The OrchestratorOverviewHow It Connects
The StudioOverviewExtensionsHow It Works
ResourcesBlogFAQAbout
Get in touch
Studio Extensions

Local tools, not cloud dependencies.

Every extension runs on your machine. Code indexing, document processing, terminal management, application observation — your agents get powerful tools without sending data to third-party services.

Code Intelligence

Structural code understanding without dumping your repo into the context window.

SCIP-based indexing builds a local graph of symbols, references, and call chains. Agents query precise answers instead of scanning thousands of lines.

  • Find all references, get definition, search by meaning
  • Interactive graph explorer (Graphology + Sigma.js)
  • Semantic vector search with local embeddings (LanceDB)
  • Cross-repo federation — see dependencies across projects
SCIP indexingGraph navigationSemantic searchCross-repo20 languages

Documents

Read, write, and convert documents locally. No cloud upload required.

PDF, DOCX, XLSX, PPTX — processed on your machine. Agents get structured data, not raw file dumps.

  • Tier 1: Local JS libraries — fast, lightweight extraction
  • Tier 2: Docling deep parsing — complex layouts, scanned pages, OCR
  • Tier 3: Remote Document Server — when you need it (opt-in)
  • Drop files into any chat — auto-routed to the right processor
PDF, DOCX, XLSX, PPTX3 processing tiersLocal firstFile drop routing

Terminal

Agents run your project's commands instead of guessing shell syntax.

Auto-detects scripts and exposes them as structured tools. Agents describe intent, the plugin resolves the command.

  • Detects from package.json, Makefile, docker-compose.yml
  • Streaming output via NDJSON — progress, warnings, results
  • Process lifecycle management — start, monitor, stop
  • Interactive shell sessions via xterm.js and node-pty
Command detectionProcess managementStreaming outputInteractive shells

Witness

Agents see your running app — without expensive screenshots.

Three observation tiers, from most token-efficient to most visual. Agents verify UI state, debug interactions, and record proof.

  • Tier 1: DOM extraction — structured data from web apps, ~50× cheaper than screenshots
  • Tier 2: OS accessibility tree — native apps on macOS, Windows, and Linux
  • Tier 3: Screenshot + vision LLM — fallback when structure isn't enough
  • Record proof — each step captured with timestamp and tier used
DOM extractionAccessibility treeVision fallback~50× fewer tokens

Goose

Run a full AI agent locally via Ollama. No cloud API needed.

Local inference with open-weight models — no API keys, no usage fees, no data leaving your machine.

  • Ollama for local inference — download models, run on your GPU
  • 9+ providers — Anthropic, OpenAI, Google, Groq, Mistral, and more
  • Full-auto mode — autonomous execution without approval prompts
  • Per-slot configuration — different provider per specialist agent
Ollama (local)9+ providersFull-auto modePer-slot config
Build your own

Same SDK, same extension points.

Third-party developers use the same typed SDK as built-in extensions. Install @swarmix/plugin-sdk and @swarmix/plugin-ui, implement the PluginAdapter interface, and ship a plugin that integrates as deeply as the built-in ones.

Five plugin types: CLI agents, MCP servers, daemons, exporters, and utilities. Each runs in full isolation — its own process, its own config namespace, its own lifecycle.

CLI agent vendorAdd a new LLM provider to the agent spawn dropdown
MCP serverExpose tools callable by any agent in the workspace
Terminal toolsDeclare shell commands agents can invoke with streaming
Workspace pageFull-tab workspace component with sidebar entry
Tab pageDynamic plugin-hosted tabs in the layout
Agent feedbackIcon and panel in the agent chat header
Config & setupSettings forms, setup wizards, status pages
File handlerRoute file drops to the right processor

Local tools. No lock-in. No data leaks.

Every extension ships with the Studio. Build your own with the same SDK.