Provider Configuration

Connecting your LLM providers

FlowMason is provider-neutral. Point it at any of the seven supported providers — or switch between them — without touching pipeline logic. All configuration lives in Custom Metadata.

How provider configuration works

Provider configuration is stored in LLMProviderConfig__mdt. The active record (the one with IsActive__c = true) is the default for all pipelines that don't specify a provider. You can override per-stage in pipeline config, or per-call from Apex.

Only one record should be active at a time. FlowMason reads the first active record it finds — activating a second record doesn't automatically deactivate the first.

Supported providers

Provider	ProviderName__c	Notes
Eden AI (default)	`edenai`	Smart router — picks best model automatically. Supports 30+ models from all providers via one API key.
Anthropic	`anthropic`	Claude 3.5, Claude 4 family. Best for reasoning-heavy tasks and long document analysis.
OpenAI	`openai`	GPT-4o, GPT-4o mini. Strong general-purpose performance and function calling.
Azure OpenAI	`azure_openai`	OpenAI models hosted in your Azure subscription. Meets data residency requirements.
Google Vertex AI	`google_vertex`	Gemini 2.5 Pro, Gemini Flash. Strong for multimodal and code generation.
AWS Bedrock	`bedrock`	Claude, Llama, Titan via AWS. Full VPC isolation available.
Ollama	`ollama`	Self-hosted local models. Zero data egress — everything stays on your infrastructure.

Setting up a provider

Go to Setup → Custom Metadata Types → LLM Provider Config → Manage Records → New, or deploy the following SFDX metadata:

XML — LLMProviderConfig__mdt

<?xml version="1.0" encoding="UTF-8"?>
<CustomMetadata xmlns="http://soap.sforce.com/2006/04/metadata"
                xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
    <label>Anthropic Claude</label>
    <values>
        <field>ProviderName__c</field>
        <value xsi:type="xsd:string">anthropic</value>
    </values>
    <values>
        <field>ModelName__c</field>
        <value xsi:type="xsd:string">claude-sonnet-4-6</value>
    </values>
    <values>
        <field>ApiKey__c</field>
        <value xsi:type="xsd:string">sk-ant-...</value>
    </values>
    <values>
        <field>IsActive__c</field>
        <value xsi:type="xsd:boolean">true</value>
    </values>
</CustomMetadata>

API key configuration for each provider

Configuration reference — all providers

// Anthropic
// ProviderName__c = 'anthropic'
// ApiKey__c       = 'sk-ant-...'

// OpenAI
// ProviderName__c = 'openai'
// ApiKey__c       = 'sk-...'

// Azure OpenAI
// ProviderName__c = 'azure_openai'
// ApiKey__c       = 'your-azure-key'
// BaseUrl__c      = 'https://your-resource.openai.azure.com/'
// ModelName__c    = 'gpt-4o'       (deployment name in Azure)

// Google Vertex AI
// ProviderName__c = 'google_vertex'
// ApiKey__c       = 'service-account-json-or-access-token'
// BaseUrl__c      = 'https://us-central1-aiplatform.googleapis.com/'
// ModelName__c    = 'gemini-2-5-pro-preview'

// AWS Bedrock
// ProviderName__c = 'bedrock'
// ApiKey__c       = 'access-key-id::secret-access-key'
// BaseUrl__c      = 'https://bedrock-runtime.us-east-1.amazonaws.com/'
// ModelName__c    = 'anthropic.claude-3-5-sonnet-20241022-v2:0'

// Ollama (self-hosted)
// ProviderName__c = 'ollama'
// BaseUrl__c      = 'http://your-server:11434/'
// ModelName__c    = 'llama3.2'
// ApiKey__c       = ''   (no key needed for local Ollama)

Production recommendation: Store API keys in Salesforce Named Credentials where possible. The MDT ApiKey__c path is convenient for development and sandbox orgs, but Named Credentials provide better security posture for production. Set ApiKey__c to callout:YourNamedCredential to route through a Named Credential.

