Skip to content

Changelogs

0.0.44

Upgrade notes

  • Take a database snapshot before you upgrade: this release adds a migration (item 9). Stacks with AutoUpdateBRConnector set to true (ec2 and lambda compute types) pick up the new latest image automatically at 00:00 UTC.
  • Updating a stack to the 0.0.44 quick-build-brconnector.yaml template changes the UserData: an ec2 instance is stopped and started once, and lambda and lambda-web stacks created with KeepEc2 set to false must be recreated. See Template updates of existing stacks.
  • lambda-web stacks now build the deployment package on a stack update and run the migrations at start-up (item 2), including the 0.0.43 migrations if they were never applied, which delete the models of retired providers (see 0.0.43).
  • The alternative ECS template issues a new admin key on update; read it from the SSM parameter /brconnector/admin-key-<stack name>. Direct access to its ALB stops; use CloudFrontURL (item 6).
  • The Lambda@Edge ARN changes: set LambdaEdgeVersionArn to the new output (item 6).
  • Image models stay unbilled until price_per_image is set in their config (item 9).
  • bedrock-converse prompt-cache tokens are now billed (item 9).
  • openai-compatible and azure-openai streams now send stream_options. If your upstream or Azure API version rejects it, set "streamUsage": false in the model config (item 9).
  • Claude Code Web stacks: the template update does not rebuild the backend image. To get item 11, change ClaudeCodeWebImageBuildId (for example from 1 to 2) in the stack update; see Known limitations.

  1. CI - Unit tests, the build and the template checks now run on every pull request.

  2. lambda-web Updates - The deployment package is now built by CodeBuild. A stack update that changes SourceRepoUrl, SourceRef or the new WebFnBuildId parameter rebuilds it and deploys it as a new Web Function revision, and so does the first update to the 0.0.44 template. SourceRef may now be a commit. See AutoUpdate.
    - A failed build fails the update, names the failing step and leaves the running revision in place.
    - The rollback builds the previous SourceRef again, so keep that ref available until the update completes. If the rollback fails too, run aws cloudformation continue-update-rollback --stack-name <stack name> --resources-to-skip WebFnPackage.
    - The Web Function now applies the database migrations at start-up (SKIP_DB_INSTALL is no longer set), under one lock that covers the whole install.

  3. EC2 Boot Log and Run Script
    - The stack’s EC2 instance no longer writes keys or passwords to its boot log or console. New stacks write the BRConnector start command to the root-only /root/brconnector-run.sh (mode 0700), and the nightly update runs bash /root/brconnector-run.sh latest. Stacks created before 0.0.44 keep updating from the command recorded at first boot.
    - After the template update, the nightly update on existing ec2 stacks also makes the boot log and the generated key files readable by root only.
    - Admins of stacks created before 0.0.44 may rotate the admin key, stored in the SSM parameter /brconnector/admin-key-<stack name>, if other users can log in to the instance or read its console output. See Admin key on stacks created before 0.0.44.
    - The PGPassword parameter is now NoEcho.

  4. Regions - ec2, lambda and ecs now work in 17 Regions: us-east-1, us-east-2, us-west-1, us-west-2, ca-central-1, sa-east-1, eu-west-1, eu-west-2, eu-west-3, eu-central-1, eu-north-1, ap-northeast-1, ap-northeast-2, ap-northeast-3, ap-southeast-1, ap-southeast-2 and ap-south-1. lambda-web works only where Lambda Web Functions are available: us-east-1, us-east-2, us-west-2, ap-northeast-1 and eu-west-1 (see AWS Capabilities by Region for the current list), and the template refuses it elsewhere. Claude Code Web needs AgentCore Runtime, so it is not available in ap-northeast-3. See Supported Region.

  5. IAM Roles and Security Groups
    - The helper EC2 role, the cleanup function’s role and the nightly lambda update role now grant only the calls they make. The broad managed policies are removed: ECR, Lambda, ECS and EC2 full access, EC2 read-only, CloudWatch Logs full access and CodeBuild admin.
    - The cleanup function can terminate only this stack’s instances.
    - The nightly update output is still not written to S3.
    - The ECS load-balancer security group is created only for the ecs compute type.

