feature-flags.md

feature flags

The umbrella crate agentkit re-exports subcrates behind feature flags.

Flags

• core
  • enables agentkit-core
• compaction
  • enables agentkit-compaction
  • implies core
• capabilities
  • enables agentkit-capabilities
  • implies core
• context
  • enables agentkit-context
  • implies core
• tools
  • enables agentkit-tools-core
  • implies capabilities
• task-manager
  • enables agentkit-task-manager
  • implies tools
• loop
  • enables agentkit-loop
  • implies tools, task-manager
• mcp
  • enables agentkit-mcp
  • implies capabilities, tools
• plugins
  • enables the information-first agentkit-plugins package loader
  • does not imply skill or MCP runtimes
• acp
  • enables agentkit-acp
  • implies loop
• adapter-completions
  • enables agentkit-adapter-completions
  • implies loop
• provider-baseten
  • enables agentkit-provider-baseten
  • implies adapter-completions
• provider-anthropic
  • enables agentkit-provider-anthropic
  • implies loop (the Messages API is not OpenAI-compatible, so this adapter does not go through adapter-completions)
• provider-cerebras
  • enables agentkit-provider-cerebras
  • implies loop (carries enough provider-specific surface — compression, version-patch header, reasoning config, rate-limit snapshot, Files + Batch — that it implements ModelAdapter directly instead of routing through adapter-completions)
• provider-groq
  • enables agentkit-provider-groq
  • implies adapter-completions
• provider-mistral
  • enables agentkit-provider-mistral
  • implies adapter-completions
• provider-ollama
  • enables agentkit-provider-ollama
  • implies adapter-completions
• provider-openai
  • enables agentkit-provider-openai
  • implies adapter-completions
• provider-openrouter
  • enables agentkit-provider-openrouter
  • implies adapter-completions
• provider-vllm
  • enables agentkit-provider-vllm
  • implies adapter-completions
• reporting
  • enables agentkit-reporting
  • implies loop
• tool-fs
  • enables agentkit-tool-fs
  • implies tools
• tool-shell
  • enables agentkit-tool-shell
  • implies tools
• tool-skills
  • enables agentkit-tool-skills
  • implies tools
• tool-compose
  • enables agentkit-tool-compose
  • implies tools

Default flags

The current default set is:

• core
• capabilities
• tools
• loop
• reporting

Typical combinations

Minimal orchestration:

• core
• capabilities
• tools
• loop

Coding agent:

• core
• capabilities
• context
• tools
• loop
• tool-fs
• tool-shell
• reporting

MCP-enabled agent:

• everything above
• mcp

ACP-exposed agent (editor-addressable over the Agent Client Protocol):

• everything needed for the host
• acp

OpenRouter-backed example host (streaming, prompt caching):

• everything needed for the host
• provider-openrouter

OpenAI-compatible provider (streaming; e.g. Baseten, Groq, Mistral, vLLM, Ollama):

• everything needed for the host
• provider-baseten / provider-groq / provider-mistral / provider-vllm / provider-ollama

Anthropic Messages API (streaming, extended thinking, server tools):

• everything needed for the host
• provider-anthropic

Cerebras Inference API (streaming, reasoning, rate-limit snapshot):

• everything needed for the host
• provider-cerebras
• plus, on agentkit-provider-cerebras directly, any of:
  • compression (msgpack + gzip request bodies)
  • predicted-outputs
  • service-tiers
  • batch (Files + Batch API)
  • experimental (umbrella for the three preview flags above)