A Claude Skill for the LangChain Deep Agents CLI
Scaffold a production-grade Claude Skill that gives Claude comprehensive, structured knowledge of the Deep Agents CLI — commands, providers, memory, sandboxes, streaming, and SDK customization. Works with Claude Code, Claude Desktop, and claude.ai.
Installation · What Gets Installed · Skill Resources · Official Documentation
The Deep Agents CLI is LangChain's terminal coding agent — a LangGraph-powered assistant with persistent memory, 20+ LLM provider integrations, sandboxed code execution, and a composable skills system. Its surface area spans interactive and non-interactive modes, a rich slash-command REPL, programmatic SDK customization, streaming output, and the Agent Client Protocol for editor integration.
This package scaffolds a Claude Skill that encodes that entire surface area as structured, on-demand knowledge. When the skill is active, Claude can accurately assist with any Deep Agents CLI task — from launching a first session to building production subagent pipelines — without requiring repeated context or pasted documentation. The skill works across all Claude surfaces: Claude Code, Claude Desktop, and claude.ai.
The skill follows the Agent Skills open standard and loads automatically when relevant. It can also be invoked directly with /deepagents-cli.
This isn't just a documentation reference. When your coding agent has this skill loaded, it can actively help you work with deepagents -- not just answer questions about it.
- Learn by building. Ask your agent to walk you through deepagents concepts, then have it set up a session, configure providers, or scaffold a skill -- all from within your current editor.
- Author custom agents. The skill includes the full SDK reference for
create_deep_agent(), middleware, subagents, and backends. Your agent can help you design and implement specialized deepagents workflows tailored to your stack. - Extend deepagents itself. Clone the deepagents repo, install this skill, and let your coding agent help you customize the framework -- adding tools, modifying behavior, or building production pipelines.
The skill covers everything from first launch to programmatic customization, so your agent grows with you.
Installs the skill into the current project's .claude/skills/ directory. Check it into version control to share it with your team.
npx deepagents-cli-claude-skill initInstalls the skill to ~/.claude/skills/ so it is available across all projects on your machine.
npx deepagents-cli-claude-skill init --personalIf you use the Claude Desktop app or claude.ai (no terminal required):
- Download
deepagents-cli-skill.zipfrom the latest release - Open Claude Desktop or claude.ai
- Go to Settings > Features > Custom Skills
- Upload the zip file
The skill works identically across all Claude surfaces. Requires a Pro, Max, Team, or Enterprise plan with code execution enabled.
| Flag | Description | Default |
|---|---|---|
--personal |
Install to ~/.claude/skills/ (personal scope) |
false |
--path <dir> |
Target directory for workspace installation | Current directory |
--force |
Skip confirmation when skill directory already exists | false |
--help |
Show usage information | — |
--version |
Show installed version | — |
To update the skill after a new release, re-run the init command. Existing files are never overwritten; only files added in the new release will be created.
.claude/skills/deepagents-cli/
├── SKILL.md Main skill definition and quick-start reference
└── references/
├── cli-reference.md Complete CLI commands, flags, and keyboard shortcuts
├── providers.md LLM provider configuration and config.toml schema
├── skills-system.md Skills creation, SKILL.md format, progressive disclosure
├── memory-and-persistence.md Memory system, AGENTS.md, SDK long-term memory
├── sandboxes.md Sandbox providers: Modal, Runloop, Daytona, LangSmith
├── sdk-customization.md create_deep_agent(), middleware, subagents, backends, HITL
├── streaming.md Streaming modes, subgraph streaming, frontend integration
├── acp.md Agent Client Protocol and editor integrations
└── workflows.md Automation patterns, CI/CD, troubleshooting
The skill follows the Agent Skills progressive disclosure model: SKILL.md is the primary context loaded by Claude, and the reference files are surfaced on-demand as the task requires them. This keeps the token footprint minimal while making the full surface area available.
Each reference file is a structured knowledge document scoped to a specific area of the Deep Agents CLI. All content is sourced from the official LangChain Deep Agents documentation and aligned with the documented API.
Purpose: Entry point for the skill. Contains a concise overview, quick-start examples, core concept summaries, and links to all reference files.
Value: Claude reads this file first when the skill is activated. It provides enough context to handle the most common tasks immediately and directs Claude to the appropriate reference file for deeper questions, keeping initial token usage minimal.
File: references/cli-reference.md
Source: CLI Overview
Complete reference for every command, flag, slash command, and keyboard shortcut in the Deep Agents CLI. Covers top-level commands (deepagents list, deepagents reset, deepagents skills), all launch flags (-M, -a, -n, -r, -q, --auto-approve, --shell-allow-list), all slash commands (/model, /remember, /threads, /trace, /clear, /tokens), keyboard shortcuts for the interactive REPL, the built-in tool inventory with approval requirements, and non-interactive mode behavior in detail.
