Skip to main content
Glama
bedrock-multi-model-mcp

bedrock-multi-model-mcp

by Sagargupta16

bedrock-multi-model-mcp

Stars Forks License: MIT Last commit CI

MCP server for AWS Bedrock - text, image, video and embeddings from any model. Use Claude, GPT, Grok, Kimi, Gemma, Mistral, Nova, Qwen, DeepSeek, and more from Claude Code (or any MCP client).

Why: Bedrock already gives one AWS account access to models from eleven text providers, but each one is a different model ID, inference profile, region and request shape. This server puts them behind eleven MCP tools and one set of AWS credentials, so an agent can ask a second model for a review, run the same prompt across five models side by side, have a judge model score those answers blind, or generate and edit an image, without leaving the chat or writing SDK code.

At a glance: 35 text models from 11 providers, 3 image models, 13 image-editing services, 1 video model and 3 embedding models, all defined as data in src/data/ and validated at startup. The 26 text models in the 2026-10-18 benchmark answered every call. Since that run, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 and, in 0.7.0, twelve models from OpenAI, xAI, Moonshot AI, Google, MiniMax, NVIDIA, Mistral and Qwen joined the registry, while the five models AWS moved to Legacy on 2026-09-30 (Llama 4 Maverick and Scout, Llama 3.3 70B, Devstral 2, DeepSeek R1) left it. 0.8.0 added Claude Fable 5.2, GPT-6 Astra and GPT-6 Luna, and 0.9.0 moved to GPT-6.2 Astra and GPT-6.2 Sol.

A real bedrock_judge run from 2026-10-18, called through scripts/call-tool.mjs with the arguments in docs/demo/judge-args.json. The output lines are excerpted from that run with the table columns aligned; the full reply also carries the judge's reasoning and all three answers. Llama 4 Maverick has since gone Legacy on Bedrock and left the registry, so swap llama4 for another model to rerun it.

Tools

Tool

Description

bedrock_ask

Send a prompt to any text model. Returns response with token counts, an estimated cost and latency. schema returns JSON that matches a JSON schema; effort sets the thinking depth; files adds local text files to the prompt; level picks a model instead of model; conversation_id keeps a follow-up conversation.

bedrock_compare

Same prompt to 2-5 models side by side. Compare quality, speed, and style. Takes schema, effort, files and level too.

bedrock_judge

Same prompt to 2-5 models, then a judge model scores the answers blind and picks a winner, optionally merging them into one answer. The verdict comes back as schema-checked JSON when the judge supports it. Takes files and level.

bedrock_panel

A short discussion between 2-5 models: they answer in turn, each seeing what the earlier ones said, with optional roles (critic, builder, fact-checker and more), 1-3 rounds and a closing summary from a moderator model. Returns the transcript. Takes files and level.

bedrock_list_models

List available text models with aliases, use cases, capabilities and the estimated price per 1M tokens.

bedrock_doctor

Check which text models work on your AWS account right now: one small billed call per model, one status each (ok, throttled, denied, invalid or error).

bedrock_generate_image

Generate images from text (Stable Image Ultra, SD3.5 Large, Stable Image Core). Saves PNG locally.

bedrock_edit_image

Edit or transform a local image with the 13 Stability AI Image Services: inpaint, outpaint, erase, remove the background, find and replace or recolor an object, three upscalers, sketch and structure control, style guide and style transfer. Saves PNG locally.

bedrock_generate_video

Start async video generation (Luma Ray 2). Output to S3. start_image_path and end_image_path use local images as the first and last frame.

bedrock_video_status

Check video generation job progress.

bedrock_embed_similarity

Embed 2+ texts (Titan, Cohere) and return a cosine-similarity matrix. images adds local images for the multimodal models.

The server also offers three MCP prompts, ready-made instructions that call the tools above: second-opinion (ask another model, default gpt-astra, and report where it agrees and disagrees), compare-models (run bedrock_compare and summarize the differences in a table) and panel-discussion (run bedrock_panel at a level, default medium, and report agreement, disagreement and a recommendation).

Estimated cost

bedrock_ask, bedrock_compare, bedrock_judge and bedrock_panel end their output with an estimated cost (Est. cost), and bedrock_list_models shows the price per 1M input and output tokens. The prices are in the price field of src/data/text-models.json, taken from the AWS Price List bulk files for us-east-1 (the AmazonBedrockFoundationModels and AmazonBedrock offers), read 2026-11-20: on-demand, standard tier, USD. For ids with the us. profile prefix the non-global price is used (AWS lists a global price about 10 percent lower). The estimate leaves out cache, batch and long-context pricing and uses the token counts Bedrock reports, so treat it as an estimate and check AWS Bedrock pricing for the bill.

