bedrockrouter

Official Python SDK and Multi-Modal CLI for BedrockRouter.

BedrockRouter gives you immediate terminal access to every AI modality on AWS Bedrock—text, vision, video, and embeddings—all routed through your single sk_m11_... key and self-hosted proxy. You can also bring your own OpenAI or Anthropic API keys and route through the same proxy.

Install

pip install bedrockrouter

---

⚡ The Multi-Modal CLI

Before you write a single line of code, bedrockrouter gives you a frictionless terminal interface to every Bedrock model.

Set your proxy key once:

export BEDROCKROUTER_API_KEY="sk_m11_..."

🤖 Autonomous Coding Agent (bedrockrouter code)

Transform your terminal into a powerful, autonomous coding assistant (similar to Claude Code or Cursor). The agent can read files, write code, run shell commands, and manage your git repository safely.

bedrockrouter code

Agent Features:

• Safe File Operations: Uses a precise "Search and Replace" architecture to modify code safely without overwriting entire files.
• Project Awareness: Automatically respects .gitignore, scans project context, and runs syntax checks.
• Git Integration: Can review diffs, commit changes, and manage version control autonomously.
• Smart Sessions: Persistent sessions with auto-cleanup (7-day TTL) and persistent token/cost tracking.

Interactive Slash Commands: Once inside the REPL, use these commands to manage the agent's context:

• /add <file> - Inject a file into the context.
• /image <file> - Add an image/diagram for vision-based coding tasks.
• /undo / /redo - Step back or forward through agent edits.
• /diff - View pending changes.
• /review - Review the current state of the project.
• /compact - Compress session history to save tokens.

Interactive Chat

Start a continuous conversation with any model (defaults to auto-routing):

bedrockrouter chat
bedrockrouter chat --model claude-3.5-sonnet

One-Shot Questions (Text & Vision)

Ask quick questions or process local documents and images directly in your terminal:

# Standard text prompt
bedrockrouter ask "Write a bash script to zip all python files"

# Vision (Image-to-text)
bedrockrouter ask "What is in this architectural diagram?" --file architecture.png

Video Understanding

Analyze video files natively (routes to Nova Pro by default):

bedrockrouter video "Summarize this meeting" --file zoom_recording.mp4

Generate Embeddings

bedrockrouter embed "Hello world" --model amazon.titan-embed-text-v2:0

---

SDK Quick Start

Your single API key seamlessly unlocks standard SDK usage. base_url defaults to the BedrockRouter proxy automatically.

Anthropic-compatible client

from bedrockrouter import BedrockRouter, Models

client = BedrockRouter(api_key="sk_m11_...")

msg = client.messages.create(
    model=Models.CLAUDE_SONNET_4_6,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
)
print(msg.content[0].text)

OpenAI-compatible client

from bedrockrouter import BedrockOpenAI, Models

client = BedrockOpenAI(api_key="sk_m11_...")

resp = client.chat.completions.create(
    model=Models.CLAUDE_SONNET_4_6,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
)
print(resp.choices[0].message.content)

Auto-routing — let the proxy pick the best model

from bedrockrouter import BedrockRouter, Models

client = BedrockRouter(api_key="sk_m11_...")

# Pass Models.AUTO — the proxy classifies the task and routes automatically
msg = client.messages.create(
    model=Models.AUTO,
    messages=[{"role": "user", "content": "Write a binary search in Python"}],
    max_tokens=1024,
)
print(msg.content[0].text)

---

Bring Your Own Key (BYOK) — Quick Start

When using 3rd-party models, initialize the client by passing your provider's API key alongside your BedrockRouter tracking key (mach11_key).

from bedrockrouter import BedrockOpenAI

client = BedrockOpenAI(
    api_key="sk-proj-...",     # 1. Your 3rd Party API Key (OpenAI/Anthropic)
    mach11_key="sk_m11_...",   # 2. Your BedrockRouter Tracking Key
)

# Request hits OpenAI directly, but usage metrics are logged to your dashboard!
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello!"}],
    provider="openai"
)
print(response.choices[0].message.content)

💡 BYOK & Auto-Routing: When using model="auto" in BYOK mode, the proxy still intelligently classifies your prompt using a lightning-fast Amazon Nova Micro router (~$0.000007), but executes the final heavy model call directly via your provided 3rd-party credentials. Pass mach11_key alongside api_key to enable this.

