Boomi Muse Connector#

Built from the latest commit on main. The one-zip download holds the archive, the descriptor and install notes; upload the archive and the descriptor to your Boomi connector group.

Docs site: Muse Connector, with the latest connector download.

A native Boomi connector for the Muse API (Meta Model API). Send prompts from a Boomi process and get completions back as documents, with token usage and estimated cost attached to every response.

This exists because the generic HTTP connector makes you hand-roll everything: auth headers, retries, response parsing, per-document error handling, and usage accounting. This connector does that once, properly, and gives you a real connection screen instead of a pile of headers.

%%{init: {"theme": "base", "fontFamily": "Josefin Sans, trebuchet ms, verdana, arial, sans-serif", "themeVariables": {"fontSize": "15px", "background": "#FBF6EE", "primaryColor": "#FBF6EE", "primaryTextColor": "#0B0B0B", "primaryBorderColor": "#9A7B3F", "lineColor": "#9A7B3F", "textColor": "#0B0B0B", "secondaryColor": "#9FDED6", "tertiaryColor": "#FBE7A1", "edgeLabelBackground": "#FBF6EE", "clusterBkg": "#FBF6EE", "clusterBorder": "#9A7B3F", "actorBkg": "#F7A8C4", "actorBorder": "#C2185B", "actorTextColor": "#0B0B0B", "actorLineColor": "#9A7B3F", "signalColor": "#0B0B0B", "signalTextColor": "#0B0B0B", "labelBoxBkgColor": "#FBE7A1", "labelBoxBorderColor": "#9A7B3F", "labelTextColor": "#0B0B0B", "loopTextColor": "#0B0B0B", "noteBkgColor": "#FBE7A1", "noteBorderColor": "#9A7B3F", "noteTextColor": "#0B0B0B", "activationBkgColor": "#9FDED6", "activationBorderColor": "#0F766E", "sequenceNumberColor": "#FBF6EE"}}}%%
flowchart LR
    P["Prompt document<br/>JSON"] --> M["Muse connector<br/>Chat Completion"]
    M -->|"/v1/chat/completions<br/>Bearer key"| API[("Meta Model API<br/>api.meta.ai")]
    API -->|"completion + usage"| M
    M --> R["Response document<br/>content, usage, cost"]
    R --> D{"Route on it"}
    D -->|profile| MAP["Map content<br/>into your target"]
    D -->|properties| PROP["Branch on<br/>tokens / cost / model"]
    class M focus
    class API external
    class D decision
    class PROP highlight
    classDef focus fill:#F7A8C4,stroke:#C2185B,stroke-width:2px,color:#0B0B0B
    classDef decision fill:#9FDED6,stroke:#0F766E,stroke-width:2px,color:#0B0B0B
    classDef highlight fill:#FBE7A1,stroke:#9A7B3F,stroke-width:1.5px,color:#0B0B0B
    classDef external fill:#FFFFFF,stroke:#0F766E,stroke-width:1.5px,stroke-dasharray:5 3,color:#0B0B0B
    classDef error fill:#FFFFFF,stroke:#D03B3B,stroke-width:2px,color:#0B0B0B

Contents: Quick start · Build · Deploy in Boomi · Connection fields · Operations · Input and output · Using results in a process · Budgets · Troubleshooting · Project layout

Quick start#

1. Get the files. Download muse-connector-1.0.0.zip from the dist/ folder. It holds the connector archive, ready to upload — no build needed. You can also build locally (see Build).
2. Upload it. In Boomi: Manage → Connector Groups → your group → Version upload, then Version change to activate. No process redeploy needed.
3. Get an API key. Create one at dev.meta.ai, under API keys.
4. Connect. New Connection component → Muse → paste the key. Test Connection hits GET /v1/models; only a rejected key fails it.
5. Run. Connector step → Chat Completion → send {"prompt": "Hello"} → map content out of the response.

Build#

Requirements: JDK 17 (compiles to Java 8 bytecode). No Gradle, no Maven — one script:

./build.sh

This fetches dependencies on first run, compiles, lints the descriptor against the SDK schema, and writes build/distributions/muse-connector.zip.

