bedrockify ⚡

One binary. One port. Full OpenAI-compatible API — chat completions, embeddings, reasoning, tool use, vision, streaming, prompt caching — all backed by Amazon Bedrock.

POST /v1/chat/completions → Bedrock Converse API (Claude, Nova, Llama, Mistral, DeepSeek…)
POST /v1/embeddings → Bedrock InvokeModel (Titan Embed, Cohere Embed v3/v4, Nova Embed v2)
GET /v1/models → Lists foundation models
GET / → Health check

---

Quick Start

# 1. Install
curl -fsSL https://github.com/inceptionstack/bedrockify/releases/latest/download/install.sh | bash
# 2. Run as daemon (chat + embeddings on port 8090)
sudo bedrockify install-daemon \
--region us-east-1 \
--model us.anthropic.claude-opus-4-6-v1 \
--embed-model amazon.titan-embed-text-v2:0
sudo systemctl daemon-reload && sudo systemctl enable --now bedrockify
# 3. Chat
curl http://127.0.0.1:8090/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"claude-opus","messages":[{"role":"user","content":"Hello!"}]}'# 4. Embed
curl http://127.0.0.1:8090/v1/embeddings \
-H "Content-Type: application/json" \
-d '{"input":"semantic search rocks","model":"amazon.titan-embed-text-v2:0"}'

---

Supported APIs

API	Endpoint	Description
Chat Completions	POST /v1/chat/completions	Non-streaming and SSE streaming
Embeddings	POST /v1/embeddings	Titan, Cohere, Nova embedding models
Models	GET /v1/models	List available Bedrock foundation models
Health	GET /	Health check with config info

Chat Completions — Supported Parameters

Parameter	Type	Description
model	string	Model ID, alias, cross-region ID, or ARN
messages	array	Chat messages (system, developer, user, assistant, tool)
max_tokens	int	Maximum output tokens
max_completion_tokens	int	Max output tokens (takes precedence over max_tokens)
temperature	float	Sampling temperature (0.0–2.0)
top_p	float	Nucleus sampling (auto-stripped when reasoning is active)
stream	bool	Enable SSE streaming
stream_options	object	{include_usage: true} — separate usage chunk
stop	array	Stop sequences
tools	array	Function definitions for tool calling
tool_choice	string/object	auto, required, or specific tool
reasoning_effort	string	low/medium/high — enable extended thinking
extra_body	object	Passthrough to Bedrock additionalModelRequestFields
extra_body.thinking	object	Interleaved thinking config
extra_body.prompt_caching	object	Prompt caching control
n	int	Number of completions (only n=1 supported)

Chat Completions — Response Fields

Field	Description
choices[].message.content	Model response text
choices[].message.reasoning_content	Chain-of-thought reasoning (when reasoning enabled)
choices[].message.tool_calls	Function calls requested by the model
choices[].finish_reason	stop, tool_calls, length, content_filter
usage.prompt_tokens	Input token count
usage.completion_tokens	Output token count
usage.prompt_tokens_details.cached_tokens	Tokens served from cache (when caching active)
usage.completion_tokens_details.reasoning_tokens	Estimated reasoning tokens

Embeddings — Supported Parameters

Parameter	Type	Description
input	string/array	Text(s) to embed
model	string	Embedding model ID
encoding_format	string	float (default) or base64
dimensions	int	Output dimensions (Nova models: 256/384/1024/3072)

---

Features

Core

• Unified proxy — chat completions + embeddings on a single port
• OpenAI-compatible API — drop-in for any OpenAI SDK client
• Streaming — SSE streaming for chat completions with stream_options.include_usage support
• Tool use — function calling via Bedrock Converse (auto, required, specific tool)
• Vision — image inputs via base64 data URLs AND remote HTTP/HTTPS URLs
• Model aliases — short names, OpenRouter IDs, bare Bedrock IDs all work
• Cross-region inference — auto-prefixes (us., eu., ap., global.) based on region

Intelligence

• Reasoning / Extended Thinking — reasoning_effort (low/medium/high) for Claude 3.7/4/4.5 with reasoning_content in responses and streaming
• DeepSeek R1 Reasoning — automatic format detection (string format for DeepSeek, object format for Claude)
• Interleaved Thinking — extra_body.thinking for Claude 4/4.5 thinking between tool calls
• Prompt Caching — extra_body.prompt_caching for up to 90% cost reduction, with ENABLE_PROMPT_CACHING env var for global default

