Bedrock Multi-Model

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 nine 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: 33 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.

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

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)

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 5.1 (access is gated per account; both failed the 2026-11-01 probe), 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-sol, gpt-terra) track the current model in that tier; version-pinned aliases (fable-5, 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 and Haiku 5.5 on 2026-10-07, so the bare claude-opus, claude-sonnet and claude-haiku now resolve to the 5.5 models.

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 Sol and Terra, Grok 4.7 and Kimi K3 ("This model doesn't support the temperature field", probed 2026-11-01). 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 5.1 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 for the models added then): all 10 Claude models, GPT-6.1 Sol and Terra, 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 Sol and Terra, Grok 4.7 and Kimi K3 (probed 2026-11-01). 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)

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.

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)

The output S3 bucket must be in us-west-2 (same region as the model).

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.

Embeddings (InvokeModel API)

bedrock_embed_similarity only sends text, so Titan Multimodal is exercised through its text path here and 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"      }    }  }}

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.gitcd bedrock-multi-model-mcpnpm installnpm 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

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.

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/"

Compare text similarity

"Use bedrock_embed_similarity to compare: 'a cat', 'a kitten', 'a car'"

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  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)    embed.ts        # InvokeModel wrapper (embeddings) + 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 registryscripts/  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

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

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. Eight of these models have since left the registry (see the 0.7.0 changelog). 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.

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 buildnode scripts/call-tool.mjs                       # list the toolsnode scripts/call-tool.mjs bedrock_ask '{"model":"nova-micro","prompt":"Say hi"}'

Troubleshooting

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