Delv
CLIby Anthropic★ 4.4

Claude Code

Anthropic's terminal-first coding agent. Full MCP support, uses the same config format as Desktop. The fastest way to add MCPs to a coding workflow.

Homepage MCP docs
A
Safety & Trust

Delv Safety Grade: A

Score 82/100 · assessed 2026-04-19

Maintainer95
Permissions45
Supply chain85
Transparency75
Incidents100

Claude Code is Anthropic's official terminal-based coding agent with full MCP support. As a first-party Anthropic product, it benefits from strong organisational backing and professional maintenance. The tool inherits the same MCP configuration format as Claude Desktop, which simplifies setup but also means it can access any MCP server you configure, including those with broad permissions. The primary safety concern is that this is an autonomous coding agent with filesystem write access, shell execution capability, and network access by design. It operates in your terminal with the same privileges you have, so misconfigured or malicious MCP servers could cause significant damage. No repository means limited transparency into implementation details, though documentation exists. The supply chain is controlled by Anthropic through direct distribution. No known security incidents. Best suited for developers who understand the implications of granting an AI agent terminal access.

Green flags

  • Official Anthropic product with professional maintenance
  • Uses same MCP config as Desktop, reducing configuration complexity
  • Full MCP protocol support (tools, resources, prompts)
  • Clear documentation at docs.claude.com
  • No known security incidents or credential leaks

Red flags

  • No public repository limits code auditability
  • Autonomous agent with full terminal privileges in working directory
  • Inherits all permissions from configured MCP servers without sandboxing
  • Can execute arbitrary shell commands as part of coding workflow
  • Filesystem write access across entire project scope

Permissions requested

Read filesWrite filesDelete filesShell executeOutbound networkRead envExternal LLM call
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

MCP capabilities

  • Tools
  • Resources
  • Prompts
  • Sampling

Platforms

macOSLinuxWindows

Config location

~/.claude/mcp.json · or via `claude mcp add`

Review

Claude Code is Anthropic's terminal-first coding agent, and it's the most direct route from MCP server to actual work. If you've configured an MCP server in Claude Desktop, the same JSON file works here. No translation layer, no adapter hell. You drop a server definition into `~/.claude/mcp.json` or run `claude mcp add`, and it's live. Full support for tools, resources, and prompts means you get the entire MCP surface area without compromise. I reach for this when I need to iterate on code without leaving the terminal. The workflow is simple: point it at a directory, ask it to refactor something, and watch it call out to your MCP servers for context. If you've wired up a database MCP, it queries schema. If you've got a filesystem server, it reads your actual project structure. The speed advantage over switching between terminal and Desktop is real when you're in flow. The CLI nature means it's scriptable. You can pipe prompts, chain commands, or slot it into CI experiments. That's not theoretical: I've seen people use it to auto-generate migration scripts by querying a database MCP, then validate them against a linting tool MCP. Desktop can't do that. Rough edges exist. Error messages when an MCP server fails to start are terse. You don't get the visual config UI that Desktop offers, so debugging a malformed server path means reading JSON and checking logs manually. The terminal UI is clean but minimal: no rich markdown rendering, no inline images. If your MCP server returns a 50-line JSON blob, you're scrolling raw text. Sampling isn't supported, which rules out certain advanced MCP patterns where the server needs to prompt the model mid-operation. That's a niche limitation, but it's there. For 95% of MCP use cases (tools, resources, prompts), it's complete. The real win is speed. No Electron overhead, no UI lag. You type, it responds, it calls your servers, it writes code. If your workflow is terminal-native and you've already invested in MCP servers, this is the fastest way to use them. If you need hand-holding or visual config, stick with Desktop.
Verdict

Best MCP client for terminal-native developers who want zero friction between MCP servers and coding work. If you live in the CLI and already have MCP servers configured, this is the obvious choice. Visual thinkers or those needing rich output should stay with Desktop.

Good at

  • Shares exact config format with Claude Desktop, no translation needed
  • Full MCP support: tools, resources, prompts all work
  • Scriptable and pipeable, fits into automation workflows
  • Fast, no Electron overhead or UI lag
  • Simple `claude mcp add` command for quick server setup

Watch out

  • No sampling support, blocks some advanced MCP patterns
  • Minimal error messages when MCP servers fail
  • No visual config UI, debugging requires reading JSON manually
  • Terminal-only output, no rich markdown or inline images
  • Less forgiving for users unfamiliar with CLI workflows