Run the tests (fake HTTP transport; no sockets):

javac --release 8 -encoding UTF-8 -nowarn \
  -cp "build/classes:build/deps/*" \
  -d build/test-driver/classes \
  src/test/java/com/boomi/connector/muse/MuseConnectorTest.java
java -cp "build/test-driver/classes:build/classes:build/deps/*" \
  com.boomi.connector.muse.MuseConnectorTest

57 checks, all passing: retry loop, parsing, budgets, error-message extraction, dry run, debug logging, tracked properties, list-models, and all three operations.

Build on GitHub (no local setup)#

• Download: every push to main rebuilds the connector; open the run → Artifacts → muse-connector for the fresh zip and descriptor.
• Release: push a tag such as muse-connector-v1.0.0; the zip is attached to a GitHub Release.
• The dist/ folder always holds the latest released zip, committed by the workflow.

Deploy in Boomi#

Upload the zip as a new connector version (connector group → version upload → version change to activate), then create a Connection component with your API key. Activating a new version needs no process redeploy.

Connection fields#

Main page

Performance page

Budgets page

Budgets are per execution, not per day. When one trips, the remaining documents are skipped with a clear BUDGET_EXCEEDED note instead of spending more. Nothing already completed is thrown away.

Under bulk concurrency a few requests can be in flight when a budget trips, so think of budgets as guardrails: no new request starts past the limit, but an execution can overshoot by a small margin.

Operations#

All three are EXECUTE type. Every knob resolves as document > operation field > connection: set a default temperature on the operation and override it in one document.

• Chat Completion — one input document in, one API call, one output document out. Documents run in order.
• Bulk Chat Completion — one API call per input document, sent with bounded parallelism (you set how many at once). Results come back in document order. A failed document is recorded against itself; the rest still run.
• List Models — asks the API which models your key can reach and returns them as one document ({"models":[{"id":...}]}). The input document is ignored; send one to trigger the call. Useful when you're deciding what to put in the connection's model field.

Operation fields (chat operations): System prompt, Temperature (0 focused, up to 2 adventurous), Max tokens (caps the response length), Response format (Text or JSON — JSON mode works best when the prompt itself mentions JSON), Dry run (build the request and return it as the document without calling the API; no key needed, no budget consumed — the response profile doesn't apply to dry-run output).

Input and output#

Send a JSON document. The simple form:

{
  "prompt": "Summarize this release note in two sentences.",
  "system": "You are a terse technical writer.",
  "temperature": 0.3,
  "maxTokens": 200,
  "responseFormat": "json"
}

Or pass a full conversation through untouched:

{
  "messages": [
    { "role": "system", "content": "You are a terse technical writer." },
    { "role": "user", "content": "Summarize this release note in two sentences." }
  ],
  "model": "muse-spark-1.3"
}

The output document, one per input:

{
  "id": "chatcmpl-abc123",
  "model": "muse-spark-1.3",
  "content": "…the model's reply…",
  "finishReason": "stop",
  "usage": {
    "promptTokens": 42,
    "completionTokens": 118,
    "totalTokens": 160
  },
  "estimatedCostUsd": "0.002528"
}

The Import wizard offers the request and response shapes as JSON profiles, so you usually don't have to type them.

Using results in a process#

There are three ways to act on a result. Use whichever suits each shape.

