magpie-portkey-bedrock

A magpie provider for AWS Bedrock models through Portkey. It runs the real @ai-sdk/amazon-bedrock transport (native Bedrock Converse protocol) and works in both magpie (via magpie's pi package bridge) and pi.

The provider id is portkey-bedrock. Its models are available to every agent connected to magpie.

中文文档：README.zh-CN.md

Features

• Native Bedrock Converse streaming through @ai-sdk/amazon-bedrock.
• Dynamic model discovery: once an account is signed in, the provider asks the gateway for its model list (GET /v1/models) and merges what your account actually serves into the catalog; failures keep the catalog.
• Text, vision, streamed tool arguments and tool results, including image results and failed tools.
• Reasoning levels, including xhigh and max where declared; signed and redacted reasoning survives replay. Thinking streams with display: summarized, so the first token arrives while the model reasons instead of after the whole reasoning pass.
• Two explicit prompt-cache checkpoints (system prompt and final conversation message) — Bedrock only caches at explicit checkpoints, so this is what makes prompt caching actually work on Anthropic models.
• Input, output, cache-read and cache-write token accounting with catalog prices.
• Cancellation, request timeouts, payload inspection and response hooks.
• HTTP status and retry headers forwarded to magpie for its routing and failover logic.

Install in magpie

# npm (after publishing)
magpie plugin add @jack-yang/magpie-portkey-bedrock

# or GitHub
magpie plugin add github:yangwudong/magpie-portkey-bedrock

# or a local checkout
git clone https://github.com/yangwudong/magpie-portkey-bedrock.git
magpie plugin add "$PWD/magpie-portkey-bedrock"

Then sign in with your Portkey gateway API key (magpie plugin login portkey-bedrock, or the Providers page). magpie stores it in plugin-auth.json under its configuration directory and supplies the account's key for each request. Multiple accounts are supported.

Select Providers → Portkey Bedrock, then choose a portkey-bedrock/… model for your agent:

magpie claude portkey-bedrock/us.anthropic.claude-opus-5-5

magpie installs Bun, the plugin's runtime dependencies and its pi runtime when needed. To reload edits in the app, turn the plugin off and on in Plugins.

Configure your gateway

The package ships this public default in config.json:

{
  "baseUrl": "https://api.portkey.ai/v1",
  "provider": "aws-bedrock",
  "headers": {}
}

Create config.local.json beside index.ts to override it. That file is ignored by git and excluded from the npm package. Loading order is config.local.json, then config.json; each selects a complete configuration file. There are no environment variables — routing comes from these files, credentials come from magpie's sign-in.

• baseUrl: your Portkey gateway's Bedrock-compatible base URL. Requests go to /model/{encoded model id}/converse-stream.
• provider: optional x-portkey-provider route. Set this to your configured Portkey AWS provider slug, such as @your-aws-provider, when using a provider integration.
• configId: optional x-portkey-config configuration id. Use a route or configuration with AWS credentials already configured in Portkey.
• headers: additional non-secret gateway headers.

This matters for updates: reinstalling an npm/GitHub plugin replaces its directory, so a config.local.json created inside it is lost and must be recreated. If you install from a local folder (magpie plugin add /path/to/checkout), config.local.json lives in your own checkout and is never touched.

Credentials belong in magpie's sign-in — never in config files. The config reader rejects API keys and secret headers embedded in JSON (apiKey, virtualKey, authorization, *-api-key, token, secret, cookie). Requests send the selected account's key as both Authorization: Bearer … and x-portkey-api-key. Restart the plugin after configuration changes.

Requests are sent uncompressed. Portkey's deployed gateway reads the body as plain JSON and does not decompress: a gzip or deflate body answers 400 with Value null at 'messages' even at 95 bytes, on both /model/{id}/converse-stream and /v1/messages. Compressing the request needs the gateway to enable request decompression first. Responses are also sent uncompressed, and Accept-Encoding: gzip, br changes nothing.

Models

models.json is an editable catalog. Entries may be model ID strings or objects with id, name, reasoning, thinkingLevelMap, input, contextWindow, maxTokens and cost. Prices are USD per million tokens, with separate input, output, cacheRead and cacheWrite values.

The included catalog covers the current Bedrock line-up plus common Amazon / DeepSeek / Meta / Mistral models. Availability, limits and prices depend on your AWS account and Portkey route; update the catalog to match them. When an account is signed in, magpie refreshes the list through the provider's refreshModels: the gateway's /models response is merged with the catalog — catalog entries keep their metadata, unknown ids are added with conservative defaults. The catalog's name is what magpie's model picker displays (e.g. Claude Opus 5.5 instead of the raw Bedrock id); the id remains what requests use.

Use in pi

From a local checkout:

pi -e ./index.ts

Use /login for portkey-bedrock, or export the variable named by apiKeyEnv. Enable only one copy of this provider per session.

OpenCode plugin vs pi extension

This package is a pi extension: the real @ai-sdk/amazon-bedrock SDK runs here and speaks native Bedrock Converse, which is what enables the prompt-cache checkpoints, reasoning-effort mapping and signed-thinking replay above. A lighter alternative is an OpenCode provider plugin (magpie's gateway converts to chat completions and a tiny fetch adds the gateway headers) — simpler, but generic chat-completions traffic cannot carry Bedrock-specific features like explicit cache points. The two approaches use different provider ids and can be installed side by side.

Development and verification

Tested with pi 1.2.0 and magpie's Bun plugin host.

bun install --ignore-scripts
bun run check   # typecheck + tests

Tests use real AWS event-stream frames against the Bedrock SDK to verify authentication, conversion, tools, images, reasoning, cache markers, usage, Unicode chunk boundaries, aborts and failures. No live credentials are used.

For integration coverage:

git clone --depth 1 https://github.com/yetone/magpie.git /tmp/magpie
MAGPIE_SOURCE_DIR=/tmp/magpie bun test test/magpie.test.ts

This starts magpie's actual internal/plugin/host.js under Bun in a temporary configuration directory and sends Anthropic requests through the pi bridge. Without MAGPIE_SOURCE_DIR, that integration test is skipped. Live Portkey access is not part of the automated tests.

npm publish creates an installable archive containing the provider, public defaults, catalog and documentation. Dependencies, tests and local overrides are excluded.

Marketplace

Tag the repository magpie-plugin for automatic discovery under Plugins → Discover → Unofficial, on GitHub. For the curated list, PR an entry into magpie-community/plugins' registry.json after publishing to npm:

{
  "package": "@jack-yang/magpie-portkey-bedrock",
  "name": "Portkey Bedrock",
  "providers": ["portkey-bedrock"],
  "summary": {
    "en": "AWS Bedrock through Portkey, with tools, vision, reasoning, prompt caching and dynamic model discovery.",
    "zh": "通过 Portkey 使用 AWS Bedrock，支持工具调用、图像、推理、提示词缓存与动态模型发现。"
  }
}

Credits

Configuration layering, secret hygiene and the streaming core are based on TianZuo555/magpie-portkey (MIT). This package uses a different provider id (portkey-bedrock), so the two can be installed side by side.

License

MIT