Related MCP server: Ollama MCP Server

Supported Models

Model data lives in src/data/ (text-models.json, image-models.json, image-edit-models.json, video-models.json, embedding-models.json) - edit those files to add or remove models without touching code.

Text (Converse API, and bedrock-mantle for Gemma 4)

Provider

Models

Aliases

Anthropic

Claude Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5 (current); Fable 5.1, Opus 5 / 4.6, Sonnet 5 / 4.6, Haiku 4.5 (previous-gen)

fable, claude-opus, claude-sonnet, claude-haiku

OpenAI

GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, GPT-6 Luna, GPT-OSS 120B, GPT-OSS 20B

gpt-astra, gpt-sol, gpt-terra, gpt-luna, gpt-oss, gpt-oss-20b

xAI

Grok 4.7

grok

Moonshot AI

Kimi K3

kimi

Mistral

Mistral Large 3, Ministral 3 14B, Magistral Small, Pixtral Large

mistral-large, ministral, magistral, pixtral

Amazon

Nova Pro, Nova 2 Lite, Nova Lite, Nova Micro

nova-pro, nova2-lite, nova-lite, nova-micro

Qwen

Qwen3 Coder Next, Qwen3 VL 235B, Qwen3 Next 80B A3B

qwen-coder, qwen-vl, qwen-next

DeepSeek

DeepSeek V3.2

deepseek

Google

Gemma 4 31B, Gemma 4 26B A4B, Gemma 4 E2B (bedrock-mantle)

gemma, gemma-4-26b, gemma-4-e2b

MiniMax

MiniMax M2.5

minimax

NVIDIA

Nemotron 3 Super 120B

nemotron

Gemma 4 is served only on Bedrock's OpenAI-compatible bedrock-mantle endpoint, not through Converse, so those entries carry "api": "mantle" and the server calls their Chat Completions API (https://bedrock-mantle.<region>.api.aws/openai/v1) instead. Every tool works with them, side by side with Converse models in bedrock_compare and bedrock_judge. They need AWS_BEARER_TOKEN_BEDROCK: AWS also accepts SigV4 on that endpoint, but this server does not sign mantle requests.

Listed on Bedrock but not in the registry: GLM 5.3 and Claude Mythos (access is gated per account; GLM 5.3 and Mythos 5.1 failed the 2026-11-01 probe), Kimi K3.1 Pro - Highspeed (launched 2026-11-04; every probe call on 2026-11-10 and 2026-11-20 was throttled), and Grok 4.3 (answers on bedrock-mantle, but Grok 4.7 supersedes it). AWS lists no Active Meta model.

Bare-tier aliases (fable, claude-opus, claude-sonnet, claude-haiku, gpt-astra, gpt-sol, gpt-terra, gpt-luna) track the current model in that tier; version-pinned aliases (fable-5.1, claude-opus-5, claude-sonnet-5, claude-haiku-4.5, claude-opus-4.6, claude-sonnet-4.6) keep the previous generations reachable. AWS announced Claude Opus 5.5 on Bedrock on 2026-09-22, Sonnet 5.5 on 2026-09-28, Haiku 5.5 on 2026-10-07 and Fable 5.2 on 2026-11-09, so the bare claude-opus, claude-sonnet and claude-haiku resolve to the 5.5 models and fable to Fable 5.2.

temperature is the one thing that differs across those entries. Anthropic deprecated temperature / top_p / top_k for Claude Opus 4.7 and later, where a non-default value returns a 400, so text-models.json marks those entries noTemperature and this server omits the parameter for them - the temperature argument is accepted but has no effect. The three 5.5 models return the same 400 (probed 2026-10-21), and so do GPT-6.1 Terra, Grok 4.7 and Kimi K3 ("This model doesn't support the temperature field", probed 2026-11-01), Fable 5.2 and GPT-6 Luna (probed 2026-11-10) and GPT-6.2 Astra and Sol (probed 2026-11-20). Gemma 4 and the other open-weight models take it. Claude Haiku 4.5, Opus 4.6 and Sonnet 4.6 still honor it, so pin claude-sonnet-4.6 or claude-haiku-4.5 if you need to set it on a Claude model.

The maxTokens value in text-models.json is the default output cap this server sends when a call omits max_tokens - deliberately below the model's own ceiling (Anthropic documents 128K max output for Fable 5.2 and 5.1 and the 5.5 and 5 models, 64K for Haiku 4.5). Pass max_tokens to go higher on a single call.