  6. Alternative ECS Template and Lambda@Edge
    - The alternative ECS template generates its admin key from a cryptographic random source (32 characters). A stack update issues a new key; read it with aws ssm get-parameter --name /brconnector/admin-key-<stack name> --query Parameter.Value --output text. The AdminApiKey output is replaced by AdminApiKeyParameter.
    - Its ALB admits only CloudFront, and the LoadBalancerURL output is removed, so direct ALB access stops; use CloudFrontURL.
    - The Lambda@Edge function no longer logs request headers and runs on nodejs24.x. It publishes version v2 with a new ARN: set LambdaEdgeVersionArn of your BRConnector stack to the new LambdaEdgeVersionArn output.

  7. Image URLs - See the FAQ.
    - Image URLs in chat requests are fetched directly from the BRConnector host, not through HTTPS_PROXY or HTTP_PROXY, and only from public addresses. Every redirect is checked again, at most 3. The response must be image/*, at most 10 MiB, and arrive within 15 seconds.
    - Deployments without direct internet egress should send images to chat models as data: URLs, which work as before.
    - Other URLs, and unsupported schemes that used to send an empty image, now get HTTP 400.
    - Content blocks whose source has s3Location get HTTP 400; send the content inline.
    - nova-canvas takes http(s) URLs, or s3:// URLs in the model’s own s3Bucket, and answers a rejected image URL with HTTP 400.

  8. Shared-key Endpoint Removed - /open/get-shared-api-key is removed, so the BRClient portal no longer fills in a key: BRClient users enter their own key. A key named __shared-api-key__ is now an ordinary key.

  9. Billing
    - bedrock-converse now bills prompt-cache reads and writes at price_cache_read and price_cache_write from the model config (per token; price_in when unset), and counts them in the recorded input tokens. 1-hour cache writes bill at price_cache_write. Set both prices on your own converse models that use prompt caching. See bedrock-converse.
    - The migration 2026-10-23-converse-cache-prices.sql fills in the two prices on the claude-opus-5-5-chat, claude-sonnet-5-5-chat and claude-haiku-5-5-chat presets where they are absent: a cache read at 0.05 of price_in for Opus and Sonnet and 0.1 for Haiku, a cache write at 1.25 of it. A price an admin has set is kept.
    - openai-compatible and azure-openai streams now ask the upstream for usage and are billed once the stream ends, whatever the finish reason. The usage-only chunk reaches only clients that asked for it. The new config key streamUsage: false turns this off for an upstream that rejects stream_options.
    - bedrock-agent calls are now billed from the token usage in the agent’s trace. The trace is not sent to clients or logged.
    - Knowledge-base calls with bedrockModelArn are now billed from an estimate of 4 characters per token over the question, the answer and the references cited in the answer; references that are not cited are not counted.
    - painter, nova-canvas and azure-openai-image now record each image generation and bill it at the new config key price_per_image (per image). When it is unset the price is 0, so set it to bill image generation.
    - A recharge amount with more than 2 decimals gets HTTP 400, and a recharge is recorded in one transaction.

  10. Feishu Bots - Feishu bots now keep a single WebSocket client per connector and close replaced or removed ones.

  11. Claude Code Web - /ping now reports the time of the last real activity, so the AgentCore idle timeout can fire, and serve delivers SIGTERM to the server for a clean shutdown. The chat agent now runs Claude Code 2.1.310 (claude-agent-sdk 0.2.174, was 2.1.1): commands such as /mcp and /cost run and show their output instead of being merged into the next message, new sessions no longer make three Warmup calls, and the cost shown in the chat matches the BRConnector bill. Resuming a session id that has no transcript returns 404 instead of 500 #169. Existing stacks get this once a stack update changes ClaudeCodeWebImageBuildId (see the upgrade notes).

