Skip to content

Releases

v0.9.0

Choose a tag to compare

@Sagargupta16 Sagargupta16 released this 20 Nov 04:43
b632c2f

Added

  • bedrock_panel tool: 2-5 models discuss one prompt in turn and the tool returns the transcript. Unlike bedrock_compare and bedrock_judge, the models answer one after another and each sees what the earlier ones said.
    • roles gives each model one of six presets (critic, builder, fact-checker, devils-advocate, simplifier, moderator), free text up to 200 characters, or null. Every model is told who else is on the panel and in which role.
    • rounds (1 to 3, default 1) lets each model answer again having seen everything so far; summarize names a model that closes with where the panel agrees and disagrees; level (cheap, medium or intense) picks a default panel of three when models is absent.
    • A model's view is one user message holding the question and the turns so far, each prefixed with the speaker's name in square brackets, its own turns marked (you). What a model said never reaches another model's system prompt.
    • A model that fails loses its turn and the panel carries on, as in bedrock_compare. The same model may be listed twice only with different roles. max_tokens and effort work as in bedrock_compare.
    • The idea comes from the group-chat round in confab; no code or dependency was taken from it.
  • GPT-6.2 Astra (us.openai.gpt-6.2-astra) and GPT-6.2 Sol (us.openai.gpt-6.2-sol), on Bedrock since 2026-11-17. Both were live-probed on 2026-11-20: they take schema and effort from low to max, and reject temperature, so they are flagged noTemperature.
  • bedrock_doctor tool: one small billed call (a one-word prompt, at most 32 output tokens) to every registry text model, or only the models you pass, with concurrency 1 to 8 (default 4). The table gives one status per model: ok, throttled, denied, invalid or error, with the latency and a short note.
  • files option on bedrock_ask, bedrock_compare, bedrock_judge and bedrock_panel: local text files are added to the prompt as fenced blocks. Files are read only from under BEDROCK_MCP_FILES_ROOT (default the server's working folder, refused when that is a home folder or a drive root), at most 10 files, 256 KB each and 1 MB in total; names that look like secret stores (.env, .npmrc, .git-credentials, private keys, anything under .ssh or .aws and the like), network and device paths, binary files and anything outside the root are refused, and one refusal fails the call before any model is billed. The content is sent to the model provider.
  • Cost estimates: every text model in src/data/text-models.json now has an optional price (USD per 1M input and output tokens). bedrock_ask, bedrock_compare, bedrock_judge and bedrock_panel show an estimated cost, and bedrock_list_models has a price column. The prices are from the AWS Price List bulk files for us-east-1, read 2026-11-20, on-demand standard tier; for ids with the us. profile prefix the non-global price was used. Cache, batch and long-context pricing are not modeled, so a figure is an estimate.
  • level (cheap, medium, intense) on bedrock_ask, bedrock_compare and bedrock_judge, with the same presets as bedrock_panel. bedrock_ask uses the first model of the level; model on bedrock_ask and models on bedrock_compare and bedrock_judge are now optional when level is given, and models wins over level. bedrock_ask rejects model together with level.
  • One-click .mcpb bundle for Claude Desktop and other MCPB hosts: npm run pack:mcpb (scripts/pack-mcpb.mjs, manifest in mcpb/manifest.json) builds bedrock-multi-model-mcp-<version>.mcpb, and a bundle job in release.yml attaches it to the GitHub release after the publish job succeeds.
  • MCP prompts: second-opinion, compare-models and panel-discussion (ready-made instructions that call bedrock_ask, bedrock_compare and bedrock_panel).
  • conversation_id on bedrock_ask: opt-in follow-up conversations. Earlier turns of the same id are sent along as context. Kept in process memory only, forgotten 30 minutes after the last turn or when the server restarts, at most 20 turns per conversation, 50 conversations and 200,000 characters per conversation; the id is 1 to 64 letters, digits, - or _.
  • New-model watcher: npm run new-models (scripts/check-new-models.mjs) lists Bedrock model cards that the catalog neither has nor has deliberately skipped in scripts/model-skips.json. A new-models job in lifecycle.yml runs it weekly and on manual runs, and opens or updates one issue.
  • Luma Ray 2 keyframes on bedrock_generate_video: start_image_path and end_image_path turn local JPEG or PNG files (512x512 to 4096x4096 pixels) into the first and last frame. The images are checked before the job is started.
  • Image inputs on bedrock_embed_similarity: images takes local JPEG or PNG paths to embed beside the texts, for the multimodal models titan-multimodal and cohere. The tool needs at least 2 inputs and at most 50 across texts and images.

Changed

  • Breaking: bare gpt-astra now resolves to GPT-6.2 Astra and gpt-sol to GPT-6.2 Sol. The intense panel of bedrock_panel uses gpt-astra, so it moves with it.

Removed

  • Breaking: GPT-6 Astra (gpt-6-astra) and GPT-6.1 Sol (gpt-6.1-sol). AWS's cards name GPT-6.2 Astra and GPT-6.2 Sol as their next generation.