Structured output and effort

schema on bedrock_ask and bedrock_compare sends a JSON schema as Converse outputConfig.textFormat, so the reply is JSON that matches it, shown pretty-printed. Object schemas need "additionalProperties": false; Bedrock rejects numeric and string length constraints (minimum, maxLength and the like) with a 400. Probed on every registry model (2026-10-21, and 2026-11-01, 2026-11-10 and 2026-11-20 for the models added then): all 10 Claude models, GPT-6.2 Astra and Sol, GPT-6.1 Terra, GPT-6 Luna, both GPT-OSS models, Grok 4.7, Kimi K3, Mistral Large 3, Ministral 3, all three Qwen3 models, DeepSeek V3.2, MiniMax M2.5, Nemotron 3 Super and the three Gemma 4 models accept it (structured_output in bedrock_list_models; for Gemma it is sent as Chat Completions response_format). Nova and Pixtral answer "This model doesn't support the outputConfig field", and Magistral Small accepts the field but ignores the schema, so this server refuses the call before it is billed. GPT-OSS puts a stray character before its JSON (B{...}); the reply parser skips anything before the first brace.

effort (low, medium, high, xhigh, max) sets the thinking depth and token spend through Converse outputConfig.effort. Every Claude model except Haiku 4.5 takes it, and so do GPT-6.1 Terra, Grok 4.7 and Kimi K3 (probed 2026-11-01), GPT-6 Luna (probed 2026-11-10) and GPT-6.2 Astra and Sol (probed 2026-11-20). Gemma 4 takes low to xhigh as Chat Completions reasoning_effort and answers max with a 400 ("not supported with this model"). Claude Opus 5.5 and Haiku 5.5 default to medium, the other Claude models to high. On 2026-10-21, "How many primes are below 100?" cost Haiku 5.5 3 output tokens at low and 60 at max, with the same answer. bedrock_ask rejects effort for a model that does not take it; bedrock_compare and bedrock_judge apply it only to the candidates that do, so one value can tune the Claude answers in a mixed list.

Image (InvokeModel API)

Model

Alias

Region

Stable Image Ultra (default)

stable-ultra

us-west-2

Stable Diffusion 3.5 Large

sd3.5

us-west-2

Stable Image Core

stable-core

us-west-2

Stability models take an aspect ratio, not pixel dimensions. The width and height arguments only select the closest supported ratio (1:1, 16:9, 9:16, 3:2, 2:3, 4:5, 5:4, 21:9, 9:21); the model picks the output pixels. The size reported by bedrock_generate_image is read back out of the saved PNG, so it is the real output size rather than what you asked for.

Image editing (InvokeModel API)

bedrock_edit_image takes a local PNG, JPEG or WebP (image_path) and an operation. Each operation is one of Stability AI's image services, called through its US inference profile (us.stability.*) from us-west-2.

Operation

Service

Needs besides the image

Also takes

inpaint

Stable Image Inpaint

prompt

mask_path, grow_mask

outpaint

Stable Image Outpaint

one of left, right, up, down (pixels)

prompt, creativity

erase

Stable Image Erase Object

-

mask_path, grow_mask

remove-background

Stable Image Remove Background

-

-

search-replace

Stable Image Search and Replace

prompt, search_prompt

grow_mask

search-recolor

Stable Image Search and Recolor

prompt, select_prompt

grow_mask

upscale-fast

Stable Image Fast Upscale

-

-

upscale-conservative

Stable Image Conservative Upscale

prompt

creativity

upscale-creative

Stable Image Creative Upscale

prompt

creativity

control-sketch

Stable Image Control Sketch

prompt

control_strength

control-structure

Stable Image Control Structure

prompt

control_strength

style-guide

Stable Image Style Guide

prompt

fidelity, aspect_ratio

style-transfer

Stable Image Style Transfer

style_image_path

prompt, composition_fidelity, style_strength, change_strength

Most also take negative_prompt, seed and style_preset. A field the operation does not take, or a missing required one, is refused before the call is billed. Without mask_path, inpaint and erase use the input image's alpha channel as the mask. Only PNG, JPEG and WebP files are read and sent, so a mistyped path never ships some other local file to Bedrock. Each call is billed per image; Conservative and Creative Upscale cost several times more than the other services (AWS Bedrock pricing).

Video (Async API)

Model

Alias

Duration

Resolution

Region

Luma Ray 2

luma-ray, ray2

5s or 9s

720p

us-west-2

The output S3 bucket must be in us-west-2 (same region as the model). start_image_path and end_image_path make the video start or end on a local JPEG or PNG (512x512 to 4096x4096 pixels); the images are checked before the job is started.

Amazon Nova Canvas/Reel and Titan Image were removed when AWS marked them end-of-life (Nova Canvas and Nova Reel EOL 2026-09-30). The registry only ships ACTIVE models: AWS blocks new customers from Legacy ones. npm run lifecycle checks every registry id against AWS's public lifecycle page (no credentials needed), and CI runs it weekly and on every registry change. npm run new-models does the opposite check: it lists the Bedrock model cards that the catalog neither has nor has skipped on purpose (in scripts/model-skips.json), and a weekly CI job opens or updates one issue with them.