  12. Docs - AutoUpdate per compute type, how to update existing stacks, and a web search budget example in Claude Code Web.

  13. Dependencies - nodemailer 10.2.1 #165 and vite 6.4.5 #166. Claude Code Web: source-map-js 1.2.2 #156, nanoid 3.3.20 #157, postcss 8.5.31 #158, react-router and react-router-dom 7.18.4 #159, baseline-browser-mapping 2.11.31 #160, browserslist 4.29.4 #161, vite 7.3.7 with esbuild 0.28 #162, aiohttp 3.14.4 #163 and pyjwt 2.15.0 #167.

  14. lambda Compute Type Fixes
    - New lambda stacks could not be reached: function URLs created since October 2025 also need the lambda:InvokeFunction permission. The template now grants it, limited to calls through the function URL, and updating an existing lambda stack adds the permission.
    - With KeepEc2 set to false, the helper instance is now terminated after setup.

0.0.43

Upgrade warning - Take a database snapshot before you upgrade: the upgrade deletes the models of the retired providers (item 4) and adds preset models (item 2). Stacks with AutoUpdateBRConnector set to true (ec2 and lambda compute types) pick up the new latest image automatically at 00:00 UTC.

  1. Claude Code Web - Claude Code in the browser at /cc, served by BRConnector and running on Amazon Bedrock AgentCore Runtime, with one microVM and one S3 workspace per API key. Its model calls go back through /v1/messages with the user’s key, so quotas and billing apply. Turn it on with EnableClaudeCodeWeb in the CloudFormation template; see Claude Code Web. #155
    - "skipWebFetchPreflight": true is written into each user’s ~/.claude/settings.json when the key is missing. A value the user has set is kept.
    - The new parameter EnableWebSearch defaults to true, also when an existing stack is updated. With Claude Code Web on, in us-east-1, us-east-2, us-west-2, eu-west-1, ap-southeast-2 or ap-northeast-1, it creates an AgentCore Gateway with a Web Search target. Only the runtime role may call the gateway.
    - Searches are billed by AWS to the account that runs the stack (about USD 7 per 1,000 queries), outside BRConnector quotas. Consider an AWS Budgets alert, or set EnableWebSearch to false.
    - While the gateway is configured, Claude Code’s built-in WebSearch tool is hidden.

  2. Preset Models on Existing Databases - At start-up (on every compute type except lambda-web, see item 4), BRConnector applies each SQL migration in src/scripts/migrations/ that the database has not had yet. Preset models that an existing database lacks are added and bound to group 1; a model name that already exists is left unchanged. #153

  3. Model Presets - Seven GPT-6 and GPT-5.6 presets on bedrock-runtime-openai #152. claude-opus-5-5-chat, claude-sonnet-5-5-chat and claude-haiku-5-5-chat on bedrock-converse, now used by NoDB mode and knowledge-base summaries. The Claude 3.x presets and NoDB defaults are removed; existing models are left as they are. #154

  4. Retired Providers - aws-executor, urls-reader and web-miner are removed; use Claude Code Web (/cc) instead.
    - The upgrade migration deletes the models on these providers and their key and group bindings. History (threads and sessions) is kept.
    - A request to a model on one of these providers gets HTTP 400 that points to Claude Code Web (/cc). Once the model is deleted, the request gets the usual unknown-model answer, and so does a smart-router rule, localLlmModel or bot connector that names it.
    - lambda-web stacks are different. Updating the stack does not deliver 0.0.43: the deployment package is built only when the stack is created, and the function does not run migrations at start-up. To run 0.0.43, deploy a new lambda-web stack; its database is set up for 0.0.43 when the stack is created. To bring an existing database up to date by hand, take a snapshot, then apply the files in src/scripts/migrations/ in file-name order: 2026-10-20-preset-models.sql, 2026-10-20b-claude-chat-presets.sql and 2026-10-22-retire-providers.sql. The last one runs:

