Skip to main content
Deep Agents Deploy takes your agent configuration and deploys it as a LangSmith Deployment: a horizontally scalable server with 30+ endpoints including MCP, A2A, Agent Protocol, human-in-the-loop, and memory APIs. Built on open standards:
  • Open source harness: MIT licensed, available for Python and TypeScript
  • AGENTS.md: open standard for agent instructions
  • Agent Skills: open standard for agent knowledge and actions
  • Any model, any sandbox: no provider lock-in
  • Open protocols: MCP, A2A, Agent Protocol
  • Self-hostable: LangSmith Deployments can be self-hosted so memory stays in your infrastructure
Deep Agents Deploy is currently in beta. APIs, configuration format, and behavior may change between releases. See the releases page for detailed changelogs.

​
Compare to Claude Managed Agents

Deep Agents DeployClaude Managed Agents
Model supportOpenAI, Anthropic, Google, Bedrock, Azure, Fireworks, Baseten, OpenRouter, many moreAnthropic only
HarnessOpen source (MIT)Proprietary, closed source
SandboxLangSmith, Daytona, Modal, Runloop, or customBuilt in
MCP support✅✅
Skill support✅✅
AGENTS.md support✅❌
Agent endpointsMCP, A2A, Agent ProtocolProprietary
Self hosting✅❌

​
What you’re deploying

deepagents deploy packages your agent configuration and deploys it as a LangSmith Deployment. You configure your agent with a few parameters:
ParameterDescription
modelThe LLM to use. Any provider works — see supported models.
AGENTS.mdThe system prompt, loaded at the start of each session.
skillsAgent Skills for specialized knowledge and actions. Skills are synced into the sandbox so the agent can execute them at runtime. See skills docs.
user/Per-user writable memory. If a AGENTS.md template is present in the user folder, the agents seeds the template per user (if the folder is empty the agents creates an empty AGENTS.md). Writable at runtime. Preloaded into the agent’s context via the memory middleware.
mcp.jsonMCP tools (HTTP/SSE). See MCP docs.
sandboxOptional execution environment. Thread-scoped sandboxes are provisioned per thread and will be re-created if the server restarts. Use scope = "assistant" if you need sandbox state that persists across threads. See sandbox providers.

​
Install

Install the CLI or run directly with uvx: 
uv tool install deepagents-cli

​
Usage

deepagents init [name] [--force]                                             # scaffold a new project
deepagents dev  [--config deepagents.toml] [--port 2024] [--allow-blocking]  # bundle and run locally
deepagents deploy [--config deepagents.toml] [--dry-run]                     # bundle and deploy
By default, deepagents deploy looks for deepagents.toml in the current directory. Pass --config to use a different path: 
deepagents deploy --config path/to/deepagents.toml
deepagents deploy fully rebuils and creates a new revision on every invocation. Use deepagents dev for local iteration.

​
deepagents init

Scaffold a new agent project: 
deepagents init my-agent
This creates the following files:
FilePurpose
deepagents.tomlAgent config — name, model, optional sandbox
AGENTS.mdSystem prompt loaded at session start
.envAPI key template (GOOGLE_API_KEY, LANGSMITH_API_KEY, etc.)
mcp.jsonMCP server configuration (empty by default)
skills/Directory for Agent Skills, with an example review skill
After init, edit AGENTS.md with your agent’s instructions and run deepagents deploy. Optionally add a user/ directory with per-user memory templates — see User Memory.