Embeddings (InvokeModel API)

Model

Alias

Dimensions

Amazon Titan Text Embeddings V2

titan-v2

1024

Amazon Titan Multimodal Embeddings

titan-multimodal

1024

Cohere Embed v4

cohere

1536

bedrock_embed_similarity takes texts, and images (local JPEG or PNG paths) for the two multimodal models, titan-multimodal and cohere; titan-v2 embeds text only and refuses an image before any call. At least 2 inputs and at most 50 across both lists, and the matrix labels an image by its file name. Cohere Embed v4 images can be up to 5 MB; WebP is not read. Text-only input to Titan Multimodal behaves like titan-v2 at the same 1024 dimensions. The Cohere Embed v3 aliases (cohere-en, cohere-multi) now resolve to Embed v4, which covers English and multilingual text. Titan embeds one text per request, so this server sends 8 at a time.

You can also pass any valid Bedrock model ID directly. Note: many foundation models require a cross-region inference profile (us. prefix) for on-demand invocation - the registry uses the working form for each model.

Model availability varies by region and account. The registry IDs are verified working in us-east-1, but individual models may not be enabled in your region or granted to your account. Models you can't access return a clear error (per-model in bedrock_compare); the rest still work. Enable models in the AWS Console under Bedrock > Model access.

Setup

Prerequisites

  • Node.js >= 22 (Node 20 reached end of life on 2026-04-30; CI runs 22, 24 and 26)

  • AWS credentials configured (bearer token, env vars, ~/.aws/credentials, SSO, or IAM role)

  • Bedrock model access enabled in your AWS account (request access in the AWS Console under Bedrock > Model access)

Install

The server is published on npm as bedrock-multi-model-mcp and listed on the MCP Registry as io.github.Sagargupta16/bedrock-multi-model-mcp, so an MCP client can start it with npx and nothing has to be cloned or built.

Configure in Claude Code

One command:

claude mcp add --transport stdio --scope user bedrock \
  --env AWS_REGION=us-east-1 \
  --env AWS_BEARER_TOKEN_BEDROCK=ABSK... \
  -- npx -y bedrock-multi-model-mcp

Or add it by hand to your Claude Code MCP settings (~/.claude.json):

{
  "mcpServers": {
    "bedrock": {
      "command": "npx",
      "args": ["-y", "bedrock-multi-model-mcp"],
      "env": {
        "AWS_BEARER_TOKEN_BEDROCK": "ABSK...",
        "AWS_REGION": "us-east-1"
      }
    }
  }
}

Configure in Claude Desktop

Same server block, in claude_desktop_config.json:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

  • Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "bedrock": {
      "command": "npx",
      "args": ["-y", "bedrock-multi-model-mcp"],
      "env": {
        "AWS_BEARER_TOKEN_BEDROCK": "ABSK...",
        "AWS_REGION": "us-east-1",
        "BEDROCK_MCP_OUTPUT_DIR": "C:\\Users\\you\\bedrock-images"
      }
    }
  }
}

Install the .mcpb bundle in Claude Desktop

Each GitHub release carries bedrock-multi-model-mcp-<version>.mcpb, an MCP Bundle that Claude Desktop and other MCPB hosts install in one click, with no Node setup or config file to edit. Download it from the releases page and open it. The install dialog asks for the Bedrock API key (leave it empty to use your AWS credential chain), the region, the output folder and the files folder. To build it yourself, run npm run pack:mcpb: it builds, then writes the file into the repo root, with its version taken from package.json.

Run from source

To change the server or pin it to a local build, clone and build it, then point the MCP config at node /path/to/bedrock-multi-model-mcp/dist/index.js instead of npx:

git clone https://github.com/Sagargupta16/bedrock-multi-model-mcp.git
cd bedrock-multi-model-mcp
npm install
npm run build

Set BEDROCK_MCP_OUTPUT_DIR explicitly for Desktop: on Windows it spawns the server from a write-protected directory (C:\Windows\System32), so the default "save next to cwd" behavior cannot apply and generated images fall back to ~/bedrock-images/.

Environment Variables

Variable

Default

Description

AWS_REGION

us-east-1

AWS region for Bedrock API calls

AWS_DEFAULT_REGION

-

Fallback region, used only when AWS_REGION is unset

AWS_BEARER_TOKEN_BEDROCK

-

Bedrock API Key bearer token (recommended). On Windows, when the MCP client passes it empty or not at all (the Claude desktop app does not pass the user's key on), the server reads it from the user environment (HKCU\Environment)

AWS_ACCESS_KEY_ID

-

AWS access key (if not using bearer token/SSO)

AWS_SECRET_ACCESS_KEY

-

AWS secret key (if not using bearer token/SSO)

AWS_PROFILE

-

Named AWS profile from ~/.aws/credentials

BEDROCK_MCP_OUTPUT_DIR

cwd (or ~/bedrock-images if cwd is not writable)

Default output directory for generated images. Overridden by the output_dir tool arg. Directory is auto-created.

