What Agent Backends Does SwarmForge Support? A Complete List of 8 LLM Providers

SwarmForge supports eight agent backends: OpenAI, Anthropic, Cohere, Grok, Groq, Google Vertex AI, Mistral, and a built-in "none" mock backend for CI/testing, selectable via the SWARMFORGE_AGENT_BACKEND environment variable.

SwarmForge's agent layer is designed for runtime flexibility, allowing developers to swap between large language model (LLM) providers without code changes. The backend selection system is implemented in Clojure/Babashka and normalizes provider names through a central mapping function.

How SwarmForge Agent Backend Selection Works

The backend configuration flows through a single entry point in swarmforge/scripts/swarmforge.bb. The normalize-agent-backend function (lines 45-52) handles case normalization and provider routing, making the system case-insensitive for user input.

The normalize-agent-backend Function

This core function transforms user-supplied backend names into canonical provider identifiers. It supports fuzzy matching and validates against the eight recognized backends.

Complete List of SwarmForge Agent Backends

Backend	Provider	Models Available
openai	OpenAI	GPT-3.5, GPT-4, GPT-4o, o1 series
anthropic	Anthropic	Claude 3 Opus, Sonnet, Haiku
cohere	Cohere	Command-R, Command-R+
grok	xAI	Grok-1, Grok-2
groq	Groq	Llama-3, Mixtral, Gemma
google	Google Vertex AI	Gemini 1.5 Pro, Gemini 1.5 Flash
mistral	Mistral AI	Mistral-7B, Mixtral 8x7B, Mistral Large
none	Mock/Stub	No external LLM (CI/testing mode)

The none backend returns hardcoded responses and executes no network calls, making it essential for automated test suites and offline development.

Configuring Your Agent Backend

Environment Variable Method

Set SWARMFORGE_AGENT_BACKEND before launching the runtime:

# Use OpenAI GPT models

export SWARMFORGE_AGENT_BACKEND=openai
swarmforge.sh run

# Use Anthropic Claude

export SWARMFORGE_AGENT_BACKEND=anthropic
swarmforge.sh run

# CI/testing with mock backend

export SWARMFORGE_AGENT_BACKEND=none
swarmforge.sh run

Programmatic Backend Override

Override the backend for scoped execution in Babashka scripts:

(require '[swarmforge.core :as sf])

(let [ctx (sf/context "/tmp/sandbox")]
  (sf/with-agent-backend ctx "anthropic"
    (sf/run-task ctx "analyze-codebase")))

The with-agent-backend macro temporarily binds the backend for the duration of the body execution, then restores the previous configuration.

Backend Validation in Test Suites

The test/swarmforge/pack_ui_test.clj file demonstrates production backend usage. The set-backend! helper function (lines 88-93) explicitly configures "grok" as the test backend:

;; From test/swarmforge/pack_ui_test.clj
(defn set-backend! [backend]
  (System/setProperty "swarmforge.agent.backend" backend))

;; Usage in test setup
(set-backend! "grok")

This confirms that grok is fully integrated into SwarmForge's recognized provider set, not merely documented but actively exercised in continuous integration.

Key Source Files for Agent Backend Implementation

File Path	Responsibility
swarmforge/scripts/swarmforge.bb	Core loader with normalize-agent-backend function; environment variable parsing
test/swarmforge/pack_ui_test.clj	Test suite validating backend selection, including Grok integration
README.md (repository root)	User-facing documentation of supported backends and configuration patterns

Backend-Specific Considerations

Rate Limits and Latency

Groq provides the lowest latency for open-weight models through dedicated inference hardware. Groq backends typically return responses 5-10x faster than API-direct alternatives for Llama-3 and Mixtral families.

Context Window Variations

Backend selection affects available context windows:

• Anthropic Claude 3: 200K tokens
• Google Gemini 1.5 Pro: 1M tokens (largest in SwarmForge's supported set)
• OpenAI GPT-4: 128K tokens
• **Cohere Command-R+: **128K tokens

The normalize-agent-backend function does not validate context requirements—your agent configurations must align with provider limits.

Authentication Patterns

Each backend expects credentials via provider-standard environment variables:

• OPENAI_API_KEY for openai
• ANTHROPIC_API_KEY for anthropic
• COHERE_API_KEY for cohere
• GROQ_API_KEY for groq
• GOOGLE_APPLICATION_CREDENTIALS for google
• MISTRAL_API_KEY for mistral

The none backend requires no authentication.

Summary

• SwarmForge supports eight agent backends: OpenAI, Anthropic, Cohere, Grok, Groq, Google Vertex AI, Mistral, and the mock none backend.
• Backend selection occurs through the SWARMFORGE_AGENT_BACKEND environment variable or the with-agent-backend Clojure macro.
• The normalize-agent-backend function in swarmforge/scripts/swarmforge.bb (lines 45-52) implements provider normalization.
• The none backend enables CI pipelines and offline development without API credentials.
• Grok integration is validated in test/swarmforge/pack_ui_test.clj through the set-backend! helper.

Frequently Asked Questions

How do I switch between agent backends in SwarmForge without restarting?

Use the with-agent-backend macro for temporary overrides within a single execution scope. For persistent changes, export SWARMFORGE_AGENT_BACKEND before launching swarmforge.sh. Runtime switching without environment changes is not supported—the backend binds at context initialization.

What happens if I specify an unsupported backend name?

The normalize-agent-backend function returns nil for unrecognized names, causing SwarmForge to fall back to the none mock backend with a warning logged to stderr. No fatal error is raised, but agent tasks will execute with stub responses.

Does SwarmForge support local LLM backends like Ollama or LM Studio?

The current normalize-agent-backend implementation in swarmforge/scripts/swarmforge.bb does not include Ollama or LM Studio in its case branches. However, the groq backend can target local GroqCloud deployments if configured with a custom GROQ_BASE_URL. Native local LLM support would require extending the normalization function.

Which backend is recommended for cost-sensitive production deployments?

Groq offers the best price-performance for open-weight models, particularly Llama-3 70B at roughly $0.59 per million tokens versus OpenAI's GPT-4 at $30 per million tokens. For proprietary model quality at lower cost, Google Gemini 1.5 Flash matches GPT-3.5 performance at roughly 20% of the price. The none backend eliminates all external costs for testing.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}