Optional Providers

Optional provider support means the consuming app opts into provider dependencies. It does not mean this repository publishes a package per adapter.

Mock

No extra dependency is required:

{:inference, "~> 0.3.0"}

GeminiEx

Install the direct Gemini API SDK in the consuming app:

{:inference, "~> 0.3.0"},
{:gemini_ex, "..."}

Configure credentials in the SDK, then use Inference.Adapters.GeminiEx. This is a model endpoint. Gemini CLI is retired and is not a compatibility route for this adapter.

Agent Session Manager

Install Agent Session Manager in the consuming app:

{:inference, "~> 0.3.0"},
{:agent_session_manager, "~> 0.12.0"}

Use Inference.Adapters.ASM with a provider atom or session reference.

ASM reports :agent_session, so clients must set admitted_kinds: [:agent_session]. Antigravity is the current Google coding-agent SDK behind the ASM family; it is distinct from the direct Gemini API provided by GeminiEx.

The adapter supports normal query calls and managed streaming sessions when the installed ASM module exposes query/3, start_session/1, stream/3, and stop_session/1. String sessions are passed to ASM as :session_id while the provider atom remains the query target; pid sessions are treated as external sessions.

The ASM adapter is intentionally common-only and completion-only. ASM owns the run-path option gate (ASM.Options.validate/2), so the adapter does not call ASM's strict-common preflight from the outside; it maps the neutral request onto ASM options, locks completion_only: true, and rejects tool-bearing requests until ASM has a documented all-provider host-tool contract.

ASM recognizes five coding-agent providers. Inference admits only providers whose SDK feature manifest proves completion_only: true:

Amp, Antigravity, and Cursor remain valid ASM providers for agent-session work, but Inference refuses them before query or stream dispatch. This is deliberate: the semantic adapter never relaxes its completion-only profile to make a provider appear compatible.

A {:json_schema, %{schema: schema}} response format becomes ASM's :output_schema option (inline JSON for Claude, a materialized schema file for Codex). Schemaless {:json, :object} mode is refused, because ASM structured output is schema-driven.

Inference.capabilities/1 answers both the completion-only and JSON Schema questions. The claims are read at runtime from ASM's own feature manifest using ASM.ProviderFeatures.common_feature/2. A provider that declares no feature — or an application that has not installed ASM at all — reports :unknown rather than assumed support. Tests can inject a stand-in through the :asm_provider_features_module adapter option, the same way :asm_module overrides the runtime module.

ReqLlmNext

Install ReqLlmNext where broad hosted-provider coverage is needed:

{:inference, "~> 0.3.0"},
{:req_llm_next, "..."}

Use Inference.Adapters.ReqLlmNext.

ReqLLM Compatibility

Inference.Adapters.ReqLLM is for compatibility with existing users. Prefer ReqLlmNext for new broad hosted-provider work.

{:inference, "~> 0.3.0"},
{:req_llm, "~> 1.10"}

The compatibility adapter supports text generation, structured object generation through generate_object/4, provider key aliasing for OpenAI, Gemini, and Anthropic, and portable tool structs that expose :name, :description, :input_schema, and :run.

For standalone clients, the compatibility adapter can still use provider-local env configured by the consuming application. For governed clients, env fallback is skipped and direct provider keys are rejected before adapter dispatch.

Missing Dependencies

If an adapter is selected but its underlying provider module is unavailable, the adapter returns Inference.Error with category :missing_dependency.