BEDROCK_MCP_FILES_ROOT

the server's working folder

The one folder the files option may read from. Paths in files are relative to it, and anything outside it is refused. When it is unset and the working folder is a home folder or a drive root, files is off until you set it.

Authentication

The server supports two auth methods:

  1. Bedrock API Key (bearer token) - Set AWS_BEARER_TOKEN_BEDROCK. The AWS SDK reads it and sends it as Authorization: Bearer, ahead of any IAM credentials. A set-but-empty value (for example an MCP config passing "${AWS_BEARER_TOKEN_BEDROCK}" from a client whose own environment lacks the variable) is ignored, so the IAM chain answers instead of every call failing with "the token is not defined".

  2. IAM credentials - Standard AWS credential chain (env vars, profile, SSO, IAM role). Every model works this way except Gemma 4, whose bedrock-mantle endpoint this server reaches only with the bearer token.

Usage Examples

Ask a single model

"Use bedrock_ask to ask kimi to explain Docker in 3 sentences"

Get JSON that matches a schema

"Use bedrock_ask with claude-haiku, effort low, and a schema with category, subcategory and amount_inr: Swiggy order, Mattar Paneer Meal, 203 rupees"

Compare models

"Use bedrock_compare with kimi, nova-pro, and mistral-large: What are the pros and cons of microservices?"

bedrock_compare runs one full generation per model in the list, and each one is billed independently - three models is three generations, five is five. This server adds no markup and sets no rates; billing is entirely AWS Bedrock's, per model and per token. Current rates: AWS Bedrock pricing and, for the Claude models, Anthropic pricing.

Let a judge pick the best answer

"Use bedrock_judge with kimi, nova-pro and deepseek: In two sentences, when should a team pick SQS over SNS?"

The candidates answer in parallel. The judge (default claude-opus, override with judge) then sees the answers shuffled and labeled A, B, C with the model names removed, scores each from 1 to 10, and names a winner. The tool maps the labels back to models and returns a ranked table, the judge's reasoning, and every answer. When the judge supports structured output (the default Opus 5.5 does), the verdict is requested against a schema whose winner can only be one of the labels, so it always parses. Pass criteria to change what is scored and synthesize: true to also get one merged answer. A candidate that fails is listed under Failed and the rest are still judged; if the judge call fails, the answers still come back. Billing is one generation per candidate plus one for the judge.

Run a panel discussion

"Use bedrock_panel with nova-pro, kimi and deepseek, roles builder, critic and fact-checker, 2 rounds, summarize with claude-haiku: Should a small team move its cron jobs to Step Functions?"

The models speak one after another in the order given, so each one sees what the earlier ones said; in round 2 and 3 each answers again having read everything so far. A model's view is a single user message with the question and the turns so far, each prefixed with the speaker's name in square brackets and its own turns marked (you). What one model said reaches the next only as that quoted text, never in a system prompt, where it would carry more weight. roles takes one entry per model: a preset (critic, builder, fact-checker, devils-advocate, simplifier, moderator), free text up to 200 characters, or null. summarize names a model that closes with a short summary of where the panel agrees and disagrees; it is told it is the moderator and is not on the panel unless it is also in models. Without models, level picks a panel of three: cheap (Nova Micro, Ministral 3 14B, Qwen3 Next 80B A3B), medium (Claude Haiku 5.5, GPT-6 Luna, Mistral Large 3) or intense (Claude Fable 5.2, GPT-6.2 Astra, Grok 4.7). The same model may take two seats only in different roles. A model that fails loses its turn, the next one carries on, and the failure is not mentioned to it. Turns run sequentially and each is billed: models x rounds generations, plus one for the summary.

Pick models by level

"Use bedrock_compare with level cheap: What is the difference between a mutex and a semaphore?"

level is cheap, medium or intense, the same three presets bedrock_panel uses, and works on bedrock_ask, bedrock_compare and bedrock_judge: bedrock_ask uses the first model of the level (nova-micro, claude-haiku or fable), the others use all three. model or models wins over level in bedrock_compare and bedrock_judge; bedrock_ask refuses both together. level never changes the judge.