​
Project layout

The deploy command uses a convention-based project layout. Place the following files alongside your deepagents.toml and they are automatically discovered: 
my-agent/
├── deepagents.toml
├── AGENTS.md
├── .env
├── mcp.json
├── skills/
│   ├── code-review/
│   │   └── SKILL.md
│   └── data-analysis/
│       └── SKILL.md
└── user/
    └── AGENTS.md
File/directoryPurposeRequired
AGENTS.mdMemory for the agent. Provides persistent context (project conventions, instructions, preferences) that is always loaded at startup. Read-only at runtime.Yes
skills/Directory of skill definitions. Each subdirectory should contain a SKILL.md file. Read-only at runtime.No
user/Per-user writable memory. If a AGENTS.md template is present in the user folder, the agents seeds the template per user (if the folder is empty the agents creates an empty AGENTS.md). Writable at runtime. Preloaded into the agent’s context via the memory middleware.No
mcp.jsonMCP server configuration. Only http and sse transports are supported in deployed contexts.No
.envEnvironment variables (API keys, secrets). Placed alongside deepagents.toml at the project root.No
mcp.json must only contain servers using http or sse transports. Servers using stdio transport are not supported in deployed environments because there is no local process to spawn.Convert stdio servers to HTTP or SSE before deploying.

​
Configuration file

deepagents.toml configures the agent’s identity and sandbox environment. Only the [agent] section is required. The [sandbox] section is optional and defaults to no sandbox.

​
[agent]

(Required) Core agent identity. For more on model selection and provider configuration, see supported models.
​
name
string
required
Name for the deployed agent. Used as the assistant identifier in LangSmith.
​
model
string
default:"anthropic:claude-sonnet-4-6"
Model identifier in provider:model format. See supported models.
deepagents.toml
[agent]
name = "research-assistant"
model = "google_genai:gemini-3.1-pro-preview"
The name field is the only required value in the entire configuration file. Everything else has defaults.
Skills, user memories, MCP servers, and model dependencies are auto-detected from the project layout — you don’t declare them in deepagents.toml:
  • Skills: the bundler recursively scans skills/, skipping hidden dotfiles, and bundles the rest.
  • User memory: if user/ exists, a single AGENTS.md is bundled as per-user memory (from user/AGENTS.md if present, otherwise empty). At runtime, each user gets their own copy (seeded on first access, never overwritten). The agent can read from and write to this file.
  • MCP servers: if mcp.json exists, it is included in the deployment and langchain-mcp-adapters is added as a dependency. Only HTTP/SSE transports are supported (stdio is rejected at bundle time).
  • Model dependencies: the provider: prefix in the model field determines the required langchain-* package (e.g., google_genai -> langchain-google-genai).
  • Sandbox dependencies: the [sandbox].provider value maps to its partner package (e.g., daytona -> langchain-daytona).

​
[sandbox]

Configure the isolated execution environment where the agent runs code. Sandboxes provide a container with a filesystem and shell access, so untrusted code cannot affect the host. For supported providers and advanced sandbox configuration, see sandboxes. When omitted or set to provider = "none", the sandbox is disabled. Sandboxes are for if you need code execution or skill script execution.
​
provider
string
default:"none"
Sandbox provider. Determines where the container runs. Supported values: "none", "daytona", "modal", "runloop", "langsmith" (private beta). See sandbox integrations for provider details.
​
template
string
default:"deepagents-deploy"
Provider-specific template name for the sandbox environment.
​
image
string
default:"python:3"
Base Docker image for the sandbox container.
​
scope
string
default:"thread"
Sandbox lifecycle scope. "thread" creates one sandbox per conversation. "assistant" shares a single sandbox across all conversations for the same assistant.
Scope behavior:
  • "thread" (default): Each conversation gets its own sandbox. Different threads get different sandboxes, but the same thread reuses its sandbox across turns. Use this when each conversation should start with a clean environment.
  • "assistant": All conversations share one sandbox. Files, installed packages, and other state persist across conversations. Use this when the agent maintains a long-lived workspace like a cloned repo.

​
.env

Place a .env file alongside deepagents.toml with your API keys: 
# Required — model provider keys
ANTHROPIC_API_KEY=sk-...
OPENAI_API_KEY=sk-...
# ...etc.