Security

  • npm audit went from 2 findings to 0, closing the 7 Dependabot alerts opened on 2026-11-13: hono 4.13.13 -> 4.13.19 (5 advisories) and @hono/node-server 2.1.5 -> 2.1.9 (2 advisories), both pulled in by @modelcontextprotocol/sdk. The stdio server never starts hono's HTTP server.
  • @aws-sdk/client-bedrock-runtime floor raised to ^3.1173.0.

v0.8.0

Choose a tag to compare

@Sagargupta16 Sagargupta16 released this 10 Nov 15:11
v0.8.0
b2e7094

Added

  • Claude Fable 5.2 (us.anthropic.claude-fable-5-2, on Bedrock since 2026-11-09), GPT-6 Astra (gpt-astra, the model AWS's card calls OpenAI's most capable) and GPT-6 Luna (gpt-luna, OpenAI's fast tier). Each was live-probed on 2026-11-10: all three take schema and effort from low to max, and reject temperature, so they are flagged noTemperature.

Changed

  • Breaking: bare fable and claude-fable now resolve to Fable 5.2. Fable 5.1 stays reachable as fable-5.1 and claude-fable-5.1.
  • @aws-sdk/client-bedrock-runtime floor raised to ^3.1165.0.

Removed

  • Breaking: Claude Fable 5 (fable-5), now two versions behind Fable 5.2.

Not added

  • Kimi K3.1 Pro - Highspeed (us.moonshotai.kimi-k3.1, on Bedrock since 2026-11-04): every probe call on 2026-11-10 was throttled ("Too many tokens"), on the US and the global profile, so it stays out until a call answers. kimi still resolves to Kimi K3.

v0.7.1

Choose a tag to compare

@Sagargupta16 Sagargupta16 released this 01 Nov 12:43
v0.7.1
383baa8

Changed

  • README setup installs from npm: claude mcp add ... -- npx -y bedrock-multi-model-mcp and an npx server block for Claude Code and Claude Desktop, instead of cloning and building. Building from source moved to "Run from source".
  • First release through release.yml with the npm trusted publisher (GitHub OIDC, no token), so this version carries npm provenance; 0.7.0 was published by hand.

v0.7.0

Choose a tag to compare

@Sagargupta16 Sagargupta16 released this 01 Nov 12:32
v0.7.0
7e93cd3

Added

  • bedrock_edit_image tool for the 13 Stability AI Image Services on Bedrock (us.stability.* profiles, us-west-2): inpaint, outpaint, erase, remove-background, search-replace, search-recolor, upscale-fast, upscale-conservative, upscale-creative, control-sketch, control-structure, style-guide, style-transfer.
    • The tool takes a local PNG, JPEG or WebP (image_path, plus mask_path or style_image_path where the service uses one) and saves the result as a PNG, the same way bedrock_generate_image does.
    • Each service's required and optional fields live in src/data/image-edit-models.json. A missing or unsupported field is refused before the billed call, and so is a file that is not one of those image formats.
    • Every operation was live-probed on 2026-11-01.
    • The three upscalers return 4x images whose PNG is 22 to 43 MB, over Bedrock's 16 MB response limit, so they default to WebP (about 2.5 to 5.5 MB). output_format overrides this per call.
  • Gemma 4 31B (gemma), Gemma 4 26B A4B (gemma-4-26b) and Gemma 4 E2B (gemma-4-e2b), served only on Bedrock's OpenAI-compatible bedrock-mantle endpoint.
    • A new "api": "mantle" registry field routes them to src/bedrock/mantle.ts, which calls Chat Completions at https://bedrock-mantle.<region>.api.aws/openai/v1. The request is built by the same buildConverseInput, so option checks, output caps and temperature rules match, and compare and judge mix them with Converse models.
    • Probed on 2026-11-01:
      • plain calls, temperature, system prompts and a JSON schema (response_format) all answered;
      • effort (reasoning_effort) takes low to xhigh, while max gets a 400 ("not supported with this model");
      • an abort at 400 ms settled at 411 ms.
    • These models need AWS_BEARER_TOKEN_BEDROCK. AWS also accepts SigV4 on mantle, but this server does not sign mantle requests and says so when the token is missing.
    • Grok 4.3, also mantle-only, is left out because Grok 4.7 supersedes it.
  • Nine text models, each live-probed on 2026-11-01 for a plain call, temperature, schema and effort before it was added:
    • GPT-6.1 Sol (gpt-sol) and GPT-6.1 Terra (gpt-terra), Grok 4.7 (grok) and Kimi K3 (kimi). All four take schema and effort and reject temperature ("This model doesn't support the temperature field"), so they are flagged noTemperature.
    • MiniMax M2.5 (minimax), Nemotron 3 Super 120B (nemotron), Ministral 3 14B (ministral) and Qwen3 Next 80B A3B (qwen-next). These take schema and temperature, not effort.
    • Magistral Small (magistral). It is not flagged structured_output: it accepted outputConfig.textFormat but ignored it, answering with fenced JSON and a different key.
  • Cohere Embed v4 (cohere.embed-v4:0, alias cohere, 1536 dimensions). Its response keys the vectors by type ({ float: [...] }) even when no type is requested, unlike v3's bare array; embed() reads both, and it also recognizes the us.cohere. inference profile.
  • scripts/check-lifecycle.mjs (npm run lifecycle) fails when a registry model is listed on AWS's public model lifecycle page. It needs no credentials. .github/workflows/lifecycle.yml runs it weekly and on PRs that touch src/data/. Against the 0.6.0 registry it flags exactly the five Legacy models below.
  • Cancellation: every tool except bedrock_generate_video passes the MCP request's abort signal to Bedrock, so a cancelled or timed-out tool call stops the request. Measured on 2026-11-01: a Nova Micro call aborted at 400 ms settled at 412 ms; the same call left to finish took 6.4 s. Video starts are excluded on purpose, because cancelling a sent StartAsyncInvoke could leave a billed job with no ARN to track.

