Duchamp

Duchamp — ad research for coding agents

Duchamp is an ad-research tool for Claude Code, Codex, and Cursor. It finds the ads curated advertisers have kept running for 90+ days, and you save, compare, and pull the research materials into your agent, on the web or over one remote MCP server: https://duchamp.app/api/mcp.

Call search_ads first, then get_ad or get_brand_anatomy. Creative files, ad copy, keyframes, and transcripts need the signed-in account. Lists and counts on the site are public. A reference prompt from compose_transfer_brief is a reference. It does not promise the same look or the same outcome. The original ad stays a link (duchamp_url). get_ad returns timestamped keyframes, the speech transcript, full body copy, the landing URL, and observed numbers, plus a stored AI analysis when one exists (format, hook, 15 emotional drivers, persona, creative targeting, product category, first product appearance — filled by a nightly batch, Foreplay-style). Treat the analysis as a reference and check it against the materials; search_ads can filter by its fields. Product facts passed to compose_transfer_brief are not stored. Saved items, boards, and watched advertiser pages are stored for the signed-in user. Those tools need a separate per-app opt-in on duchamp.app/saved/connect.

38 tools are registered: 26 ad-research tools, 6 Performance tools, 2 label tools, and 4 marketing workspace tools. The table below is that set.

Performance (your own Meta ad accounts, read only): connect an ad account on duchamp.app/performance, group accounts into workspaces (one per client or brand), then allow the app on duchamp.app/performance/connect. list_workspaces → get_performance_insights with a workspace_id or account_ids (group by ad name, copy, creative, landing page, headline, campaign, ad set, account, workspace or Audience Segment — New Audience, Engaged Audience, Existing Customer, sorted by the workspace's campaign-name rules or Duchamp's defaults; by a Meta breakdown at account level — country, age, gender, platform, placement, Meta customer segment; or compare rule groups) → get_performance_timeseries / get_performance_creative / get_performance_ad. Scores (hook, watch, click, convert, viral) are percentiles among every ad (or group) of the workspace in the same period — they are not an industry benchmark. Numbers use the ad set attribution setting and impression-time reporting unless you pick a window with attr_click (1d, 7d, 28d) and attr_view (none, 1d) — Meta no longer returns 7-day or 28-day view. Besides spend, ROAS, CPA, CTR and video rates there are website and Shop purchases, landing page views, add to cart, checkout, leads, drop-off rates and your custom conversions (cc_<id>_total, _value, _cost, _roas, listed by get_performance_metrics with a workspace_id).

Team boards (read only): list_boards with scope: "team" or "all" adds the team's boards (marked scope: "team"); pass board_scope: "team" to get_board, export_board and compose_transfer_brief, and use list_saved with scope: "team" for teammates' shared ads. This needs a separate team access per app on duchamp.app/saved/connect, and only teammates who turned on AI sharing appear — no notes, saved titles, link items or comments.

• Website: duchamp.app · English overview: duchamp.app/en
• Connect page: duchamp.app/mcp
• MCP endpoint: https://duchamp.app/api/mcp (Streamable HTTP, OAuth 2.1 with a free Duchamp account, no API key)
• For AI readers: llms.txt · llms-full.txt

What you get

Do not invent numbers, dates, prices, or reviews. Transcripts are raw speech-to-text and can mishear names. A person should check what the agent concludes.

Connect in one line

sh -c "$(curl -fsSL https://duchamp.app/connect.sh)"

The script finds Claude Code and Codex, registers the server, and opens your browser to sign in with a free Duchamp account. There is no token to copy.

Manual setup

Claude Code

claude mcp add duchamp --scope user --transport http https://duchamp.app/api/mcp

Then in a new session run /mcp, choose duchamp and press Authenticate.

Codex

codex mcp add duchamp --url https://duchamp.app/api/mcpcodex mcp login duchamp

Cursor — add to .cursor/mcp.json, then press Connect next to duchamp in Cursor's MCP settings:

{  "mcpServers": {    "duchamp": { "url": "https://duchamp.app/api/mcp" }  }}

Any other MCP client — add a remote server named Duchamp with transport Streamable HTTP and URL https://duchamp.app/api/mcp. The first connection opens a browser for OAuth.

Plugin install (skills + agent)

The plugin adds eleven skills (nine active, two paused) and a judge agent on top of the MCP server.

Claude Code:

/plugin marketplace add duchamp-app/duchamp/plugin install duchamp@duchamp

Codex:

codex plugin marketplace add duchamp-app/duchampcodex plugin add duchamp@duchamp

Then sign in once: Claude Code /mcp → duchamp → Authenticate; Codex codex mcp login duchamp.

