Skip to content
Wayland Docs

Wayland Core Overview

Wayland Core is the agent engine that powers the Wayland desktop app, shipped on its own as a command-line tool. It is a Rust LLM tool-use agent: it connects to a model provider, autonomously invokes local tools (file I/O, shell, search, sub-agents), and completes tasks end to end. The desktop app bundles it; this section covers running it directly from a terminal.

What it does

Section titled “What it does”
  • Multi-provider: Anthropic, OpenAI and compatibles (DeepSeek, Ollama), Google Gemini, AWS Bedrock, and Google Vertex AI native, with every other provider through the OpenAI-compatible layer.
  • ProviderCompat layer: provider quirks handled through configuration, not hardcoded branches.
  • ~60+ built-in tools: A universal core (Read, Write, Edit, Bash, Grep, Glob, plus orchestration: Delegate, Spawn, ToolSearch) is always present. Around 40 more cover git, SQL, web fetch, cloud CLI wrappers (kubectl, gcloud, aws_cli), API integrations (github_api, gitlab_api, linear_api, notion_api), pdf_extract, todo, and others. Around 15 integration tools are availability-gated under a NO-STUBS contract: they appear only when their key, binary, or feature is configured. See Built-in Tool Reference.
  • MCP client: connect to any Model Context Protocol server over stdio, SSE, or streamable-http, including dynamic injection at runtime.
  • Skills: named prompt snippets with variable substitution, shell expansion, conditional activation, and per-skill model or permission overrides.
  • Hooks: event-driven automation on the tool lifecycle (auto-format, lint, audit).
  • Sub-agent spawning: parallel task execution through the Delegate tool.
  • Sessions: save and resume conversation history.
  • Persistent memory: per-project memory with auto-indexing across sessions.
  • Plan mode: a read-only exploration mode for designing a plan before acting.
  • Context compression: three-tier automatic compaction (microcompact, autocompact, emergency), plus output compaction and TOON encoding.
  • Prompt caching: Anthropic cache_control for large cost reductions on repeated context.
  • Profiles: named profiles with extends for quick provider and model switching.
  • OAuth login: --login/--logout flags are present in the CLI surface. End-to-end functionality is not confirmed. See the caution note on the providers page before relying on this flow.
  • AGENTS.md injection: hierarchical loading of project instructions with @include support.

The subcommand map

Section titled “The subcommand map”

Beyond the default agent and REPL, the CLI exposes focused subcommands:

models, plugin, mcp-serve, swarm, workflow (alias forgeflows), project-context, init, acp, agent, cron, self-update, setup, and auth. See the CLI Subcommand Reference for each.

When to use Core over the desktop

Section titled “When to use Core over the desktop”

Reach for the engine when you want it headless: a one-shot prompt in a script, an agent step in CI, a long-running agent on a server, or the engine embedded in your own program over the JSON stream protocol. For channels, teams, scheduling UI, voice, and image generation, the desktop app is the fuller surface. Both read the same config and run the same loop.

Next

Section titled “Next”