%%{init: {"theme": "base", "fontFamily": "Josefin Sans, trebuchet ms, verdana, arial, sans-serif", "themeVariables": {"fontSize": "15px", "background": "#FBF6EE", "primaryColor": "#FBF6EE", "primaryTextColor": "#0B0B0B", "primaryBorderColor": "#9A7B3F", "lineColor": "#9A7B3F", "textColor": "#0B0B0B", "secondaryColor": "#9FDED6", "tertiaryColor": "#FBE7A1", "edgeLabelBackground": "#FBF6EE", "clusterBkg": "#FBF6EE", "clusterBorder": "#9A7B3F", "actorBkg": "#F7A8C4", "actorBorder": "#C2185B", "actorTextColor": "#0B0B0B", "actorLineColor": "#9A7B3F", "signalColor": "#0B0B0B", "signalTextColor": "#0B0B0B", "labelBoxBkgColor": "#FBE7A1", "labelBoxBorderColor": "#9A7B3F", "labelTextColor": "#0B0B0B", "loopTextColor": "#0B0B0B", "noteBkgColor": "#FBE7A1", "noteBorderColor": "#9A7B3F", "noteTextColor": "#0B0B0B", "activationBkgColor": "#9FDED6", "activationBorderColor": "#0F766E", "sequenceNumberColor": "#FBF6EE"}}}%%
flowchart TD
    IN["Prompt documents<br/>one or many"] --> MC["Muse: Chat / Bulk Chat"]
    MC -->|Application Error| ERR["Error handling<br/>BAD_INPUT, budget, API error"]
    MC --> OK["Response documents"]
    OK --> USE{"Use the result"}
    USE -->|map| A["Option A: response profile<br/>map content and usage"]
    USE -->|branch| B["Option B: document properties<br/>route on tokens, cost, model"]
    USE -->|report| C["Option C: tracked properties<br/>Process Reporting"]
    class MC focus
    class USE decision
    class B highlight
    class ERR error
    classDef focus fill:#F7A8C4,stroke:#C2185B,stroke-width:2px,color:#0B0B0B
    classDef decision fill:#9FDED6,stroke:#0F766E,stroke-width:2px,color:#0B0B0B
    classDef highlight fill:#FBE7A1,stroke:#9A7B3F,stroke-width:1.5px,color:#0B0B0B
    classDef external fill:#FFFFFF,stroke:#0F766E,stroke-width:1.5px,stroke-dasharray:5 3,color:#0B0B0B
    classDef error fill:#FFFFFF,stroke:#D03B3B,stroke-width:2px,color:#0B0B0B

Option A: the response profile (recommended for maps)#

On the operation, click Import. You get a JSON profile with content, finishReason, usage/promptTokens and friends. Use Profile Element parameters in shapes:

Option B: dynamic document properties (no profile needed)#

Every output document carries these as document properties — read them with Document Property → Dynamic Document Property in Decision, Route, Map, or Notify shapes:

Handy when you want to branch without a map: a Decision shape can route on museTotalTokens Greater Than 4000, or kill an expensive loop when museEstimatedCostUsd climbs.

Option C: tracked properties (Process Reporting)#

The same values are declared as tracked properties, so they're visible and searchable in Process Reporting without any extra shapes:

Budgets#

See Connection fields — budgets are per execution. One more pattern worth knowing: put a cheap model on the connection for bulk classification work and override the model per document for the few that need the strong one. The museEstimatedCostUsd property makes the spend visible per document either way.

Pooling, honestly#

The pool lives in the Atom's JVM, one per connection, shared by every execution on that node. A cluster of three nodes has three pools — there is no such thing as a cross-node pool, and this connector doesn't pretend otherwise. What you do get: no per-request connection setup, idle connection eviction, and per-connection pool sizing.

Troubleshooting#

Project layout#

muse-connector/
├── README.md                  # this file
├── build.sh                   # the whole build: compile, lint, package
├── dist/
│   └── muse-connector-1.0.0.zip   # released archive, ready to upload
└── src/
    ├── main/java/com/boomi/connector/muse/   # the connector
    │   ├── MuseConnector.java       # entry point, operation routing, pool cache
    │   ├── MuseConnection.java      # connection fields, validation, defaults
    │   ├── MuseClient.java          # HTTP, retries, parsing, debug logging
    │   ├── MuseBaseOperation.java   # request building, budgets, tracked props
    │   ├── MuseExecuteOperation.java    # Chat Completion
    │   ├── MuseBulkChatOperation.java   # Bulk Chat Completion
    │   ├── MuseListModelsOperation.java # List Models
    │   └── MuseBrowser.java         # profiles, Test Connection
    ├── main/resources/META-INF/
    │   └── connector-descriptor.xml # the Boomi UI: fields, pages, operations
    └── test/java/…/MuseConnectorTest.java  # 57 checks, fake HTTP transport

Notes#

• SDK 2.31.0, strict class loading. connector-sdk-api is provided by the runtime and intentionally left out of the archive.
• A wrong API key fails the whole execution loudly (ConnectorException). A bad document, a rate limit that survives retries, or a tripped budget fails just that document and the rest keep going.
• Built with the boomi-connectordev skill. Where this project deviates from its supported path (javac instead of Gradle), the reason is written down at the deviation, not just here.