• ads-research: find working ads, open one ad with its materials, break down an advertiser, compare your page, and keep a saved library. This is the skill to use.
• Research skill packs — workflows the agent runs end to end. Say "use the skill on …":
  • evergreen-hunt: ads that ran 90+ days in a niche, clustered by hook, with transcripts for the strongest.
  • weekly-competitor-sweep: your watched advertisers' last 7 days as a Slack-ready Monday brief.
  • creative-test-autopsy: what one advertiser is testing now vs what it keeps scaling.
  • board-to-brief: a named board turned into a production brief with three angles grounded in the saved ads.
  • hook-mining: a hook bank for a niche and format, ranked by how long the ads ran.
  • domain-war-room: a competitor's domain turned into its ads, hooks, landing pages, paid partners and attack angles.
  • angle-theft-detector: your winning creatives from Performance (your own Meta ad accounts) matched to lookalike ads from other advertisers, with how crowded each angle is.
• connect-extension: install and connect the Duchamp Chrome extension (opens the Chrome Web Store page in your Chrome, you click Add and Connect), then confirm the first save from the Meta Ad Library.
• make and launch-kit: paused. They call tools that are not registered. Do not call those tools. Use ads-research.
• visual-judge agent (Claude Code): a fresh-context reviewer. The make skill that used to call it is paused.

Everything runs on your own Claude Code or Codex subscription.

Tools

The ad research tools are not on a daily fetch meter. 38 tools are in the table.

Example prompts

First prompts (replace the [ ] parts):

• Brand snapshot — "Using the Duchamp MCP, look up [advertiser name or homepage] (find_advertisers_by_domain for a homepage, or facets.pages of search_ads for a name). Show the 6 newest live ads with search_ads (page_id, status active, sort recent, all true) and link each duchamp_url. Then 3 bullets: formats, offers, the most common angle."
• UGC script rewrite — "search_ads for live selfie-style video ads in [niche], pick the longest-running, pull its transcript with get_ad and rewrite it as a 30-second script for [your product]. Keep the structure, change the words."
• Offer scan — "search_ads for live ads in [niche] and list every offer (discounts, bundles, free shipping, guarantees, trials), ranked by how often they appear, with one ad for each of the top 3."
• Format mix check — "Count live ads in [niche] by medium (video, image, carousel) with search_ads, show the longest-running example of each and the under-used format."
• Ad copy swipe — "Find 10 image ads in [niche] that ran 90+ days with search_ads, pull their copy with get_ad and rewrite the best 5 for [your product]."

More:

• "Find long-running ads in my niche with search_ads. List the advertiser, days running, and duchamp_url. Do not invent impression counts or revenue."
• "Open this ad with get_ad, then get_brand_anatomy for the advertiser. Read the keyframes and transcript, then tell me the hook and the offer. Link duchamp_url."
• "Watch my page with watch_brand and is_mine true, watch a competitor, then compare_brands. At most four pages."

How it works

1. Check the connection with get_connection_status. If the saved library is not allowed for this app, open the link it returns before saving.
2. Search with search_ads.
3. Open one ad with get_ad, or one advertiser with get_brand_anatomy.
4. Compare with compare_brands (at most four page ids) or save with the library tools after the saved-library opt-in.
5. Hand off with compose_transfer_brief when the user asks. It is a reference prompt. It does not promise the same ad.

Limits & account

• Sign in with a free Duchamp account on the first connection (OAuth in your browser). No API key.
• The ad-research tools are not counted toward a daily limit.
• Pro is listed at $19/month or $190/year (5 competitor pages of your own collected daily, unlimited MCP). Once payments open, Free keeps search unlimited and opens 10 ads a day over MCP. Payments are not open yet, so everything is free with no limits for now. See duchamp.app/en/pricing.
• Items appear on their release date. Ids not listed on the site are not available.

Privacy

Duchamp stores your account, usage records, saved items, boards, and watched advertiser pages. If you connect an ad account to Performance, Duchamp stores that account's ad metadata and daily performance numbers (and an encrypted read-only access token) for your account only — never in the public library or a team — and deletes them when you turn the account off, disconnect, or remove the app in Facebook settings. Performance numbers go to an AI app only after you allow that app on the Performance connect page. Product facts passed to compose_transfer_brief are not stored. Your rendered files stay on your machine. Duchamp does not call a model on your requests. Speech transcripts are made once per video when it is collected. Duchamp does not reproduce the ad. Signed-in ad research can include creative URLs on Duchamp's mirror. Reference labels do not include creative files.

What Duchamp does not do:

• No generation of human faces, human likenesses, face swaps, deepfakes, or synthetic voices. Duchamp does not create synthetic media of real people.
• No voice cloning, no audio or music generation, no adult content, no AI companion or relationship features.
• Templates: no display or re-hosting of other people's posts or accounts. Templates carry no original wording, images, or account names.
• Ads research: signed-in callers get creative URLs on Duchamp's mirror. Inline images in tool results are low-resolution previews for the agent. When showing the user, link duchamp_url. Reference labels still do not include creative files. Duchamp does not reproduce the ad.
• No scheduling or publishing. Performance reads only ad accounts you connect yourself (Meta, read only, ads_read) for performance reports. No server-side content generation. No engagement automation, and no access to any account the user does not own.
• No aggregation or resale of personal data, and no performance guarantees.

Rights in the original posts stay with their rights holders. To ask for a template to be taken down, write to [email protected] with the original URL.

Links

• duchamp.app · English overview · Connect · Pricing
• llms.txt · llms-full.txt
• Company: Notique, Seoul · Contact: [email protected]

License

MIT. See LICENSE.