Add local files to the prompt

"Use bedrock_ask with claude-sonnet and files ['src/index.ts']: Summarize what this file does"

files works on bedrock_ask, bedrock_compare, bedrock_judge and bedrock_panel. Paths are relative to BEDROCK_MCP_FILES_ROOT (default the server's working folder), and files outside that folder are refused. Up to 10 text files, 256 KB each and 1 MB in total; names like .env, .npmrc, .git-credentials and private keys, anything under .ssh or .aws, network paths (\\host\share), binary files and missing files are refused, and a refusal fails the call before any model is billed. A desktop client can start the server in your home folder: the default is refused there, so set BEDROCK_MCP_FILES_ROOT to a project folder. The content goes to the model provider, as part of the prompt, and with conversation_id it is sent again on each follow-up.

Follow up in a conversation

"Use bedrock_ask with kimi and conversation_id design-review: Outline a retry policy for an SQS consumer"
"Use bedrock_ask with kimi and conversation_id design-review: Now add a dead-letter queue to that"

Calls with the same conversation_id send the earlier turns along as context. The id is 1 to 64 letters, digits, - or _. Conversations live in the server's memory only: they end 30 minutes after the last turn or when the server restarts, and hold at most 20 turns. Without conversation_id every call stands alone.

Check which models work on your account

"Use bedrock_doctor"
"Use bedrock_doctor with models nova-pro and kimi"

Sends one small billed call (a one-word prompt, at most 32 output tokens) to every text model in the registry, or only the models you name, and returns a table with ok, throttled (a rate or quota limit right now), denied (no model access on this account), invalid (not in the registry, or not offered in this region) or error, plus latency and a short note. concurrency (1 to 8, default 4) sets how many run at once.

Generate an image

"Use bedrock_generate_image with stable-ultra: A futuristic Tokyo street at night, neon lights, rain"

Edit an image

"Use bedrock_edit_image to remove the background from ./photo.png"
"Use bedrock_edit_image with search-replace on ./street.png: replace the car with a blue pickup truck"

Generate a video

"Use bedrock_generate_video: Closeup of ocean waves crashing on rocks at sunset, s3_uri: s3://my-bucket/videos/"
"Use bedrock_generate_video with start_image_path ./first.png and end_image_path ./last.png: The camera pulls back from the door, s3_uri: s3://my-bucket/videos/"

Either keyframe image alone also works.

Compare text similarity

"Use bedrock_embed_similarity to compare: 'a cat', 'a kitten', 'a car'"
"Use bedrock_embed_similarity with titan-multimodal to compare the text 'a cat' with images ./cat.png and ./car.png"

Image inputs need a multimodal model (titan-multimodal or cohere).

Architecture

The spec behind the picture is docs/diagrams/architecture.yaml, rendered with orrery.

A tool call resolves its model alias against the registry, picks the request shape and region from that entry, and sends it through a cached SDK client for that region. The request carries the MCP call's abort signal, so when the client cancels a tool call (or gives up on it), the server stops waiting on Bedrock: a Nova Micro call aborted at 400 ms settled at 412 ms on 2026-11-01, against 6.4 s for the same call left to finish. bedrock_generate_video is the exception, because cancelling a StartAsyncInvoke that was already sent could leave a billed job running with no ARN to track it.

src/
  index.ts          # MCP server - tool registration, stdio transport
  judge.ts          # bedrock_judge - blind labels, judge prompt, verdict parsing
  panel.ts          # bedrock_panel - roles, each model's view of the discussion, rounds, summary
  doctor.ts         # bedrock_doctor - one small call per text model, status per model
  files.ts          # `files` option - root folder, limits, refusals, prompt blocks
  cost.ts           # Estimated cost from the registry `price`
  conversation.ts   # conversation_id - in-memory follow-up turns with size and time limits
  prompts.ts        # MCP prompts: second-opinion, compare-models, panel-discussion
  models.ts         # Loads + validates registry data, resolves aliases
  types.ts          # Zod schemas for model data validation
  bedrock/
    client.ts       # Shared region/auth config, one cached SDK client per region
    converse.ts     # Converse API wrapper (text models), dispatches mantle models
    mantle.ts       # bedrock-mantle Chat Completions wrapper (Gemma 4)
    image.ts        # InvokeModel wrapper (image models)
    edit.ts         # InvokeModel wrapper (Stability image-edit services)
    video.ts        # Async invoke wrapper (video models), Luma Ray 2 keyframes
    embed.ts        # InvokeModel wrapper (embeddings, text and image inputs) + cosine similarity
  data/
    text-models.json       # Text model registry (data, not code)
    image-models.json      # Image model registry
    image-edit-models.json # Image-edit service registry (operation, required and optional fields)
    video-models.json      # Video model registry
    embedding-models.json  # Embedding model registry
scripts/
  call-tool.mjs     # Minimal stdio MCP client: list tools or call one
  benchmark.mjs     # Latency benchmark across every text model
  check-lifecycle.mjs # Fails if a registry model is Legacy or EOL on AWS's lifecycle page
  check-new-models.mjs # Lists Bedrock model cards the catalog neither has nor skips on purpose
  model-skips.json  # Model cards deliberately not in the catalog, with the reason
  pack-mcpb.mjs     # Builds the one-click .mcpb bundle (npm run pack:mcpb)
mcpb/
  manifest.json     # Bundle manifest; its version is stamped from package.json at pack time

Model definitions are data, kept in src/data/*.json and validated against Zod schemas at load time, so the catalog can be updated without changing logic. The build copies these JSON files into dist/data/.

