Best for
- Developers wanting to leverage a powerful coding agent within MCP clients
- Builders creating AI applications requiring complex agent orchestration
- Hobbyists and teams seeking a customizable, config-driven agent harness
D
Dexto
Community
An open agent harness for AI applications—ships with a powerful coding agent.
GitHub source repository ↗
52FMRS · D
Dexto is an agent harness—the orchestration layer that turns LLMs into reliable, stateful agents that can take actions, remember context, and recover from errors. It provides configuration-driven agent definitions (YAML), session management, tool orchestration, memory, multimodal support, and observability. Dexto can run locally, cloud, or hybrid, with CLI, Web UI, REST API, and Discord/Telegram integrations. It can also act as an MCP server over stdio, exposing your agent to clients like Claude Code or Cursor.
Tools
agent-spawner
Spawn specialized sub-agents to handle complex subtasks.
mcp-tools
Access connected tools, prompts, and resources via MCP protocol.
Setup
Install: curl -fsSL https://dexto.ai/install | bash (macOS/Linux/WSL) or irm https://dexto.ai/install.ps1 | iex (Windows PowerShell). For MCP mode, ensure you pass --mode mcp and specify the agent with --agent coding-agent, and add --auto-approve and --no-elicitation for non-interactive operation. Example MCP config: {"command":"npx","args":["-y","dexto","--mode","mcp","--agent","coding-agent","--auto-approve","--no-elicitation"]}.
claude_desktop_config.json
{
"command": "npx",
"args": [
"-y",
"dexto",
"--mode",
"mcp",
"--agent",
"coding-agent",
"--auto-approve",
"--no-elicitation"
]
}
Fit and risk
Not for
- Users needing official support or enterprise guarantees
- Deployments requiring non-commercial licensing (Elastic License 2.0 restricts use)
- Non-developers seeking a no-code quick setup
Required permissions
- Access to local file system (depending on configuration)
- May execute shell commands and run tests
- Requires LLM API keys or local model setup
Risks and side effects
- Dexto is in beta and features may change
- Agents can execute arbitrary code, posing security risks
- Telemetry collects anonymous usage data (can be disabled)
Troubleshooting
- Ensure Node.js and pnpm are installed
- Check logs at `~/.dexto/logs/`
- For MCP mode, `--auto-approve` and `--no-elicitation` are required
- If connection fails, verify MCP server command and args
- Use `DEXTO_LOG_LEVEL=debug` for detailed output
Use cases
Run the coding agent inside Claude Code or Cursor
Build autonomous agents
Create multi-agent systems
Deploy MCP client capabilities
Supported clients
Claude DesktopFull support
CursorFull support