Open-Source Providers

MIRASTACK ships two production-ready open-source AI providers covering the two leading frontier model APIs. Both are available under the AGPL v3 license and are built on the official vendor Go SDKs.

Provider Models License
openai GPT-4o, GPT-4, o1, o3, and all OpenAI-compatible endpoints AGPL v3
anthropic Claude Opus, Sonnet, Haiku (all generations) AGPL v3

Providers connect the engine to an AI inference backend. The engine’s internal LLM Router dispatches every LLM request to a registered provider over gRPC. Providers are the only components that make actual HTTP calls to external AI APIs — the engine itself never calls OpenAI or Anthropic directly.

openai — OpenAI Provider

Wraps the OpenAI Chat Completions API using the official OpenAI Go SDK. Supports both single-turn prompt completion and multi-turn agentic mode with full tool calling support.

Compatible backends: Any OpenAI-compatible API endpoint, including:

  • OpenAI API (api.openai.com)
  • Azure OpenAI Service (with appropriate endpoint URL)
  • LM Studio (local, OpenAI-compatible)
  • vLLM (OpenAI-compatible serving mode)
  • Ollama (OpenAI-compatible serving mode)

Configuration

miractl provider config-set openai endpoint https://api.openai.com/v1
miractl provider config-set openai api_key sk-...
miractl provider config-set openai model gpt-4o
Key Required Description
endpoint Yes API base URL. Change this to point at any OpenAI-compatible server.
api_key Yes API key. Stored encrypted in Kine.
model Yes Model name (e.g. gpt-4o, gpt-4-turbo, o1, o3-mini).

Supported models

Any model available through the configured endpoint. Tested with:

  • gpt-4o — recommended for agentic workflows
  • gpt-4o-mini — cost-optimised for high-volume tasks
  • o1, o3 — reasoning models for complex multi-step decisions
  • gpt-4-turbo — longer context window

Using as default provider

miractl settings set llm.default_provider openai

Source

mirastack-provider-openai on GitHub

anthropic — Anthropic Provider

Wraps the Anthropic Messages API using the official Anthropic Go SDK. Supports all Claude models with full tool calling support for agentic Lane 2 and Lane 3 workflows.

Configuration

miractl provider config-set anthropic endpoint https://api.anthropic.com
miractl provider config-set anthropic api_key sk-ant-...
miractl provider config-set anthropic model claude-sonnet-4-20250514
Key Required Description
endpoint Yes Anthropic API base URL.
api_key Yes Anthropic API key. Stored encrypted in Kine.
model Yes Model name.

Supported models

Any model available through the Anthropic API. Tested with:

  • claude-opus-4-* — most capable, best for complex agentic reasoning
  • claude-sonnet-4-* — recommended balance of capability and speed
  • claude-haiku-4-* — fastest and most cost-efficient for high-frequency calls

Using as default provider

miractl settings set llm.default_provider anthropic

Source

mirastack-provider-anthropic on GitHub

Running Multiple Providers

You can register both providers simultaneously and route different workflows to different providers:

# Register both
miractl provider list
# NAME        VERSION  STATUS
# openai      2.0.0    healthy
# anthropic   2.0.0    healthy

# Set the default for agentic loops
miractl settings set llm.default_provider anthropic

# A specific workflow step can override the provider in the YAML:
#   steps:
#     - name: analyze
#       type: llm
#       provider: openai

Installing the OSS Providers

All providers are published as Docker images at ghcr.io/mirastacklabs-ai/<provider-name>.

Running a Provider

# Set the engine gRPC address
export MIRASTACK_ENGINE_ADDR=engine-host:9090

# Start the provider
./mirastack-provider-openai
# or
docker run -e MIRASTACK_ENGINE_ADDR=engine-host:9090 \
           ghcr.io/mirastacklabs-ai/mirastack-provider-openai:latest

The provider self-registers with the engine automatically — no engine restart needed. Configure its API credentials at runtime:

miractl provider config-set openai api_key sk-...
miractl provider config-set openai model gpt-4o

Verify the provider is healthy:

miractl provider list
# NAME    VERSION  STATUS
# openai  2.0.0    healthy

Self-Hosted and Local LLM Backends

Both providers can be pointed at self-hosted or local inference servers. The openai provider works with any OpenAI-compatible API, making it the right choice for:

  • LM Studio — set endpoint to http://localhost:1234/v1, leave api_key empty
  • Ollama — set endpoint to http://localhost:11434/v1, leave api_key empty
  • vLLM — set endpoint to your vLLM server URL, set model to the loaded model name

This means MIRASTACK can run fully air-gapped with no external AI API calls — purely on your own hardware.

Building Your Own Provider

If you want to connect MIRASTACK to an inference backend not covered here, see the Provider Development Guide.

Back to top ↑