Compatibility

• Application Inference Profiles — pass ARN as model ID for cost tracking (no alias mangling)
• Developer role — developer messages treated as system (OpenAI compatibility)
• Message coalescing — consecutive same-role messages automatically merged (Bedrock requirement)
• No-prefill handling — models that reject assistant-ending conversations (e.g. claude-opus-4-6) get automatic user continuation
• temperature/topP conflict — auto-stripped for models that reject both simultaneously (Claude 4.5, Haiku 4.5, Opus 4.5)
• Extra body passthrough — unknown extra_body keys forwarded to Bedrock additionalModelRequestFields

Embedding Models

• Amazon Titan Embed v2 — 1024 dimensions, English + multilingual
• Cohere Embed v3/v4 — English, multilingual, latest v4 (1536 dims)
• Amazon Nova Multimodal Embeddings v2 — configurable dimensions (256/384/1024/3072)
• base64 encoding — encoding_format: "base64" for compact responses

Infrastructure

• Auth — IAM/SigV4 (default) or Bedrock API key (bearer token)
• Adaptive retries — max 8 attempts with adaptive backoff
• Config — TOML file, env vars, CLI flags (layered priority)
• Systemd daemon — install-daemon subcommand
• Self-update — update subcommand

---

Chat Models

Alias	Bedrock Model
claude-opus	anthropic.claude-opus-4-6-v1
claude-sonnet	anthropic.claude-sonnet-4-6-v1
claude-haiku	anthropic.claude-haiku-4-5-20251001-v1:0
claude-3.5-sonnet	anthropic.claude-3-5-sonnet-20241022-v2:0
nova-pro	amazon.nova-pro-v1:0
nova-lite	amazon.nova-lite-v1:0
nova-micro	amazon.nova-micro-v1:0
llama-4-maverick	meta.llama4-maverick-17b-instruct-v1:0
deepseek-r1	deepseek.deepseek-r1-v1:0
mistral-large	mistral.mistral-large-2407-v1:0

OpenRouter-style IDs (e.g. anthropic/claude-opus-4.6) also work. Cross-region inference prefixes (us., eu., ap.) are added automatically based on your --region.

---

Embedding Models

Model ID	Dims	Notes
amazon.titan-embed-text-v2:0	1024	Default, English + multilingual
cohere.embed-english-v3	1024	Cohere v3, English
cohere.embed-multilingual-v3	1024	Cohere v3, multilingual
cohere.embed-v4:0	1536	Cohere v4, latest
amazon.nova-2-multimodal-embeddings-v1:0	256–3072	Nova v2, configurable dimensions

---

Reasoning / Extended Thinking

Enable chain-of-thought reasoning for Claude 3.7/4/4.5 and DeepSeek R1. The model's thinking is returned in a separate reasoning_content field.

curl http://127.0.0.1:8090/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{ "model": "claude-sonnet", "messages": [{"role": "user", "content": "Which is larger: 9.11 or 9.8?"}], "max_completion_tokens": 4096, "reasoning_effort": "low" }'

Response includes reasoning_content:

{
"choices": [{
"message": {
"role": "assistant",
"content": "9.8 is larger than 9.11.",
"reasoning_content": "Comparing 9.11 and 9.8:\n9.8 = 9.80\n9.80 > 9.11"
}
}],
"usage": {
"completion_tokens_details": {"reasoning_tokens": 12}
}
}

Level	Budget	Use Case
low	30% of max_tokens (min 1024)	Quick comparisons, simple reasoning
medium	60% of max_tokens (min 1024)	Standard analysis
high	max_tokens - 1	Complex math, coding, deep analysis

DeepSeek R1 uses a different format automatically — no changes needed on the client side.

Streaming also works — reasoning deltas arrive as reasoning_content in delta objects before content deltas.

---

Interleaved Thinking

For Claude 4/4.5, enable thinking between tool calls using extra_body.thinking:

response=client.chat.completions.create(
model="claude-sonnet",
messages=[{"role": "user", "content": "What is 15 * 17?"}],
max_tokens=2048,
extra_body={"thinking": {"type": "enabled", "budget_tokens": 4096}}
)
print(response.choices[0].message.reasoning_content) # Chain of thoughtprint(response.choices[0].message.content) # Final answer

Note:budget_tokens can exceed max_tokens — bedrockify auto-adjusts max_tokens so the request is valid.

