← Back to directory
K

Kode CLI

Community
Multi-model terminal AI coding agent that acts as an MCP client to connect external tool servers
GitHub source repository ↗
★ 5.2k Stars Category · Dev Tools Very popular
44FMRS · D
Reliability
5/20
Security and permissions
3/20
Maintenance
11/20
Documentation
13/20
Setup experience
12/20

Kode is a feature-rich, multi-model terminal AI coding CLI whose 'MCP Extensions' functionality is client-side — it connects to external MCP servers to extend its own capabilities rather than exposing itself as an MCP server callable by other clients. The provided source material contains no manifest or tool list for Kode acting as an MCP server, so the tools and transport fields are left empty. The default YOLO permission mode and reliance on third-party plugins/model integrations are the main risk areas to evaluate before adoption.

Read the FMRS scoring method →

Kode (@shareai-lab/kode) is a terminal-based AI coding agent CLI billed as a "unit agent for every human & computer task." It is not itself an MCP server exposing tools — it functions as an MCP client: users add and manage connections to external MCP servers via kode mcp add, kode mcp list, kode mcp remove, or by editing a project-root .mcp.json file (e.g. {"my-server": {"type": "sse", "url": "..."}}), which extends what Kode itself can do. Beyond MCP, Kode orchestrates 20+ AI models across distinct role pointers (main conversation, subagent/task, context compaction, quick utilities), supports reusable subagent templates invoked via @run-agent-*, loads skills/plugins from .kode/skills, and is compatible with the AGENTS.md / legacy CLAUDE.md project-instruction standards. It separately ships a kode-acp command that exposes an Agent Client Protocol (ACP) stdio JSON-RPC interface for editor clients like Toad and Zed — a distinct protocol from MCP. By default it runs in "YOLO mode," skipping most permission prompts; kode --safe enables approval checks and, on Linux, a bwrap sandbox. The README notes the project is undergoing a major maintenance and refactoring phase from July through December 2026.

Tools

The tool list has not been reviewed yet.

Setup

  1. Install globally: npm install -g @shareai-lab/kode (a China mirror registry and a @dev channel are also documented), or download a Bun-compiled standalone binary from GitHub Releases.
  2. Run kode for interactive mode, or kode -p "..." file.js for one-shot tasks.
  3. To let Kode connect to an external MCP server and gain its tools, run kode mcp add interactively, or edit the project's .mcp.json to add an entry such as {"my-server": {"type": "sse", "url": "http://127.0.0.1:3333/sse"}}; verify with kode mcp list.
  4. Configure model pointers (main/task/compact/quick) and API keys via /model or the kode models command family; profiles can be exported/imported as YAML (without plaintext keys) for team sharing.

Fit and risk

Best for

  • Developers who want multi-model orchestration inside a terminal coding agent
  • Teams already using the AGENTS.md standard or migrating from similar terminal AI coding tools
  • Users who want to extend a coding agent's capabilities by connecting it, as an MCP client, to custom external tool servers

Not for

  • Scenarios that need a service exposed AS an MCP server for Claude Desktop or other MCP clients to call — Kode does not provide that here
  • Security-sensitive environments requiring default-strict permission checks, since YOLO mode is on by default and --safe must be enabled manually
  • Simple, single-fixed-model use cases that don't need multi-model orchestration

Required permissions

  • Read and write files in the working directory and subdirectories (editing/creating code)
  • Execute shell/Bash commands (no per-command confirmation by default in YOLO mode)
  • Make network requests: calls to configured model APIs, optional web tools, plugin/skill marketplace downloads, optional update checks
  • (Optional) run Bash commands inside a Linux bwrap sandbox with networking disabled by default, when --safe or KODE_SYSTEM_SANDBOX=1 is set

Risks and side effects

  • Default YOLO mode skips permission prompts, so the model can modify files or run destructive commands without per-action approval
  • Externally connected MCP servers and marketplace-installed plugins/skills come from third parties and can introduce untrusted code execution paths
  • Multi-model orchestration means code and context are sent to whatever third-party model APIs the user configures, raising data-residency/privacy considerations
  • The README states the project is in a major refactoring/maintenance phase (July–December 2026), so interfaces, commands, or config formats may change

Troubleshooting

  1. If installing with --no-optional breaks the bundled ripgrep binary, install system `rg` or set the KODE_RIPGREP_PATH environment variable
  2. If a connection to an external MCP server fails, check the address/type (e.g. sse) in the project's .mcp.json or .mcprc and confirm status with kode mcp list
  3. For model call issues, run kode models list to verify the main/task/compact/quick pointers resolve to correctly configured model profiles
  4. On Linux, if commands are unexpectedly restricted (e.g. no network), check whether --safe or KODE_SYSTEM_SANDBOX=1 is active and adjust as needed
  5. If AGENTS.md/CLAUDE.md instructions aren't being picked up, confirm the files sit on the path from the repo root to the current working directory and stay within the 32 KiB cap (adjustable via KODE_PROJECT_DOC_MAX_BYTES)

Use cases

Understand a codebase, edit files, run commands, and orchestrate multi-step development workflows from the terminal
Assign different AI models to different task types (architecture, implementation, expert review) within one session
Define reusable subagent templates under .kode/agents for delegated tasks like code review or refactors
Connect external MCP servers (filesystem, database, browser, etc.) to extend Kode's own available toolset
Integrate Kode into ACP-compatible editors (Zed, Toad) via the kode-acp stdio JSON-RPC interface

Supported clients

Supported clients have not been confirmed yet.