design-agent

Content Studio is the design agent LangChain runs, now open source. Add your brand and deploy it to your team's Slack with Managed Deep Agents. Anyone can ask it like they'd ask a designer, attach what they have, and get finished, on-brand files back, from a resized ad to a launch film.

Since August it has shipped over 3,000 assets for us, posts with its work drew about 4.5M views on X, and three in four requests come from outside marketing. How we built it goes deeper into each part.

See it work

We asked Content Studio for its own launch film. It went full Rocky.

https://github.com/user-attachments/assets/ee5b98e0-7ec6-4b24-ada6-1f27087947a3

How it works

Content Studio is a Deep Agents agent deployed on Managed Deep Agents. Managed Deep Agents runs the infrastructure: Slack, sandboxes, memory, schedules and tracing. This repo is the design agent on top. The model makes the creative calls, and code renders, checks and delivers.

• Slack. Mention it in a thread with a brief and any files. MDA's Slack app hands the message and attachments to the agent, and the finished files come back to the same thread. While it works, one message updates step by step, a footer shows what's running and the cost so far, and anything you add mid-job reaches the work in progress.
• A computer per thread. Each thread gets its own LangSmith sandbox: a Linux machine with Chrome, FFmpeg, HyperFrames, D2 and your fonts, built from one pinned recipe (sandbox/setup.sh). Everything it makes is code in a project folder there (a slide is HTML, a film is an animation timeline), so a change is an edit and a re-render, not a restart. API keys never enter it.
• Skills and a design wiki. The always-on prompt is small. For each ask it first reads a router skill that picks the right workflow, from a resized ad to a launch film, out of 35 skills. The wiki holds the craft (type, colour, layout, charts) and your brand pack holds the facts. Skills load only when a job needs them, served by MDA from Context Hub.
• Your tools. Video is built in HyperFrames, which treats a film as a web page on a timeline. Paper is a shared canvas where you edit a storyboard or layout directly. Image, voice and music models run as tools outside the sandbox, and MCP servers like Notion, Linear, GitHub and Figma connect through MDA, with each person signed in as themselves.
• Async subagents. On a big job the main agent plans, splits the work and starts subagents in the background on the same sandbox and files. MDA wakes it when one finishes, fails or stalls, so it never sits waiting and can keep talking to you. Films go storyboard, animatic, then production, and code makes every part pass review before assembly.
• Memory. Tell it a rule once ("never use orange on LinkedIn") and it saves it that turn. MDA gives it two layers of files: team memory everyone shares and personal memory for each person, and each layer's index loads into every run.
• It gets better. After a job, learn_from_job writes notes on what worked against the skill it used. Later jobs read and rate those notes, and the ones that don't hold up are dropped.

Watch it explain itself

We asked it to explain its own harness as a nature documentary, starring a pixel parrot. It walks through the computer, the tools and the context behind its work.

https://github.com/user-attachments/assets/65eac8e4-0fac-4a5a-9cca-02e26a0a8baa

The agent made it from this ask and a few rounds of notes:

Make a short explainer of how you create videos in code, end to end, for our blog post on the media harness we built on Deep Agents. Narrate it like a David Attenborough nature documentary, starring the pixel parrot from our promo, with your subagents as its flock. Keep it technical enough for an engineer, in real terms: skills, wiki, tools, apps. Storyboard it before you build, and end on "Built on Managed Deep Agents".

Read the architecture, or start changing it: instructions.md · skills/ · Customization · Evals · Operations · Contributing.

Why it doesn't look like AI

• It looks at its work. It renders, then looks at the image before it posts anything.
• Your brand is code. Layouts only take your tokens: your colours, fonts and spacing.
• It checks what a machine can count. Colours, fonts, contrast, sizes and chart math are checked in code on every render. Taste stays with your designers.

Quickstart

Building it yourself	Starting here
Months on a harness, a sandbox, skills and a Slack UX	Fork it, add your brand, deploy with one command
Polished work that looks like a hundred others	Work that follows your brand, checked on every render
Design requests wait in a queue	Anyone asks in Slack: a graphic in minutes, a long deck overnight
Every team fixes the same mistakes	Fix a skill once, and every asset after it follows

This repository is a Managed Deep Agents project, so it deploys to your LangSmith workspace as it is. You need a LangSmith account (Managed Deep Agents is in public beta on LangSmith Cloud) and a model key:

curl -fsSL https://raw.githubusercontent.com/langchain-ai/design-agent/main/install.sh | bash
cd design-agent
uv run design-agent config set LANGSMITH_API_KEY=lsv2_… ANTHROPIC_API_KEY=sk-ant-…
uv run design-agent sandbox publish   # once per workspace, 10-15 min: the agent's computer
uv run mda deploy                     # the agent itself, and its Slack app

The installer gets uv if you don't have it, clones the repository and installs its dependencies; git clone and uv sync do the same.

The two last commands do different things. sandbox publish builds the Linux machine every thread works on, as a snapshot in your LangSmith workspace; nothing you can talk to exists yet. mda deploy deploys the agent, which clones that machine for each thread. Publish again only when sandbox/setup.sh changes.

