsyncAinfera(api_key="ainf_sk_live_...") as client:
        agents = await client.agents.list()
        score = await client.trust.get_score(agents[0].id)
        print(score.overall_score)

asyncio.run(main())

Resources

The client exposes five sub-resources:

Resource	What it does
client.agents	Deploy, list, update, kill, and delete agents
client.trust	Read trust scores, history, anomalies, and record events
client.billing	Inspect billing overview and metered usage
client.execution	Invoke agents and inspect execution runs
client.registry	Search the public agent registry (no auth required)

Agents

client.agents.list(page=1, per_page=20)          # list[Agent]
client.agents.get(agent_id)                      # Agent
client.agents.create(name=..., framework=...)    # Agent
client.agents.update(agent_id, name="renamed")   # Agent
client.agents.kill(agent_id)                     # emergency stop
client.agents.unkill(agent_id)                   # resume
client.agents.delete(agent_id)                   # permanent delete

Trust

The Ainfera trust score is a geometric mean across five dimensions:

• Reliability — uptime, latency, error rate
• Quality — output correctness, groundedness, user feedback
• Cost-efficiency — cost per task vs. benchmark
• Security — prompt injection resistance, data-handling compliance
• Compliance — regulatory posture, audit completeness

Any single dimension at zero collapses the overall score. There is no gaming this.

client.trust.get_score(agent_id)        # TrustScore
client.trust.get_history(agent_id)      # list[TrustHistoryEntry]
client.trust.get_anomalies(agent_id)    # list[Anomaly]
client.trust.record_event(agent_id, event_type="hallucination", data={...})
client.trust.recompute(agent_id)        # TrustScore (forced recompute)
client.trust.badge_url(agent_id)        # str (embed as SVG)

Grade boundaries:

Grade	Score
AAA	≥ 900
AA	≥ 800
A	≥ 700
BBB	≥ 600
BB	≥ 500
B	≥ 400
CCC	< 400 → auto-quarantine

Execution

run = client.execution.invoke(agent_id, input={"query": "..."})
client.execution.get_run(run.id)
client.execution.list_runs(agent_id, limit=20)

Registry (no auth required)

from ainfera_sdk import Ainfera

public = Ainfera(api_key=None)  # registry endpoints are public
public.registry.search(q="research", framework="langchain")
public.registry.get(agent_id)
public.registry.stats()

Billing

client.billing.overview()                              # BillingOverview
client.billing.usage(start="2026-04-01", end="2026-04-30")

Error handling

All non-2xx responses raise typed exceptions inheriting from AinferaError, exposing .status_code, .message, and .response_body:

from ainfera_sdk import AuthError, NotFoundError, RateLimitError, ServerError

try:
    client.agents.get("ag_missing")
except NotFoundError:
    ...
except AuthError:
    ...
except RateLimitError:
    ...
except ServerError:
    ...

Exception	HTTP status
AuthError	401, 403
ValidationError	400, 422
NotFoundError	404
RateLimitError	429
ServerError	5xx
ApiError	anything else non-2xx

Frameworks

Supported now:

• LangChain
• CrewAI
• OpenClaw (Ainfera-native, Alpha Q3 2026)

Coming Q4 2026:

• AutoGen (Microsoft)
• LlamaIndex
• Google ADK

Examples

See examples/:

• quickstart.py — list agents and fetch a trust score
• deploy_langchain.py — deploy a LangChain agent and invoke it
• trust_monitoring.py — pull trust history and record events

Development

git clone https://github.com/ainfera-ai/sdk-python.git
cd sdk-python
pip install -e ".[dev]"

# Tests (HTTP-mocked, no real API calls)
python -m pytest tests/ -v

# Lint + type check
ruff check ainfera_sdk/ --fix
ruff format ainfera_sdk/
mypy ainfera_sdk/ --strict

Links

• Homepage: https://ainfera.ai
• API docs: https://api.ainfera.ai/docs
• CLI: https://github.com/ainfera-ai/cli
• Issues: https://github.com/ainfera-ai/sdk-python/issues

Maintainer

Hizrian Raz — founder, Ainfera Pte. Ltd.

License

Apache 2.0 — see LICENSE. Copyright © Ainfera Pte. Ltd.