    BEGIN;
DELETE FROM eiai_key_model WHERE model_id IN (SELECT id FROM eiai_model WHERE provider IN ('aws-executor', 'urls-reader', 'web-miner'));
DELETE FROM eiai_group_model WHERE model_id IN (SELECT id FROM eiai_model WHERE provider IN ('aws-executor', 'urls-reader', 'web-miner'));
DELETE FROM eiai_model WHERE provider IN ('aws-executor', 'urls-reader', 'web-miner');
COMMIT;

  5. Model Id Override Removed - A model string name/<id> that is not itself a configured model name now gets HTTP 400, for every key. This also applies to Feishu bot, knowledge-base and smart-router model ids that contain /, and to azure-openai models. A model_id field in the request body is ignored. To use an upstream model id, configure a model for it; a configured model name may contain / (FAQ).

  6. Billing
    - Spend over the monthly quota is now drawn from the balance.
    - The balance can go below zero by at most the last calls. After that, calls are refused until a recharge, or until the next month when the monthly quota is above 0.
    - Charges are atomic, and a recharge adds to the stored balance.
    - A recharge reaches all instances within 60 seconds, through the cache.
    - The month_fee shown for a key keeps last month’s value until the key’s first charge in the new month.

  7. Session-Id - A Session-Id longer than 128 characters gets HTTP 400 on every route.

  8. User APIs
    - /user/model/list returns only name, provider, price_in, price_out, multiple and group_name.
    - /user/thread/list returns only the caller’s own threads, and limit is capped at 100 (on /admin/thread/list too).
    - orderBy on the session and payment lists accepts a fixed set of values and falls back to the default order for anything else.

  9. API Keys and Logs
    - New API keys are generated from a cryptographic random source.
    - SQL parameters and whole error objects are no longer logged, outside debug mode.

0.0.42

  1. bedrock-runtime-anthropic Provider - New provider for Claude Code and other Anthropic Messages clients. POST /v1/messages is passed through to Bedrock’s native /anthropic/v1/messages route, so system, tools, thinking, cache_control and context_management reach the model unchanged. It bills input, cache read, 5-minute and 1-hour cache writes, and output, and keeps each session on one region and credential so that prompt caching keeps hitting. Six Claude presets are added for new databases. #147

  2. bedrock-runtime-openai Provider - New provider that passes OpenAI Chat Completions requests through to Bedrock’s /openai/v1/chat/completions route, for OpenAI-format models such as Kimi K3 and GLM. Seven presets are added for new databases. #141

  3. Anthropic-format Errors on /v1/messages - Errors on /v1/messages now come back as {"type":"error","error":{...}} with a matching status, for example 401 authentication_error or 403 permission_error, so Anthropic SDK clients can handle them. Other routes are unchanged. #147

  4. AWS Lambda Web Functions - New lambda-web ComputeType in the CloudFormation template that runs BRConnector on Lambda Web Functions: no adapter, native response streaming, billing by active CPU, and requests up to 900 seconds. The endpoint defaults to IamAuth, so only the CloudFront distribution can reach it, and every request is also checked against a shared origin secret. The existing lambda ComputeType is unchanged. #145 #148

  5. Concurrency Fix: One Provider Instance per Request - Concurrent requests no longer share provider state or model data. #144

  6. bedrock-converse Fixes
    - temperature and top_p are no longer sent to Claude models that reject them (Opus 4.x and Claude 5.x). A ValidationException on a single-region model is no longer retried. #146
    - Bedrock API keys and AK/SK credentials are scoped to their own model. #138
    - The thinking parameter is read from the Anthropic request body. #122

  7. BRClient Portal - With no API key saved, the portal turns on BRConnector mode and opens the settings page. #142

  8. Container Runs as Non-root - Both Docker images now run as the node user, the health check no longer depends on curl (the container used to report unhealthy), and the Feishu bot writes its temporary files to the system temp directory. #150

  9. Build and Dependencies - devDependencies are kept out of dist/package.json #140. Dependency updates: #109 #110 #111 #114 #115 #116 #128 #132 #133 #134 #135 #136 #137 #143

Upgrade notes: preset models are inserted only into new databases. On an existing database, add the models you want in the admin UI. Admin Price-In / Price-Out are per 1 million tokens, while the cache prices in a model’s config are per token.