---

Bring Your Own Key (3rd Party Providers)

Route requests to external providers using your own API keys. Pass your provider key as api_key and your BedrockRouter identity key as mach11_key.

Header	Value	Purpose
Authorization	Bearer {api_key}	Your 3rd party provider key (OpenAI, Anthropic, etc.)
X-Mach11-Key	{mach11_key}	Your BedrockRouter identity key (sk_m11_...)

Use the provider parameter on each request to tell the proxy which provider to route to.

OpenAI (OpenAI-compatible client)

from bedrockrouter import BedrockOpenAI

client = BedrockOpenAI(
    api_key="sk-...",            # Your OpenAI API key
    mach11_key="sk_m11_...",     # Your BedrockRouter identity key
)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    provider="openai",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

Anthropic (Anthropic-compatible client)

from bedrockrouter import BedrockRouter

client = BedrockRouter(
    api_key="sk-ant-...",        # Your Anthropic API key
    mach11_key="sk_m11_...",     # Your BedrockRouter identity key
)

msg = client.messages.create(
    model="claude-haiku-4-5",
    provider="anthropic",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)
print(msg.content[0].text)

Native AWS Bedrock (default)

When using only your BedrockRouter key — no mach11_key, no provider flag — requests route to AWS Bedrock through your linked AWS account:

from bedrockrouter import BedrockRouter, Models

client = BedrockRouter(api_key="sk_m11_...")

msg = client.messages.create(
    model=Models.CLAUDE_SONNET_4_6,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
)
print(msg.content[0].text)

---

Auto-Routing

Pass model="auto" (or Models.AUTO) and the proxy automatically classifies your prompt using Amazon Nova Micro and routes it to the best model — balancing cost and accuracy.

How it works

Your request (model="auto")
        ↓
Nova Micro classifies the task (~$0.000007)
        ↓
Routes to the best model:
  SIMPLE_CHAT     → nova-micro          (cheapest)
  FACTUAL_QA      → nova-lite
  CODING          → qwen3-coder-30b
  MATH_REASONING  → deepseek-r1
  CREATIVE        → claude-sonnet-4-6
  ANALYSIS        → llama-4-maverick-17b
  COMPLEX         → claude-sonnet-4-6
  HIGH_STAKES     → claude-opus-4-6     (most capable)

Usage

from bedrockrouter import BedrockRouter, Models

client = BedrockRouter(api_key="sk_m11_...")

# Anthropic format
msg = client.messages.create(
    model=Models.AUTO,   # or model="auto"
    messages=[{"role": "user", "content": "Explain quicksort"}],
    max_tokens=1024,
)
print(msg.content[0].text)

from bedrockrouter import BedrockOpenAI, Models

client = BedrockOpenAI(api_key="sk_m11_...")

# OpenAI format
resp = client.chat.completions.create(
    model=Models.AUTO,   # or model="auto"
    messages=[{"role": "user", "content": "Explain quicksort"}],
    max_tokens=1024,
)
print(resp.choices[0].message.content)

Cost tracking

Both the classifier call and the routed model call are tracked separately on your dashboard:

Entry	What
nova-micro (classifier)	~$0.000007 per request
claude-sonnet-4-6 (or chosen model)	normal model cost

Auto-routing constants

Constant	String value	Notes
Models.AUTO	"auto"	Primary alias
Models.BR_AUTO	"br-auto"	Alternative alias, same behaviour

---

CLI — Model Discovery

All 100+ AWS Bedrock models are built into the package as static values (no AWS credentials needed for listing).