  • Text: Uses the Converse API - unified interface across the text models, except Gemma 4, which uses bedrock-mantle Chat Completions

  • Image: Uses InvokeModel with the Stability request format (prompt and aspect ratio)

  • Image editing: Uses InvokeModel with the Stability AI Image Services request format (base64 input image plus per-service fields)

  • Video: Uses StartAsyncInvoke - requires S3 bucket for output

Design decisions

Decision

Why

Converse API for every text model on bedrock-runtime

One request and response shape across ten providers, so adding a text model is a JSON entry, not a new code path.

Gemma 4 through bedrock-mantle, behind the same options

Gemma 4 has no Converse API. Its requests are built by the same buildConverseInput (option checks, output cap, temperature rules) and only the transport changes, so compare and judge mix it with Converse models.

Model catalog as JSON + Zod

A malformed or half-edited registry fails at startup, not on the first call. Capabilities are a closed enum, so a typo like tool-use is rejected at load.

Per-model noTemperature flag

Claude Opus 4.7 and later, GPT-6.x, Grok 4.7 and Kimi K3 return a 400 for temperature; the flag drops the parameter for exactly those entries.

Measured effort and structured_output flags

Each was probed live on every registry model, so an unsupported option fails with the list of models that take it, before a billed call. Ids outside the registry pass through for Bedrock to judge.

Cached SDK client per region

Text and embeddings run in AWS_REGION, while the Stability and Luma models are only served from us-west-2.

Promise.allSettled in bedrock_compare

One model without access, or one throttled call, reports its own error while the rest still answer.

Blind, shuffled answers in bedrock_judge

LLM judges favor answers by position and by name. The judge sees only labels in a random order, and the tool maps them back to models afterwards.

Cancelled calls abort the Bedrock request

A client that gave up on a tool call should not leave the server waiting on, and paying for, an answer nobody reads. Video starts are left alone, since a cancelled start could leave a billed job with no ARN.

Weekly lifecycle check

AWS blocks new customers from Legacy models months before their EOL date, so scripts/check-lifecycle.mjs compares the registry with AWS's lifecycle page every week and on every registry change.

Image size read from the saved PNG

Stability models take an aspect ratio and choose the pixels, so echoing the requested size back would be wrong.

Edit services as data, checked before the call

Each service's required and optional fields live in image-edit-models.json, so bedrock_edit_image refuses a missing or unsupported field before Bedrock bills the image.

Input files checked by their first bytes

bedrock_edit_image reads local paths, so only a file that starts like a PNG, JPEG or WebP is sent to Bedrock.

Region from the job ARN is validated

The region picks the request host, so an unchecked value from a crafted ARN could send the signed request to another server.

Benchmark

Every text model in the registry on 2026-10-18 (the 0.5.0 catalog), same prompts, same machine, one run of scripts/benchmark.mjs. Nine of these models have since left the registry (see the 0.7.0 and 0.8.0 changelogs). The latency includes the network round trip from that machine to the region, so read it as a ranking within this run rather than a figure for your own setup. Raw per-call data is in docs/benchmark/results.json; rerun it on your own account with npm run benchmark (every call is billed at the model's normal rate), or re-render the table from saved data with node scripts/benchmark.mjs --from docs/benchmark/results.json.

Measured 2026-10-18 in us-east-1 by scripts/benchmark.mjs: 2 prompts x 3 runs per model, temperature 0 where accepted, 256-token output cap. 26 of 26 models answered, 156 of 156 calls succeeded. Latency is the median client wall-clock time for the whole call, sorted fastest first. "Hit cap" counts answers cut off at 256 tokens, which also caps their token counts.