On the LangSmith Gateway? Skip the provider key and run uv run design-agent config set DESIGN_AGENT_MODEL=langsmith:anthropic/claude-opus-5-5. The agent then uses LANGSMITH_GATEWAY_API_KEY, else LANGSMITH_API_KEY, and image generation goes through the gateway too when there is no OPENAI_API_KEY. For any other OpenAI-compatible endpoint, set DESIGN_AGENT_MODEL_BASE_URL and DESIGN_AGENT_MODEL_API_KEY_ENV (Models).

The first deploy sets up the agent's Slack app. Authorize it in your workspace, then mention the agent in a channel. It's how we run ours at LangChain. Try changes on a separate copy with uv run mda deploy --name design-agent-test. In a workspace with agent mode, uv run mda deploy --agent-environment production (or staging, development) deploys into that environment of the agent, and each environment keeps its own traces: open the agent in LangSmith and pick the environment. uv run mda dev traces to its Local environment. Deploy with MDA covers keys, what ships and checking it works.

Or ask your coding agent. Open the repo in Claude Code, Codex or Cursor and say what you want. AGENTS.md routes each request to a step-by-step skill:

Say	It will
"Deploy it to our Slack."	publish the sandbox, check what ships, and deploy after your yes
"Build our brand pack from our website and brand guide."	draft the pack, validate it, and ask you to review it
"Teach it to make our weekly newsletter header."	add a skill for it and try it on a real request
"Did that change make it better?"	run the evals before and after, and show you the numbers

Try it locally

The demo runs the real agent, sandbox and checks with a scripted model and no API keys. It needs Docker, and leaves a launch card in three sizes and a 4-second teaser in workspace/out/demo/:

uv run design-agent sandbox build
uv run design-agent demo

With a model key in .env, ask for something real. Files land in workspace/out/<thread>/, and the last line tells you what the run cost:

uv run design-agent ask "A LinkedIn card announcing our design review service, 1200x627."

Make it yours

This repo has two kinds of skills. The ones in .agents/skills/ are for your coding agent, to set up, brand, deploy and change this repo. The ones in skills/ are what Content Studio reads while it works. Customization covers both.

Want	Add	Guide
Your brand	workspace/brand/ (until then, a neutral sample brand), then redeploy	brand-onboarding skill
Generated images	OPENAI_API_KEY, or the LangSmith Gateway	Models
Voice, music, sound effects	ELEVENLABS_API_KEY	sound skill
The Paper canvas	PAPER_MODE=local with Paper Desktop, or bridged when deployed	Paper
Web search, your docs, other tools	MCP servers in config/mcp.toml	Customization

Once it's yours, ask it in Slack like you'd ask a designer:

• "A launch card for our new feature, in LinkedIn, X and Instagram sizes."
• "A Rocky-themed promo film for our launch, with music and sound effects."
• "A chart of our weekly signups from this CSV." (attach the file)
• "A YouTube thumbnail for this video."
• "Caption this webinar clip." (attach the clip)

Everything on this page uses LangChain's brand, because that's the one we run. Licensed fonts go in private-fonts/, which Git ignores.

Questions

What does it cost? The code is free (Apache 2.0). You pay your model provider: at LangChain the median request costs about 50 cents in model spend, and long films and decks cost more. Managed Deep Agents hosting is paid; it runs the agent, the sandboxes, the Slack app and the schedules that you'd otherwise build and operate yourself.

Will it use my brand? Yes. Add your brand pack and every file follows it. The brand-onboarding skill builds the pack with you from your brand guide, website, files or a Paper file.

Where do I start? With the graphic your team asks for every week. Let people brief the agent, review the work and ask for changes in Slack, then teach it the next thing.

Do I need Slack? No. Call the deployment's API, open it in LangSmith Studio, or use the terminal (design-agent ask). On Managed Deep Agents, long jobs that work in the background (films, parallel looks) report back in Slack; over the API, keep to requests that finish in one turn.

Is it safe with my keys? The sandbox never holds a credential and has no network by default. Host tools do the fetching, generating and delivering, and drop results in as files.

Which models? Claude Opus 5.5 by default. OpenAI and Google models, any OpenAI-compatible endpoint and the LangSmith Gateway work too. See Models.

Can I run it on my own servers? Yes. It's Apache 2.0, and Self-hosting runs the same agent with Docker, an HTTP API and your own Slack app.

Why two names? Content Studio is what we call the agent at LangChain, and the name its Slack app gets unless you set DESIGN_AGENT_SLACK_NAME. The repository, package and command are design-agent.

Does it replace a designer? No. It takes the repetitive requests, so designers spend their time on the work only they can do, and their judgment lives on in its skills and checks.

Build your own. Fork it, add your brand tokens, fonts and skills, and deploy it to your team's Slack. Quickstart. If it's useful, a star helps other teams find it.

Built with

Also: LangSmith · Chrome Headless Shell · FFmpeg · D2 · Apache ECharts · sharp. Logos belong to their owners. HyperFrames and Paper are partner projects we build on.

---

Apache 2.0 · Notices · Changelog · Security · Code of conduct