0.0.41

  1. Anthropic Messages API Compatibility - Added a new endpoint POST /v1/messages that is fully compatible with the Anthropic Messages API. Clients using the Anthropic SDK can now connect directly without modification. Supports text, tool use, thinking content blocks, and both streaming and non-streaming modes.

  2. x-api-key Header Support - Added support for the x-api-key header (Anthropic SDK default) in addition to the existing Authorization: Bearer header.

  3. Bug Fix: claude-opus-4 temperature/topP - Removed deprecated temperature and topP parameters for claude-opus-4 and later models, which caused invalid_request_error.

  4. Build Fix - Removed hardcoded local path alias for kui-vue in vite.config.js that caused build failures on non-author machines. Downgraded vite from experimental rolldown-vite to stable ^6.3.3.

0.0.40

  1. AWS Bedrock Bearer Token Support - Added support for AWS Bedrock API Key (Bearer Token) authentication method in addition to traditional AKSK credentials.

  2. Multi-tool Call Fix - Fixed an issue where multiple tool calls in a single response were not handled correctly in streaming mode.

0.0.38

  1. Bedrock App Profiles Support - Added support for Bedrock application profiles, allowing better configuration management for different deployment scenarios. #87

  2. PostgreSQL Database Upgrade - Upgraded standalone PostgreSQL database from version 16.3 to 16.10 for improved performance and security. #89

  3. Zero-Length Messages Fix - Fixed an issue where Bedrock would fail when processing zero-length messages. #90

  4. Tool Use Enhancement - Improved tool use functionality with bug fixes and stability improvements. #91 #92

  5. General Improvements - Various bug fixes and performance optimizations.

0.0.37

  1. bedrock-converse Provider Fix - Fixed an issue with Claude new models: ‘The model returned the following errors: temperature and top_p cannot both be specified for this model. Please use only one.’

  2. Frontend Caching Enhancement - Added 1-month caching for frontend pages served by koa-static-server.

0.0.36

  1. anthropic_beta Features Support - Added support for anthropic_beta features, including:
    - Compatible with Claude 3.7 Sonnet: token-efficient-tools-2025-02-19, output-128k-2025-02-19
    - Compatible with Claude Sonnet 4: context-1m-2025-08-07

These features are automatically integrated into the backend logic without requiring separate configuration. Reference: AWS Bedrock Anthropic Claude Parameters

  1. Thinking Mode Enhancement - In thinking mode, content is now set to empty string instead of null.

0.0.35

  1. New Provider: gemini - Added a new Google Gemini provider. Thanks to “daniexon” for the contribution.

  2. Enhancement: Tool Calls for openai_compatible Provider - The openai_compatible provider now supports passing tool calls to the model. Thanks to “bdavj” for the contribution.

  3. Format Mapping: Bedrock Converse finish_reason - Mapped Bedrock Converse finish_reason to OpenAI format for improved compatibility.

  4. New Parameter Support: max_completion_tokens - Added support for the max_completion_tokens parameter.

  5. New Docker Files - Added two new Dockerfiles.

0.0.34

  1. Bug Fixes - Fixed several minor bugs to improve system stability.

0.0.33

  1. New Provider: bedrock-agent - Added a new provider for Amazon Bedrock Agents. This is a simple implementation that allows you to interact with your Bedrock Agents. Documentation

  2. New Provider: azure-openai-image - Added support for Azure OpenAI image generation models. Documentation

  3. Removed Providers - Removed several deprecated providers (bedrock-claude3, bedrock-mistral, bedrock-llama3) as they can be fully replaced by the bedrock-converse provider.

  4. Docker Optimization - Removed AWS CLI commands from the Docker file to reduce image size.