Agent

1MCP

1MCP is the unified MCP runtime. 1mcp serve aggregates your MCP servers, and CLI mode adds a thinner agent-facing workflow for Codex, Claude, Cursor, and similar tool-using agents.

Why 1MCP

Most MCP setups eventually hit two kinds of sprawl:

• Configuration sprawl: every client needs its own MCP wiring, auth choices, and filtering rules.
• Agent sprawl: autonomous sessions carry too many tools and schemas into context up front.

1MCP addresses both:

• 1mcp serve gives you one aggregated runtime in front of many MCP servers.
• CLI mode lets agents discover tools progressively with instructions, inspect, and run.
• Static servers can load at startup, while template servers are created from per-client or per-session context.
• Presets, filters, and instruction aggregation keep the same runtime adaptable across clients and projects.

Quick Start for Agent Users

This page is optimized for AI agent users. The 5-minute outcome is simple: start a real 1mcp serve runtime, connect your agent with cli-setup, then verify the instructions -> inspect -> run workflow.

You — setup: install 1MCP, add one upstream server, and start the runtime:

npm install -g @1mcp/agent
1mcp mcp add context7 -- npx -y @upstash/context7-mcp
1mcp serve

You — client setup: in a second shell, connect your agent to CLI mode:

1mcp cli-setup --codex
# or
1mcp cli-setup --claude --scope repo --repo-root .

AI agent — after startup: follow this workflow, skipping instructions if hooks already delivered the current instructions. Keep the runtime above running; you do not need to repeat these agent commands manually. Verify that startup delivers guidance and establishes session ownership. With a named or implicitly discovered local runtime, supported Codex sessions resolve their owner and assigned project automatically. Claude and hosts without automatic owner lookup add their own bootstrap --project-owner <own-file> to the commands below; each worker uses its own credential.

# Only if current target guidance is missing or needs refreshing:
1mcp instructions
1mcp inspect context7
1mcp inspect context7/query-docs
1mcp run context7/query-docs --args '{"libraryId":"/mongodb/docs","query":"aggregation pipeline"}'

If you want the full walkthrough (with success criteria and off-ramps), use the Quick Start guide.

For a given agent, choose one mode only. If you switch that agent to CLI mode, remove its old direct MCP configuration first.

Work Across Projects and Worktrees

Use one local runtime across repositories and Git worktrees. After configuring an installed backend with mcp setup, the agent selects a checkout on each call:

1mcp inspect codegraph --project /absolute/checkout

For the session’s assigned checkout, omit --project; explicit overrides do not change that default. Follow Working Across Projects and Worktrees (中文) for setup, project sets, independent workers, optional preparation, and safe cleanup.

Why CLI Mode Exists

CLI mode is the primary workflow for agent-style sessions. It keeps MCP as the backend protocol but narrows what the agent sees at each step:

• instructions explains the current runtime and recommended flow
• inspect lets the agent discover only the server or tool it needs
• run executes one selected tool after schema inspection

That gives agent loops a smaller working surface without giving up the unified runtime behind 1mcp serve.

Choose Another Path

Stdio Proxy

Use 1mcp proxy when you want the broadest client compatibility without giving up project context.

It is the recommended fallback after CLI mode because it:

• works with the stdio transport that most AI clients already support
• keeps project context through .1mcprc
• supports template MCP servers resolved from project or session context
• is easier to roll out with one-time global setup plus per-project config

Direct stdio mode is not the recommended path. It is mainly useful for debugging because 1MCP startup is slower than a thin standalone stdio setup.

Direct MCP Attachment

Direct MCP attachment is still supported for clients that want to talk to the aggregated runtime over streamable HTTP.

Examples:

{
  "mcpServers": {
    "1mcp": {
      "url": "http://127.0.0.1:3050/mcp?app=cursor"
    }
  }
}

claude mcp add -t http 1mcp "http://127.0.0.1:3050/mcp?app=claude-code"

Use this path if your client already speaks MCP natively, can work without project context, and you do not want CLI mode. For Codex, Claude, Cursor, and similar agent loops, prefer CLI mode first and proxy second.

Runtime Operators

Use the deeper docs if you are configuring or deploying the runtime itself:

• Configuration
• Authentication
• Architecture

Contributors

• Development guide
• CONTRIBUTING.md

How It Works

flowchart LR
    A[User or Agent] --> B[1mcp serve]
    B --> C[Static servers loaded at startup]
    B --> D[Template servers resolved from client or session context]
    A --> E[CLI mode: instructions -> inspect -> run]
    E --> B
    F[Direct streamable HTTP client] --> B
    G[stdio-compatible client] --> H[1mcp proxy]
    H --> B

1MCP runs as an aggregated runtime behind 1mcp serve. Static servers are prepared from startup configuration, template servers are materialized when client context is known, and the runtime can use async loading for early HTTP listener availability and lazy loading for a stable tool surface. Instruction aggregation, presets, and notifications sit alongside that runtime rather than outside it.

Lazy loading is an opt-in stable tool-surface compatibility mode. It keeps the backend discovery and invocation surface at tool_list, tool_schema, and tool_invoke so capable agents can discover tools progressively without replacing their MCP tool table. Any explicitly enabled internal management tools remain directly exposed. Lazy loading reduces the initial schema payload, but it does not reduce backend connections or processes, make synchronous startup bind earlier, or repair orphaned proxy processes. See #392 for the async late-server visibility contract.

Core Capabilities

• Unified runtime for many MCP servers behind one serve process
• CLI mode for progressive discovery with 1mcp instructions, 1mcp inspect <server>, 1mcp inspect <server>/<tool>, and 1mcp run <server>/<tool> --args '<json>'
• Template servers for per-client or per-session resolution
• Opt-in async loading for early HTTP listener availability when clients can reconcile capability changes
• Opt-in lazy loading for a stable progressive-discovery tool surface and smaller initial schemas
• Opt-in automatic recovery for owned stdio backends, with health/status visibility and operator restart controls
• Instruction aggregation across static and template-backed servers
• Presets, filters, and preset change notifications
• proxy for maximum compatibility with project context and template-server support
• Direct streamable HTTP MCP access for native HTTP clients that do not need project context

Common Use Cases

• Give a coding agent one stable runtime but a smaller working surface.
• Share the same MCP inventory across Cursor, Claude Code, Codex, and internal tooling.
• Expose context-specific template servers per repo, branch, or session.
• Centralize auth, filtering, presets, and runtime lifecycle instead of rebuilding them in ad hoc scripts.

Contributing / License

Contributions are welcome. See CONTRIBUTING.md for the development workflow and LICENSE for the Apache 2.0 license.