Open Multi-Agent Canvas
An open-source multi-agent chat interface that manages multiple agents in one dynamic conversation and adds MCP servers for deep research and general tasks.
Per-dimension scores and reasoning
Static review based on README only. MIT license and CopilotKit attribution are present. However, the MCP Agent lets users run arbitrary local commands via stdio and connect to arbitrary external SSE servers, with no documented least-privilege, user confirmation, data-flow explanation, key protection, or rollback; deductions apply because no trust control is evidenced in files.
README gives basic install steps but no error handling, failure messages, or dependency-locking evidence; dependency availability is inferred from standard pnpm/poetry flow only.
Target scenarios (travel, research, general MCP tasks) and environment requirements (pnpm, poetry, port 8123, tunnel) are clear; but agent capability boundaries and trigger precision are entirely undocumented.
Information architecture is clear (frontend/agent split), install notes concrete, MIT license stated, and maintenance path explicit (consolidated into CopilotKit monorepo). However, naming stability is weak (MCP mis-expanded as 'Multi-Channel Protocol' instead of Model Context Protocol), no changelog, no FAQ, no documented known limitations, and this repo itself is effectively unmaintained as a snapshot.
As a multi-agent canvas with MCP configuration it offers differentiation; output usability rests on screenshots and prose, unverifiable statically, hence mid-range scores.
Claims (built-in MCP Agent, two external agent repos, CopilotCloud requirement) have no reproduction path or tests and link only back to CopilotKit-owned resources, lacking third-party corroboration; fact/inference separation is acceptable but the evidence base is narrow.
- This repo is consolidated into the CopilotKit monorepo; this is a historical snapshot with no future security fixes here.
- MCP stdio mode allows arbitrary local command execution; review MCP server configuration and sandboxing before deployment.
- Connecting to third-party public MCP servers (composio, mcp.run) routes data through external services; assess privacy risk yourself.
- Storage and transmission protection for API keys (OPENAI/LANGSMITH/CopilotCloud) is undocumented.
- README mis-expands MCP as 'Multi-Channel Protocol'; documentation accuracy is questionable.
What does this agent do, and when should you use it?
Open Multi-Agent Canvas, created by the CopilotKit team, is an open-source chat interface for managing multiple agents within a single dynamic conversation. It is built with Next.js, LangGraph, and CopilotKit and targets travel planning, research, and general-purpose tasks through MCP servers. The repository contains a Next.js frontend plus an optional Python MCP Agent backend installed via Poetry and served by the LangGraph dev server on port 8123. The actual agents (Travel Agent and AI Researcher) live in separate repositories and can be run independently or deployed on LangSmith. A built-in MCP Agent connects to MCP servers via Standard IO (local commands such as Python scripts) or SSE (external MCP-compatible servers). Note that the project has been consolidated into the CopilotKit monorepo, with the latest version under examples/showcases/multi-agent-canvas.
The frontend is a Next.js app that uses CopilotKit to connect multiple LangGraph agents — the CoAgents Travel Agent and the CoAgents AI Researcher — and coordinates their output in one chat surface. The built-in MCP Agent is a general-purpose agent configured through the "MCP Servers" button in the top right of the interface: Standard IO mode runs local commands (e.g., Python scripts), while SSE mode connects to external MCP-compatible servers; public servers such as mcp.composio.dev and mcp.run are also supported. The MCP Agent backend lives in the agent folder, is managed with Poetry, and is launched with the langgraph dev command, then wired to the frontend via a Remote Endpoint / Local Development tunnel pointing at port 8123.
- Product developers who want a reference architecture for building a multi-agent chat UI with CopilotKit and LangGraph.
- Researchers who need a research agent to continuously gather and synthesize material, using the CoAgents AI Researcher.
- Teams that want agents to invoke local scripts or external tools via MCP, including Composio and mcp.run servers.
- Travelers who want an agent to help plan itineraries within a single conversation, using the CoAgents Travel Agent.
- Engineering teams that need one unified frontend to manage multiple LangGraph agents deployed on LangSmith.
What are this agent's strengths and limitations?
- Manages multiple agents inside one dynamic conversation instead of separate interfaces per agent.
- Ships a built-in MCP Agent with both Standard IO and SSE server modes, plus support for public servers like Composio and mcp.run.
- Clear, reusable stack: Next.js frontend, LangGraph agents, and CopilotKit, with agents deployable independently to LangSmith.
- Official Travel Agent and AI Researcher agents provide ready-made source code to fork and customize.
- Strong CopilotKit ecosystem lock-in: a Copilot Cloud API key is mandatory to run the project.
- The backend requires OPENAI_API_KEY, creating a runtime dependency on OpenAI.
- The repo has been consolidated into the CopilotKit monorepo and no longer accepts issues/PRs here, a maintenance-path risk.
- License metadata is unknown; the README states MIT but you should verify the LICENSE file yourself.
- Running the MCP Agent backend additionally requires Poetry and a LangGraph dev server, making local deployment non-trivial.
How do you install or deploy this agent?
- Prerequisite: install pnpm and obtain a Copilot Cloud API key from dashboard.operations.copilotkit.ai. 2. Frontend: rename frontend/example.env to .env with NEXT_PUBLIC_CPK_PUBLIC_API_KEY=..., then run: cd frontend && pnpm i && pnpm run build && pnpm run start. 3. (Optional) MCP Agent backend: rename agent/example.env to .env with OPENAI_API_KEY=... and LANGSMITH_API_KEY=..., then run: cd agent && poetry install && poetry run langgraph dev --host localhost --port 8123 --no-browser. 4. For a tunnel, open another terminal, select Remote Endpoint, then Local Development, copy the command and change the port to the LangGraph server's 8123.
How do you use this agent?
After starting the frontend, click the "MCP Servers" button in the top right to configure custom servers: choose Standard IO to run local commands (e.g., Python scripts) or SSE to connect to external MCP-compatible servers; you can also connect public servers like mcp.composio.dev or mcp.run. If you run the local MCP Agent backend, point a Remote Endpoint / Local Development tunnel at the LangGraph server on port 8123. You can then invoke the travel agent, researcher agent, and general MCP agent within one conversation. Copilot Cloud is required to run the project.