Salesforce Docs Portal

Publishing Custom LLM Providers to Anypoint Exchange

If your LLM provider or gateway is not one of the built-in Model Proxy providers, publish it to Anypoint Exchange as an llm-metadata.json classifier. Model Proxy can then route traffic to that provider. You create the classifier, confirm the asset version is free, and publish it with the Exchange API.

Before You Begin

Before getting started, make sure you have:

  • curl installed

    To install, see Install curl.

  • jq installed

    To install, see Install jq.

  • An Anypoint connected app that:

    • Acts on its own behalf with the client_credentials grant type.

    • Has the Exchange Contributor or Exchange Administrator scope assigned to the target business group.

Create the Classifier

The llm-metadata.json classifier is a JSON object. Only platform is required. Unknown properties aren’t accepted, so use only the documented fields.

Field Type Required Description

platform

String

Yes

Provider or platform identifier, such as Gemini, OpenAI, or a custom name.

name

String

No

Human-readable platform name.

inputFormat

String

No

Wire format: openai, anthropic, or gemini.

kind

String

No

Platform type, for example llm-gw.

authentication

Array

No

Supported provider authentication methods.

models

Array

No

Model ID strings or detailed model objects.

defaultUrl

Array

No

Named provider URL entries. Each entry requires both name and url.

capabilities

Object

No

Platform-wide supported features.

policyRef

Object

No

Adapter policy coordinates. Requires groupId and assetId. version is optional.

provenance

Object

No

Origin information. Supports id, createdAt, updatedAt, version, status, and url. Use RFC 3339 date-time strings for timestamps.

Authentication

Each authentication entry requires kind. Supported values are:

  • none

  • apiKey

  • amazonSigv4

Authentication entries can include non-secret defaults such as header names, token URLs, scopes, or timeouts. Never include actual API keys, passwords, client secrets, or tokens. If kind is none, it must be the only authentication entry.

Authentication Kind Optional Configuration

apiKey

headerName (required when a configuration is present), apiKeyPrefix, headersToRemove, and additionalHeaders

Models

The models field is an array. Each item can be either of these forms:

  • A string containing only the provider’s model identifier, such as "gemini-2.5-flash".

  • A detailed object. The id field is required. All other fields are optional.

Model Field Type Description

id

String

Provider-specific model identifier.

label

String

Human-readable model name.

description

String

Summary of the model and its intended use.

limits

Object

Context, output, and prompt-caching limits.

capabilities

Object

Features supported by this specific model.

pricing

Object

Prices per million tokens.

lifecycle

Object

Release and retirement information.

The nested objects support these fields:

Object Field Type Description

limits

maxInputTokens

Integer

Maximum number of input tokens.

limits

maxOutputTokens

Integer

Maximum number of generated tokens.

limits

supportsPromptCaching

Boolean

Whether prompt caching is supported.

limits

maxCachedTokens

Integer

Maximum number of tokens that can be cached.

pricing

inputPerMillion

Number

Price per million input tokens.

pricing

outputPerMillion

Number

Price per million output tokens.

pricing

cachedInputPerMillion

Number

Price per million cached input tokens.

pricing

reasoningPerMillion

Number

Price per million reasoning tokens.

lifecycle

status

String

Provider-defined state, such as active, deprecated, or sunset.

lifecycle

releasedAt

String

Release date or timestamp.

lifecycle

sunsetAt

String or null

Known retirement date, or null.

This example shows both model forms.

{
  "models": [
    "gemini-2.5-flash",
    {
      "id": "gemini-2.5-pro",
      "label": "Gemini 2.5 Pro",
      "description": "A multimodal model for complex reasoning tasks.",
      "limits": {
        "maxInputTokens": 1000000,
        "maxOutputTokens": 8192,
        "supportsPromptCaching": true,
        "maxCachedTokens": 500000
      },
      "capabilities": {
        "toolCalling": true,
        "structuredOutput": true,
        "imageInput": true,
        "reasoning": {
          "supported": true,
          "type": "adaptive",
          "effortLevels": ["low", "medium", "high"]
        }
      },
      "pricing": {
        "inputPerMillion": 1.25,
        "outputPerMillion": 5.0,
        "cachedInputPerMillion": 0.3125,
        "reasoningPerMillion": 5.0
      },
      "lifecycle": {
        "status": "active",
        "releasedAt": "2025-03-25T00:00:00Z",
        "sunsetAt": null
      }
    }
  ]
}

Capabilities

Declare capabilities at the platform level or within a detailed model. The boolean flags are:

  • embeddings

  • toolCalling

  • structuredOutput

  • codeExecution

  • citations

  • batchProcessing

  • imageInput

  • pdfInput

Reasoning support isn’t a boolean flag. Declare it with a nested reasoning object that describes whether the platform or model supports reasoning and how it’s controlled:

Field Type Required Description

supported

Boolean

Yes

Whether the platform or model supports reasoning.

type

String

No

The reasoning mode the provider uses, for example adaptive.

effortLevels

Array of strings

No

The reasoning effort levels that the provider accepts, for example low, medium, and high.

In this example, the model supports reasoning and accepts three effort levels:

"capabilities": {
  "reasoning": {
    "supported": true,
    "type": "adaptive",
    "effortLevels": ["low", "medium", "high"]
  }
}

If the platform or model doesn’t support reasoning, omit the reasoning object or set supported to false.

With the exception of a custom authentication method’s provider-specific configuration, nested objects don’t accept undocumented fields.