Removed

  • Breaking: Llama 4 Maverick (llama4), Llama 4 Scout, Llama 3.3 70B, Devstral 2 and DeepSeek R1.
    • AWS moved all five to Legacy on 2026-09-30, with EOL on 2027-03-30, and "new customers can't use Legacy models". So the llama4 examples in the tool descriptions and README failed for new accounts.
    • AWS lists no Active Meta model, so Llama leaves the catalog.
  • Breaking: these Active but superseded models are removed:
    • Claude Opus 4.8 and 4.7 (claude-opus-4.8, claude-opus-4.7); Opus 5 and 5.5 cover them, and Opus 4.6 stays as the Opus that still takes temperature.
    • Mistral Small 2402 (mistral-small), now covered by Ministral 3.
    • Titan Text Embeddings V1 (titan-v1).
    • Cohere Embed English and Multilingual v3. Their aliases (cohere-en, cohere-english, cohere-multi, cohere-multilingual) now resolve to Embed v4.
  • The raw-HTTP bearer-token fallback (withBearerFallback, bedrockFetch) and its 7 tests.
    • With AWS_BEARER_TOKEN_BEDROCK set, the SDK already sends that token, ahead of any IAM keys, so the retry resent the same rejected credential and could never succeed.
    • Its bedrock_video_status branch also requested /async-invoke/<id> where the API takes the full ARN.
  • The Amazon image request shape and the image registry's format field. Every Active text-to-image model on Bedrock is Stability's (Nova Canvas reached EOL on 2026-09-30), so an image id outside the registry now gets the Stability shape too.
  • The rimraf dev dependency; npm run clean uses fs.rmSync.

Fixed

  • A set-but-empty AWS_BEARER_TOKEN_BEDROCK broke every call with "request could not be signed with token since the token is not defined".
    • This happens with an MCP config that passes "${AWS_BEARER_TOKEN_BEDROCK}" from a client whose own environment lacks the variable: the SDK picks bearer auth whenever the variable exists.
    • The server now drops an empty value at startup, so the IAM credential chain answers.
  • bedrock_compare reported success when every model failed; it now sets isError in that case. Partial failures still return the answers that came back.
  • Two images finished in the same millisecond overwrote each other. File names now end in an 8-character random suffix.
  • max_tokens on bedrock_ask, bedrock_compare and bedrock_judge accepted fractions, and seed on bedrock_generate_image accepted fractions and negatives, all of which Bedrock then rejected. The tool schemas now refuse them.
  • Titan embeddings ran one request after another (50 texts, 50 round trips in a row). They now run 8 at a time.
  • server.json described 26 text models; the registry has 33 now.

Security

  • npm audit went from 5 findings to 0, closing the 6 open Dependabot alerts:
    • body-parser 2.3.0 -> 2.4.0 (GHSA-27r3-jxmh-p828, high) and fast-uri 3.1.8 -> 3.1.9 (4 advisories), both through @modelcontextprotocol/sdk 1.32.2.
    • glob and minimatch left with rimraf.
  • Manifest floors raised to @aws-sdk/client-bedrock-runtime ^3.1161.0 and @modelcontextprotocol/sdk ^1.32.2.

Changed

  • One loadRegistry and indexByAlias in src/models.ts replace the four copies of registry loading and alias indexing. One errorResult in src/index.ts replaces six copies of the error branch. firstImage and saveImage in src/bedrock/image.ts are shared by image generation and editing.
  • Examples in the tool descriptions, README and scripts use kimi instead of llama4. The effort and schema descriptions list the models that take them.
  • release.yml skips npm publish when npm already has the version. The first version is published by hand (npm attaches a trusted publisher only to an existing package), so its tag now goes on to the MCP Registry step, and a rerun after a failed registry step no longer stops at npm. package.json repository.url uses the git+https form npm normalizes to, so npm publish no longer auto-corrects it.