File: references/providers.md
Source: Model Providers
Configuration reference for all 20+ supported LLM providers including OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Ollama, Groq, DeepSeek, Mistral AI, Perplexity, xAI, OpenRouter, and more. Covers API key environment variables per provider, the ~/.deepagents/config.toml schema in full, model resolution order (-M flag → config default → recent → API key detection), and provider-specific parameters for local models such as Ollama.
File: references/skills-system.md
Source: Skills
Detailed guide to the Deep Agents skills system: the SKILL.md format and frontmatter fields, directory conventions for user-level (~/.deepagents/<agent>/skills/) and project-level (.deepagents/skills/) skills, the progressive disclosure loading model, override behavior when user and project skills share a name, and the full CLI for creating, listing, inspecting, and deleting skills.
File: references/memory-and-persistence.md
Source: Long-term Memory
Covers both layers of the Deep Agents memory system: the file-based layer (AGENTS.md and memories/) that persists across sessions per agent, and the SDK layer (InMemoryStore, AsyncPostgresStore) for programmatic long-term memory backends. Includes the agent directory layout, the /remember command and its optional context argument, deepagents reset behavior, and the distinction between agent-scoped and project-scoped AGENTS.md files.
File: references/sandboxes.md
Source: Sandboxes
Setup and configuration guide for all four supported sandbox providers: Modal (GPU and ML workloads), Runloop (disposable devboxes), Daytona (fast cold starts for TypeScript and Python development), and LangSmith (cloud-hosted execution). Covers --sandbox, --sandbox-setup, and --sandbox-id flags, provider-specific environment variables, setup script patterns, sandbox reuse across sessions, and the security isolation properties of each provider.
File: references/sdk-customization.md
Sources: Customization, Subagents, Backends, Human-in-the-Loop, Middleware
Programmatic usage of the Deep Agents SDK via create_deep_agent(). Covers the full function signature, model configuration, custom tool registration, system prompt customization, the middleware stack (request and response hooks for logging, rate limiting, and guardrails), subagent patterns (spawning agents as tools within a parent agent), backend types (local, sandbox, ACP), Human-in-the-Loop approval gates with interrupt and resume patterns, context management, and structured output configuration.
File: references/streaming.md
Sources: Streaming Overview, Streaming Frontend
Streaming output reference for both CLI and SDK usage. Covers the three streaming modes (text, events, full), subgraph streaming for multi-agent pipelines where each agent emits its own stream, Server-Sent Events integration for web frontends, WebSocket alternatives, and the frontend streaming protocol including event types and client-side consumption patterns.
File: references/acp.md
Source: Agent Client Protocol
Reference for the Agent Client Protocol (ACP) — the open standard that allows editors and external clients to communicate with a running Deep Agents instance over a local socket. Covers the ACP server (deepagents serve), the request and response protocol, capability negotiation, and integration guides for Zed, VS Code, JetBrains IDEs, and Neovim.
File: references/workflows.md
Practical reference for automation and scripted usage. Covers non-interactive mode recipes for common tasks, CI/CD integration patterns with environment configuration, piped input workflows using stdin, multi-step subagent orchestration, shell access configuration with --shell-allow-list, output formatting with -q and --no-stream for downstream piping, and a troubleshooting index for the most common runtime errors.
This skill is built to the Agent Skills open standard and the Claude Code Skill specification:
| Requirement | Implementation |
|---|---|
SKILL.md with YAML frontmatter |
name, description, and compatibility fields present; description written for accurate auto-invocation matching |
| Narrow, outcome-focused scope | Skill scoped entirely to Deep Agents CLI knowledge; single domain, single responsibility |
| Co-located supporting files | references/ directory co-located with SKILL.md inside the skill directory |
| Progressive disclosure | SKILL.md is the primary context under 500 lines; reference files are loaded on-demand by topic |
| Flexible invocation | Claude loads the skill automatically when relevant; user can also invoke with /deepagents-cli |
| Personal and workspace scopes | --personal targets ~/.claude/skills/; default targets <project>/.claude/skills/ |
| Never overwrites existing files | The installer skips any file that already exists at the target path |
All technical content in this skill is sourced from and cross-referenced against the official LangChain Deep Agents documentation:
Claude Code Skill specification: https://code.claude.com/docs/en/skills Agent Skills open standard: https://agentskills.io/specification
- Claude Code install: Node.js >= 18
- Claude Desktop / claude.ai install: No requirements — just download the zip and upload
- A Claude subscription (Pro, Max, Team, or Enterprise)
MIT — see LICENSE for details.