List all models

bedrockrouter models list

Filter by provider

bedrockrouter models list --provider anthropic
bedrockrouter models list --provider amazon
bedrockrouter models list --provider meta
bedrockrouter models list --provider mistral
bedrockrouter models list --provider cohere
bedrockrouter models list --provider minimax
bedrockrouter models list --provider moonshot
bedrockrouter models list --provider google
bedrockrouter models list --provider nvidia
bedrockrouter models list --provider openai
bedrockrouter models list --provider qwen
bedrockrouter models list --provider zai

Filter by keyword

bedrockrouter models list --filter sonnet
bedrockrouter models list --filter llama
bedrockrouter models list --filter nova
bedrockrouter models list --filter minimax
bedrockrouter models list --filter deepseek
bedrockrouter models list --filter gemma
bedrockrouter models list --filter qwen
bedrockrouter models list --filter nemotron

Show only thinking / reasoning models

bedrockrouter models list --thinking

Show only models with short aliases

bedrockrouter models list --alias-only

Search by name

bedrockrouter models find "sonnet 4.5"
bedrockrouter models find "opus thinking"
bedrockrouter models find "haiku"
bedrockrouter models find "kimi"
bedrockrouter models find "llama 70b"
bedrockrouter models find "llama 4"
bedrockrouter models find "gemma"
bedrockrouter models find "glm"

Live-fetch from AWS Bedrock (requires AWS credentials)

bedrockrouter models fetch
bedrockrouter models fetch --filter sonnet
bedrockrouter models fetch --filter minimax
bedrockrouter models fetch --region us-west-2
bedrockrouter models fetch --ids-only        # one model ID per line

---

SDK Usage

Basic (non-streaming) — Anthropic format

from bedrockrouter import BedrockRouter, Models

# api_key is all you need — base_url defaults to the BedrockRouter proxy
client = BedrockRouter(api_key="sk_m11_...")

msg = client.messages.create(
    model=Models.CLAUDE_SONNET_4_5,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
)
print(msg.content[0].text)

Streaming — Anthropic format

with client.messages.stream(
    model=Models.CLAUDE_SONNET_4_6,
    messages=[{"role": "user", "content": "Tell me a story"}],
    max_tokens=1024,
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Basic (non-streaming) — OpenAI format

from bedrockrouter import BedrockOpenAI, Models

# api_key is all you need — base_url defaults to the BedrockRouter proxy
client = BedrockOpenAI(api_key="sk_m11_...")

resp = client.chat.completions.create(
    model=Models.NOVA_PRO,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1024,
)
print(resp.choices[0].message.content)

Streaming — OpenAI format

stream = client.chat.completions.create(
    model="meta.llama4-maverick-17b-instruct-v1:0",
    messages=[{"role": "user", "content": "Tell me a joke"}],
    max_tokens=512,
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

With system prompt

msg = client.messages.create(
    model=Models.CLAUDE_OPUS_4_6,
    system="You are a helpful Python tutor.",
    messages=[{"role": "user", "content": "Explain list comprehensions."}],
    max_tokens=512,
)

Thinking / reasoning models

msg = client.messages.create(
    model=Models.CLAUDE_SONNET_4_6_THINKING,
    messages=[{"role": "user", "content": "Solve: 2x + 5 = 13"}],
    max_tokens=2048,
)

Use any Bedrock model (native ID)

All 100+ Bedrock models work through the proxy — just pass the native model ID directly:

# Meta Llama 4
msg = client.messages.create(
    model="meta.llama4-maverick-17b-instruct-v1:0",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=512,
)

# Google Gemma 3
msg = client.messages.create(
    model="google.gemma-3-27b-it",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=512,
)

# Qwen3
msg = client.messages.create(
    model="qwen.qwen3-32b-v1:0",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=512,
)

# NVIDIA Nemotron
msg = client.messages.create(
    model="nvidia.nemotron-super-3-120b",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=512,
)

---

Available Models

Anthropic Claude (short aliases via BedrockRouter proxy)