Example LLM Provider Asset

Create a file named llm-metadata.json:

{
  "platform": "Geminiano",
  "name": "Geminiano",
  "inputFormat": "gemini",
  "kind": "llm-gw",
  "authentication": [
    {
      "kind": "apiKey",
      "defaultConfiguration": {
        "headerName": "X-goog-api-key",
        "headersToRemove": ["Authorization"]
      }
    }
  ],
  "models": ["gemini-2.5-flash"],
  "defaultUrl": [
    {
      "name": "default",
      "url": "https://generativelanguage.googleapis.com/v1beta/"
    }
  ],
  "capabilities": {
    "toolCalling": true,
    "structuredOutput": true,
    "imageInput": true,
    "pdfInput": true,
    "reasoning": {
      "supported": true
    }
  }
}

Publish the LLM Asset with the Exchange API

  1. Set the coordinates for the asset you want to publish:

    export ORGANIZATION_ID="YOUR_ORGANIZATION_ID"
export GROUP_ID="YOUR_GROUP_ID"
export ASSET_ID="my-custom-llm"
export ASSET_VERSION="1.0.0"
export VERSION_GROUP="v1"
export ASSET_NAME="My Custom LLM"

    The organization ID identifies the business group that receives the asset. The group ID is the asset’s Maven group ID and is commonly the same value. Use a lowercase, URL-safe asset ID.

    Exchange versions are immutable. If a version already exists, publish a new semantic version.

  2. Export your credentials in the current shell or retrieve them from a secret manager. Don’t commit them to source control or paste them into documents.

    read -r -p 'Anypoint client ID: ' ANYPOINT_CLIENT_ID
read -r -s -p 'Anypoint client secret: ' ANYPOINT_CLIENT_SECRET
echo
export ANYPOINT_CLIENT_ID ANYPOINT_CLIENT_SECRET
    Shell history can expose values. Avoid commands such as export ANYPOINT_CLIENT_SECRET=literal-secret.

  3. Obtain a bearer token. The steps use OAuth 2.0 client credentials and the production Anypoint host, https://anypoint.mulesoft.com.

    TOKEN_RESPONSE=$(curl --fail-with-body --silent --show-error --location \
  --request POST \
  'https://anypoint.mulesoft.com/accounts/api/v2/oauth2/token' \
  --header 'Content-Type: application/x-www-form-urlencoded' \
  --data-urlencode "client_id=$ANYPOINT_CLIENT_ID" \
  --data-urlencode "client_secret=$ANYPOINT_CLIENT_SECRET" \
  --data-urlencode 'grant_type=client_credentials')

ANYPOINT_TOKEN=$(printf '%s' "$TOKEN_RESPONSE" | jq -er '.access_token')
export ANYPOINT_TOKEN

    Don’t print or log TOKEN_RESPONSE or ANYPOINT_TOKEN.

  4. Confirm your new provider version is available:

    curl --silent --show-error \
  --output /dev/null \
  --write-out 'HTTP %{http_code}\n' \
  --header "Authorization: Bearer $ANYPOINT_TOKEN" \
  "https://anypoint.mulesoft.com/exchange/api/v2/assets/$GROUP_ID/$ASSET_ID/$ASSET_VERSION"

    Make sure you receive a 404 error code.

    • HTTP 404: The version is available.

    • HTTP 200: The version already exists. Change ASSET_VERSION before publishing.

    • HTTP 401 or HTTP 403: The request wasn’t authorized. See Troubleshooting Error Codes.

  5. From the directory that contains llm-metadata.json, publish the asset to Exchange:

    curl --fail-with-body --silent --show-error --location \
  --request POST \
  "https://anypoint.mulesoft.com/exchange/api/v2/organizations/$ORGANIZATION_ID/assets/$GROUP_ID/$ASSET_ID/$ASSET_VERSION" \
  --header 'Accept: application/json' \
  --header 'x-sync-publication: true' \
  --header "Authorization: Bearer $ANYPOINT_TOKEN" \
  --form 'files.llm-metadata.json=@llm-metadata.json;type=application/json' \
  --form "properties.versionGroup=$VERSION_GROUP" \
  --form 'type=llm' \
  --form "name=$ASSET_NAME"

    Don’t set the multipart Content-Type header manually. curl --form generates the required boundary and header.

    If your file has a different name, replace @llm-metadata.json with @<YOUR_FILE_NAME>.

  6. Verify the published asset:

    curl --fail-with-body --silent --show-error \
  --header 'Accept: application/json' \
  --header "Authorization: Bearer $ANYPOINT_TOKEN" \
  "https://anypoint.mulesoft.com/exchange/api/v2/assets/$GROUP_ID/$ASSET_ID/$ASSET_VERSION" \
  | jq '{groupId, assetId, version, name, type, status}'

  7. Open the asset in Exchange and confirm that the published version is visible and that its llm-metadata.json classifier is downloadable.

  8. After testing, unset the sensitive shell variables:

    unset ANYPOINT_CLIENT_ID ANYPOINT_CLIENT_SECRET ANYPOINT_TOKEN TOKEN_RESPONSE
unset ORGANIZATION_ID GROUP_ID ASSET_ID ASSET_VERSION VERSION_GROUP ASSET_NAME

    Rotate any client secret you pasted into chat, logs, tickets, or shell history before using this process in automation.

Troubleshooting Error Codes

401 Unauthorized

Obtain a new token and verify the client ID and secret.

403 Forbidden

In Access Management > Connected Apps, add Exchange Contributor or Exchange Administrator for the target business group. Save the app, obtain a new token, and retry.

409 Conflict

The version already exists. Publish a new semantic version.

400 Bad Request

Check the JSON field names and types, type=llm, and properties.versionGroup.

See Also