Parity

Agent control plane for MCP + coding runtimes + multi-agent teams (TypeScript). UI inspired by OpenHands Agent Canvas.

cp .env.example .env
npm run install-all
npm run dev

• UI: http://localhost:3000
• API: http://localhost:5005/api/health

Separate processes: npm run server · npm run client

Working example:

Workspace: Files → Select folder (native dialog on the API host → real path). Or paste an absolute path → Set.

Quick tour

1. Chat — LLM + workspace tools + connected MCP
2. Servers — stdio / Streamable HTTP MCP
3. Tools / Playground — inspect and invoke tools
4. Workflows — multi-step runs (tool, agent, parallel, synthesize, handoff, team) + HITL
5. Teams — named agent roster + hierarchical director → workers → synthesis
6. Observability — timeline, tokens, evals

Multi-agent teams

Parity ships default agents (director, researcher, coder, reviewer, synthesizer).

Surface	How
Chat	run_team tool (or delegate_task with an agent name)
UI	Teams nav — run task, edit roster, view recent runs
API	POST /api/teams/run, GET /api/agents, workflow step type: "team"
Workflows	Steps: agent, parallel, synthesize, handoff, team

Shared team state (plan, messages, artifacts) is stored per run and appears in observability events.

How it works

POST /api/chat → runAgentTurn → one streamText session (ReAct tool steps inside that call).

Message	Typical path
"hi"	Chat LLM only
Edit / run code	file_editor / terminal / grep / …
“Where is X?”	codebase_search after Reindex (Voyage embeddings)
Multi-agent task	run_team → director plan → parallel workers → synthesis
Browser MCP	Connected Playwright (etc.) tools

RAG is opt-in (Reindex in Files). Chat model ≠ embedding model.

Docker

docker compose up --build