Constant	Alias	Model
Models.CLAUDE_HAIKU_4_5	claude-haiku-4-5	Claude Haiku 4.5
Models.CLAUDE_HAIKU_4_5_THINKING	claude-4.5-haiku-thinking	Claude Haiku 4.5 (thinking)
Models.CLAUDE_SONNET_4_5	claude-sonnet-4-5	Claude Sonnet 4.5
Models.CLAUDE_SONNET_4_5_THINKING	claude-4.5-sonnet-thinking	Claude Sonnet 4.5 (thinking)
Models.CLAUDE_SONNET_4_6	claude-sonnet-4-6	Claude Sonnet 4.6
Models.CLAUDE_SONNET_4_6_THINKING	claude-4.6-sonnet-medium-thinking	Claude Sonnet 4.6 (thinking)
Models.CLAUDE_OPUS_4_5	claude-opus-4-5	Claude Opus 4.5
Models.CLAUDE_OPUS_4_5_THINKING	claude-4.5-opus-high-thinking	Claude Opus 4.5 (thinking)
Models.CLAUDE_OPUS_4_6	claude-opus-4-6	Claude Opus 4.6
Models.CLAUDE_OPUS_4_6_THINKING	claude-4.6-opus-high-thinking	Claude Opus 4.6 (thinking)

Anthropic Claude (native Bedrock IDs)

Constant	Model ID
Models.CLAUDE_SONNET_4	anthropic.claude-sonnet-4-20250514-v1:0
Models.CLAUDE_OPUS_4	anthropic.claude-opus-4-20250514-v1:0
Models.CLAUDE_35_SONNET_V2	anthropic.claude-3-5-sonnet-20241022-v2:0
Models.CLAUDE_35_SONNET_V1	anthropic.claude-3-5-sonnet-20240620-v1:0
Models.CLAUDE_35_HAIKU	anthropic.claude-3-5-haiku-20241022-v1:0
Models.CLAUDE_3_OPUS	anthropic.claude-3-opus-20240229-v1:0
Models.CLAUDE_3_SONNET	anthropic.claude-3-sonnet-20240229-v1:0
Models.CLAUDE_3_HAIKU	anthropic.claude-3-haiku-20240307-v1:0

Amazon

Constant	Model ID
Models.NOVA_MICRO	amazon.nova-micro-v1:0
Models.NOVA_LITE	amazon.nova-lite-v1:0
Models.NOVA_PRO	amazon.nova-pro-v1:0
Models.NOVA_PREMIER	amazon.nova-premier-v1:0
Models.NOVA_2_LITE	amazon.nova-2-lite-v1:0
Models.TITAN_TEXT_LITE	amazon.titan-text-lite-v1
Models.TITAN_TEXT_EXPRESS	amazon.titan-text-express-v1
Models.TITAN_TEXT_PREMIER	amazon.titan-text-premier-v1:0

Meta Llama

Constant	Model ID
Models.LLAMA4_MAVERICK_17B	meta.llama4-maverick-17b-instruct-v1:0
Models.LLAMA4_SCOUT_17B	meta.llama4-scout-17b-instruct-v1:0
Models.LLAMA33_70B	meta.llama3-3-70b-instruct-v1:0
Models.LLAMA32_90B	meta.llama3-2-90b-instruct-v1:0
Models.LLAMA32_11B	meta.llama3-2-11b-instruct-v1:0
Models.LLAMA31_405B	meta.llama3-1-405b-instruct-v1:0
Models.LLAMA31_70B	meta.llama3-1-70b-instruct-v1:0
Models.LLAMA31_8B	meta.llama3-1-8b-instruct-v1:0

Mistral

Constant	Model ID
Models.MISTRAL_LARGE_3	mistral.mistral-large-3-v1:0
Models.MISTRAL_LARGE	mistral.mistral-large-2407-v1:0
Models.MAGISTRAL_SMALL	mistral.magistral-small-2506-v1:0
Models.DEVSTRAL_2	mistral.devstral-2-v1:0
Models.PIXTRAL_LARGE	mistral.pixtral-large-2411-v1:0
Models.MISTRAL_NEMO	mistral.mistral-nemo-2407-v1:0
Models.MINISTRAL_8B	mistral.ministral-8b-2410-v1:0
Models.MINISTRAL_3B	mistral.ministral-3b-2410-v1:0
Models.MIXTRAL_8X7B	mistral.mixtral-8x7b-instruct-v0:1

Google

Constant	Model ID
Models.GEMMA_3_27B	google.gemma-3-27b-it
Models.GEMMA_3_12B	google.gemma-3-12b-it
Models.GEMMA_3_4B	google.gemma-3-4b-it