Model

Provider

Median latency (s)

Median output tokens

Median output tokens/s

Hit cap

OK

Nova Lite

Amazon

1.13

56

50

0

6/6

Nova Micro

Amazon

1.18

59

46

0

6/6

Mistral Large 3 (675B)

Mistral

1.25

61

50

0

6/6

Qwen3 Coder Next

Qwen

1.29

71

37

0

6/6

Llama 4 Maverick 17B

Meta

1.35

87

65

0

6/6

Llama 4 Scout 17B

Meta

1.36

99

71

0

6/6

Nova Pro

Amazon

1.49

61

43

0

6/6

GPT-OSS 20B

OpenAI

1.50

206

114

2

6/6

Llama 3.3 70B Instruct

Meta

1.52

107

66

0

6/6

Nova 2 Lite

Amazon

1.55

80

53

0

6/6

GPT-OSS 120B

OpenAI

1.82

152

88

1

6/6

Devstral 2 (123B)

Mistral

1.88

63

35

0

6/6

Pixtral Large

Mistral

2.06

85

42

0

6/6

Claude Sonnet 4.6

Anthropic

2.10

77

36

0

6/6

Qwen3 VL 235B

Qwen

2.15

58

28

0

6/6

Mistral Small

Mistral

2.31

98

42

0

6/6

DeepSeek R1

DeepSeek

2.35

256

109

6

6/6

Claude Haiku 4.5

Anthropic

2.38

179

71

3

6/6

Claude Opus 4.7

Anthropic

2.40

98

39

0

6/6

Claude Sonnet 5

Anthropic

2.69

98

36

0

6/6

Claude Opus 4.8

Anthropic

2.79

111

39

0

6/6

Claude Opus 4.6

Anthropic

3.42

80

24

0

6/6

Claude Fable 5

Anthropic

4.18

134

30

0

6/6

Claude Opus 5

Anthropic

4.46

154

30

1

6/6

DeepSeek V3.2

DeepSeek

4.60

55

13

0

6/6

Claude Fable 5.1

Anthropic

5.61

127

21

0

6/6

Try a tool without an MCP client

scripts/call-tool.mjs spawns the built server over stdio, the same way Claude Code does, and calls one tool with AWS credentials from your shell:

npm run build
node scripts/call-tool.mjs                       # list the tools
node scripts/call-tool.mjs bedrock_ask '{"model":"nova-micro","prompt":"Say hi"}'

Troubleshooting

Error

Cause

Fix

The provided model identifier is invalid.

The ID does not resolve on Bedrock - usually a missing us. cross-region inference profile prefix, or a model that is not offered in your region.

Use a registry alias (bedrock_list_models), or add the us. prefix to the full ID.

`temperature` is deprecated for this model.

A Claude 4.7-or-later model was reached without a registry entry, so the server sent its default temperature of 0.7. Anthropic deprecated temperature / top_p / top_k for Claude Opus 4.7 and later, and a non-default value returns a 400 there. Earlier models (Haiku 4.5, Opus 4.6, Sonnet 4.6) still accept it.

Call it through a registered alias, or add a registry entry with "noTemperature": true in src/data/text-models.json.

<model> does not support structured output / does not take effort

This server's registry says the model rejects that option (measured 2026-10-21 and 2026-11-01).

Pick a model from the list in the message, or drop schema / effort.

This model doesn't support the outputConfig field.

A model outside the registry rejected schema or effort.

Drop the option for that model.

This model doesn't support the temperature field.

A model outside the registry that takes no temperature (GPT-6.x, Grok and Kimi do this) got the server's default 0.7.

Call it through a registered alias, or add a registry entry with "noTemperature": true.

Could not load credentials from any providers

No bearer token reached the server (or it was empty) and no IAM credentials resolved.

Put the real token in the MCP config's env, or configure an AWS profile, SSO or IAM role.

AccessDeniedException

Model access has not been granted to your account.

AWS Console > Bedrock > Model access, request the model, retry once it shows Access granted.

bedrock_video_status reports Failed

Bedrock returns the reason on the job itself.

Read the Reason: line printed under the status and fix what it names.

Error strings are surfaced verbatim from Bedrock. Note that the temperature one is a registry gap in this server, not an IAM or model-access problem - there is nothing to fix in AWS for it.

Contributing

See CONTRIBUTING.md for setup, the two required checks, and the rules for adding a model (registry data + a test assertion + a live probe of the model ID). Security issues go through the private route in SECURITY.md, not a public issue.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables MCP-compatible applications to manage and use local or cloud-hosted models, including listing, pulling, generating, chatting, embeddings, and web search/fetch through a unified tool interface.
      817 npm
      AGPL 3.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Exposes selected AWS resources — Lambda functions, S3, DynamoDB, SQS, EventBridge, OpenAPI endpoints and remote MCP servers — as narrowly scoped MCP tools that Claude Code agents running on Amazon Bedrock AgentCore can call, with IAM-enforced permissions on each target.
      80 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables generating AI videos, images, audio, and staged story episodes directly from Claude, ChatGPT, Cursor, Grok, or any MCP client, with 61 tools covering model catalogs, generation jobs, and media transforms such as upscaling, lipsync, motion control, and video editing. Hosted over Streamable HTTP with OAuth sign-in, so no API key or local installation is required.
      MIT