---

Prompt Caching

Reduce costs by up to 90% and latency by up to 85% for repeated prompts. Works with Claude and Nova models.

response=client.chat.completions.create(
model="claude-sonnet",
messages=[
{"role": "system", "content": "Long system prompt..."},
{"role": "user", "content": "Question"}
],
extra_body={"prompt_caching": {"system": True}}
)
# Check cache hitifresponse.usage.prompt_tokens_details:
print(f"Cached tokens: {response.usage.prompt_tokens_details.cached_tokens}")

Option	Effect
{"system": true}	Cache system prompt
{"messages": true}	Cache conversation history
{"system": true, "messages": true}	Cache both

Global default: Set ENABLE_PROMPT_CACHING=true env var to enable caching for all requests (per-request extra_body overrides).

Requirements: prompt ≥1,024 tokens. Cache TTL is 5 minutes (resets on each hit).

---

Application Inference Profiles

Pass an Application Inference Profile ARN as the model ID for per-application cost tracking:

curl http://127.0.0.1:8090/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{ "model": "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/my-app", "messages": [{"role": "user", "content": "Hello!"}] }'

ARNs pass through directly — no alias resolution or cross-region prefix applied.

---

Usage

OpenClaw Configuration

{
"llm": {
"baseUrl": "http://127.0.0.1:8090/v1",
"model": "claude-opus",
"embeddingBaseUrl": "http://127.0.0.1:8090/v1",
"embeddingModel": "amazon.titan-embed-text-v2:0"
}
}

Any OpenAI SDK

importopenaiclient=openai.OpenAI(
base_url="http://127.0.0.1:8090/v1",
api_key="not-used", # bedrockify uses AWS credentials
)
# Chatresponse=client.chat.completions.create(
model="claude-opus",
messages=[{"role": "user", "content": "Hello!"}],
)
# Chat with reasoningresponse=client.chat.completions.create(
model="claude-sonnet",
messages=[{"role": "user", "content": "Solve: 15 * 17"}],
max_completion_tokens=4096,
reasoning_effort="low",
)
print(response.choices[0].message.reasoning_content)
# Embeddingsembeddings=client.embeddings.create(
model="amazon.titan-embed-text-v2:0",
input="semantic search rocks",
)
# Embeddings (base64 format)embeddings=client.embeddings.create(
model="amazon.titan-embed-text-v2:0",
input="semantic search rocks",
encoding_format="base64",
)

---

Configuration

Config File

Copy bedrockify.example.toml to bedrockify.toml (next to binary, CWD, or ~/.config/bedrockify/):

region = "us-east-1"model = "us.anthropic.claude-opus-4-6-v1"embed_model = "amazon.titan-embed-text-v2:0"port = 8090host = "127.0.0.1"# Optional: Bedrock API key instead of IAM# bearer_token = "ABSK..."

CLI Flags

--region AWS region (default: us-east-1)
--model Default chat model (default: us.anthropic.claude-sonnet-4-6)
--embed-model Default embedding model (default: amazon.titan-embed-text-v2:0)
--port Port to listen on (default: 8090)
--host Bind host (default: 127.0.0.1)
--bearer-token Bedrock API key (alternative to IAM/SigV4)
--base-url Custom Bedrock endpoint URL
--init Write example bedrockify.toml and exit
--version Show version

Environment Variables

Variable	Description
AWS_BEARER_TOKEN_BEDROCK	Bedrock API key (overrides config)
ENABLE_PROMPT_CACHING	Set true to enable prompt caching globally
Standard AWS env vars	AWS_ACCESS_KEY_ID, AWS_REGION, etc.

---

Architecture

Client (OpenAI SDK / OpenClaw / Hermes / curl)
│
▼
bedrockify :8090
│
┌────┴────────────────────────┐
│ │
▼ ▼
POST /v1/chat/completions POST /v1/embeddings
│ │
▼ ▼
BedrockConverser BedrockEmbedder
(Converse API) (InvokeModel API)
│ │
├─ Reasoning config ├─ Titan Embed v2
├─ Prompt caching ├─ Cohere Embed v3/v4
├─ Tool config ├─ Nova Embed v2
├─ Message coalescing └─ base64 encoding
├─ topP conflict resolution
├─ No-prefill handling
└─ Image fetching
│ │
▼ ▼
Amazon Bedrock Amazon Bedrock
(Claude/Nova/Llama/ (Titan/Cohere/
Mistral/DeepSeek) Nova)

---

IAM Permissions

Minimum IAM policy:

{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"bedrock:InvokeModel",
"bedrock:InvokeModelWithResponseStream",
"bedrock:Converse",
"bedrock:ConverseStream",
"bedrock:ListFoundationModels"
],
"Resource": "*"
}
]
}

---

Building from Source

git clone https://github.com/inceptionstack/bedrockify
cd bedrockify
go build -o bedrockify ./cmd/bedrockify/
go test ./... -v # 202 tests

---

License

MIT — see LICENSE