NVIDIA

Constant	Model ID
Models.NEMOTRON_SUPER_120B	nvidia.nemotron-super-3-120b
Models.NEMOTRON_NANO_30B	nvidia.nemotron-nano-3-30b
Models.NEMOTRON_NANO_12B	nvidia.nemotron-nano-12b-v2
Models.NEMOTRON_NANO_9B	nvidia.nemotron-nano-9b-v2

Qwen

Constant	Model ID
Models.QWEN3_VL_235B	qwen.qwen3-vl-235b-a22b
Models.QWEN3_NEXT_80B	qwen.qwen3-next-80b-a3b
Models.QWEN3_CODER_NEXT	qwen.qwen3-coder-next
Models.QWEN3_CODER_30B	qwen.qwen3-coder-30b-a3b-v1:0
Models.QWEN3_32B	qwen.qwen3-32b-v1:0

Z.AI (GLM)

Constant	Model ID
Models.GLM_5	zai.glm-5
Models.GLM_4_7	zai.glm-4.7
Models.GLM_4_7_FLASH	zai.glm-4.7-flash

Other Providers

Constant	Model ID	Provider
Models.COMMAND_R_PLUS	cohere.command-r-plus-v1:0	Cohere
Models.COMMAND_R	cohere.command-r-v1:0	Cohere
Models.JAMBA_15_LARGE	ai21.jamba-1-5-large-v1:0	AI21 Labs
Models.DEEPSEEK_R1	us.deepseek.r1-v1:5	DeepSeek
Models.DEEPSEEK_V3_2	deepseek.v3.2	DeepSeek
Models.GPT_OSS_120B	openai.gpt-oss-120b-1:0	OpenAI OSS
Models.GPT_OSS_20B	openai.gpt-oss-20b-1:0	OpenAI OSS
Models.KIMI_K2_5	kimi-k2.5	Moonshot
Models.KIMI_K2_THINKING	kimi-k2-thinking	Moonshot
Models.KIMI_K2_6	kimi-k2.6	Moonshot
Models.MINIMAX_M2_5	minimax-m2.5	MiniMax
Models.PALMYRA_X5	writer.palmyra-x5-v1:0	Writer
Models.PALMYRA_X4	writer.palmyra-x4-v1:0	Writer

Run bedrockrouter models list to see all 100+ models with pricing.

---

Two Clients — One Proxy

BedrockRouter exposes two fully compatible clients so you can use whichever SDK style you prefer:

	BedrockRouter	BedrockOpenAI
Format	Anthropic Messages API	OpenAI Chat Completions API
Response	msg.content[0].text	resp.choices[0].message.content
Streaming	client.messages.stream(...)	stream=True iterator
Models	All 100+ Bedrock models	All 100+ Bedrock models
BYOK	provider="anthropic" + mach11_key	provider="openai" + mach11_key
CLI Agent	bedrockrouter code — autonomous coding assistant	bedrockrouter code — autonomous coding assistant

---

SDK vs Anthropic SDK

Feature	Anthropic SDK	bedrockrouter
Install	pip install anthropic	pip install bedrockrouter
Import	from anthropic import Anthropic	from bedrockrouter import BedrockRouter
Client	Anthropic(api_key=...)	BedrockRouter(api_key=...) — base_url optional
API call	client.messages.create(...)	client.messages.create(...) (identical)
Models	Anthropic only	100+ models: Anthropic, Meta, Amazon, Mistral, Cohere, Google, NVIDIA, Qwen, MiniMax, Kimi, DeepSeek, Z.AI, OpenAI OSS...
BYOK (3rd party keys)	No	mach11_key + provider — route with your own OpenAI or Anthropic keys
Auto-routing	No	model=Models.AUTO — proxy picks best model automatically
CLI	None	bedrockrouter chat, ask, video, embed, code, and models list
Cost tracking	None	Built into proxy — classifier + routed model tracked separately
OpenAI format	No	BedrockOpenAI client included

Migrating from the Anthropic SDK requires only changing the import — base_url is pre-configured.

---

Links

• Website: bedrockrouter.com
• PyPI: pypi.org/project/bedrockrouter