# Required for deploy and LangSmith sandbox
LANGSMITH_API_KEY=lsv2_...

# Optional — sandbox provider keys
DAYTONA_API_KEY=...
MODAL_TOKEN_ID=...
MODAL_TOKEN_SECRET=...
RUNLOOP_API_KEY=...

​
Sandbox providers

Set [sandbox].provider in deepagents.toml and add the required env vars to .env. For available providers, see sandbox integrations. For lifecycle patterns and SDK usage, see sandboxes.

​
Deployment endpoints

The deployed server exposes:
  • MCP: call your agent as a tool from other agents
  • A2A: multi-agent orchestration via A2A protocol
  • Agent Protocol: standard API for building UIs
  • Human-in-the-loop: approval gates for sensitive actions
  • Memory: short-term and long-term memory access

​
Examples

A content writing agent with per-user preferences that the agent can update:
deepagents.toml
[agent]
name = "deepagents-deploy-content-writer"
model = "google_genai:gemini-3.1-pro-preview"

my-content-writer/
├── deepagents.toml
├── AGENTS.md
├── skills/
│   ├── blog-post/SKILL.md
│   └── social-media/SKILL.md
└── user/
    └── AGENTS.md          # writable — agent learns user preferences
A coding agent with a LangSmith sandbox for running code:
deepagents.toml
[agent]
name = "deepagents-deploy-coding-agent"
model = "google_genai:gemini-3.1-pro-preview"

[sandbox]
provider = "langsmith"
template = "coding-agent"
image = "python:3.12"

​
User Memory

User memory gives each user their own writable AGENTS.md that persists across conversations. To enable it, create a user/ directory at your project root: 
user/
└── AGENTS.md          # optional — seeded as empty if not provided
If the user/ directory exists (even if empty), every user gets their own AGENTS.md at /memories/user/AGENTS.md. If you provide user/AGENTS.md, its contents are used as the initial template; otherwise an empty file is seeded. At runtime, user memory is scoped per user via custom auth (runtime.server_info.user.identity). The first time a user interacts with the agent, their namespace is seeded with the template. Subsequent interactions reuse the existing file — the agent’s edits persist, and redeployments never overwrite user data.

​
How it works

  1. Bundle time — the bundler reads user/AGENTS.md (or uses an empty string) and includes it in the seed payload.
  2. Runtime (first access) — when the agent sees a user_id for the first time, it writes the AGENTS.md template to the store under that user’s namespace. Existing entries are never overwritten.
  3. Preloaded — the user AGENTS.md is passed to the memory middleware, so the agent sees its contents in context at the start of every conversation.
  4. Writable — the agent can update it using the edit_file tool. The shared AGENTS.md file and skills folder are read-only.

​
Permissions

PathWritableScope
/memories/AGENTS.mdNoShared (assistant-scoped)
/memories/skills/**NoShared (assistant-scoped)
/memories/user/**YesPer-user (user_id-scoped)

​
User identity

The user_id is resolved from custom auth via runtime.user.identity. The platform injects the authenticated user’s identity automatically — no need to pass it through configurable. If no authenticated user is present, user memory features are gracefully skipped for that invocation.

​
Gotchas

  • AGENTS.md and skills are read-only at runtime. Edit source files and redeploy to update them. The per-user AGENTS.md at /memories/user/AGENTS.md is the exception — it is writable by the agent.
  • Full rebuild on deploy: deepagents deploy creates a new revision on every invocation. Use deepagents dev for local iteration.
  • Sandbox lifecycle: Thread-scoped sandboxes are provisioned per thread and will be re-created if the server restarts. Use scope = "assistant" if you need sandbox state that persists across threads.
  • MCP: HTTP/SSE only. Stdio transports are rejected at bundle time.

​
Limitations

  • MCP: HTTP/SSE only. Stdio transports are rejected at bundle time.
  • No custom Python tools. Use MCP servers to expose custom tool logic.
Edit this page on GitHub or file an issue.
Connect these docs to Claude, VSCode, and more via MCP for real-time answers.