Eden AI — the smart router

Eden AI is the default provider because it removes the "which model?" decision. Set ModelName__c = '@edenai' and Eden AI selects the best available model for each call based on task type, latency, and cost. You can also pin to a specific model through Eden AI:

Eden AI model selection

// Eden AI is the default. It routes to the best available model.
// To use a specific model via Eden AI:
// ModelName__c = 'anthropic/claude-sonnet-4-6'    (Claude)
// ModelName__c = 'openai/gpt-4o'                  (GPT-4o)
// ModelName__c = 'google/gemini-2-5-pro-preview'  (Gemini)
// ModelName__c = '@edenai'                         (smart router — default)

// In a pipeline stage config:
// "provider": "edenai",
// "model": "anthropic/claude-opus-4-7"

Switching providers

Switching the active provider is a metadata change — no code changes, no pipeline edits. This is what "provider-neutral" means in practice:

Apex — provider swap

// Change the active provider in FM_Config__mdt:
FMConfigAdmin.upsertConfig(new Map<String, Object>{
    'key'      => 'defaultProvider',
    'value'    => 'openai',
    'valueType'=> 'string',
    'category' => 'provider'
});

// Or override per-stage in pipeline JSON:
// "config": {
//     "provider": "anthropic",
//     "model": "claude-opus-4-7"
// }

// In a test — no permission needed:
FMConfig.setForTest('defaultProvider', 'openai', 'string');

Retry and timeout settings

All retry behavior is controlled through FM_Config__mdt. The defaults are sensible for most scenarios, but you can tune them for high-volume or latency-sensitive pipelines:

FM_Config__mdt — retry settings

// All retry + timeout settings are in FM_Config__mdt:
// providerMaxRetries       = 3       — retry count for transient failures
// providerBackoffBaseMs    = 1000    — base backoff before first retry (ms)
// providerBackoffMaxMs     = 30000   — ceiling for retry backoff (ms)
// providerBackoffMultiplier= 2.0     — exponential multiplier
// providerTimeoutMs        = 60000   — HTTP callout timeout (ms)

// Read and override in Apex:
Integer retries = FMConfig.getInteger('providerMaxRetries', 3);
FMConfig.setForTest('providerMaxRetries', '5', 'number');  // in tests

Key	Default	Description
`defaultMaxTokens`	2000	Default max_tokens for LLM calls
`defaultTemperature`	0.7	Default temperature (0 = deterministic, 1 = creative)
`providerMaxRetries`	3	Retry count for transient failures
`providerTimeoutMs`	60000	HTTP callout timeout (ms)
`providerBackoffBaseMs`	1000	Base wait before first retry
`providerBackoffMultiplier`	2.0	Exponential backoff multiplier

Cost tracking and pricing

FlowMason tracks token usage on every LLM call. Pricing rates are seeded in FM_Provider_Pricing__mdt and used to calculate cost attribution per execution:

Apex — cost calculation

// Calculate the cost of an LLM call:
Decimal cost = ProviderResponse.calculateCost(
    'anthropic/claude-sonnet-4-6',   // model identifier
    1500,                             // input tokens
    800                               // output tokens
);

// Pricing is seeded in FM_Provider_Pricing__mdt.
// Add or override rows via SFDX:
// Model_Name__c    = 'anthropic/claude-opus-4-7'
// Input_Cost__c    = 0.000015   (per token)
// Output_Cost__c   = 0.000075   (per token)
// Currency__c      = 'USD'

Add or update pricing rows by deploying FM_Provider_Pricing__mdt metadata. Rows are matched by exact Model_Name__c first, then by provider prefix fallback. See Governance & Audit for cost reporting queries.

Troubleshooting

HTTP 401 — authentication failed

The API key in ApiKey__c is invalid or expired. Update the key in the CMT record. API keys are never returned by FMProviderAdmin.listProvidersAdmin for security reasons — you must edit the record directly in Setup.

Timeout errors on large inputs

Lower defaultMaxTokens to reduce response time. Or pin to a faster model (e.g., anthropic/claude-3-5-haiku-20241022 or openai/gpt-4o-mini) via ModelName__c.

Multiple active provider records

FlowMason uses the first active record it finds. If you have multiple active records, deactivate all but one to ensure predictable routing. The admin UI (c-fm-provider-admin) shows all records and their active state.

What's next

Governance & Audit

Query cost reports, audit logs, and token usage by provider.

Unit Testing

Mock providers in tests so you never burn tokens running the test suite.