ODI Open Design Intelligence Open source

Open-source design intelligence / MCP

Better design judgment for coding agents.

Connect ODI once. Your agent can search 1,320 reviewed design references, reason from the decisions behind them, and build with evidence instead of guessing.

Connect the MCP View on GitHub
Corpus
1,320 references
Interface
9 MCP tools
Access
No account
Runtime
No model inference

Example output

See what your agent gets back.

A useful reference includes the decisions to adapt, the evidence behind them, and the product questions it cannot answer.

Example MCP trace recommend_designs / explain_design / design_synthesis Evidence boundary: explicit

Input / product brief

Design a clear personal finance dashboard for young adults. Prioritize monthly budgets, upcoming bills, and accessible spending alerts.

Retrieved direction

  • IBM Structured financial information hierarchy
  • Linear Restrained geometry and compact controls
  • Sentry Explicit states and accessible feedback

Agent handoff

Layout
Overview first, transaction detail on demand
Signals
Text and shape before color
Unknown
Account linking and financial guidance
Missing evidence remains missing

The agent can separate observed decisions, derived guidance, and product assumptions before it writes code.

MCP capabilities

What your agent can call.

Nine focused tools cover discovery, evidence inspection, design direction, and review.

Callable tools
9
Design jobs
4

  1. Discover

    Find relevant systems, related references, and recurring decisions.

    Calls
    search_designsfind_similarfind_patterns

  2. Inspect

    Retrieve canonical evidence and understand why a reference fits.

    Calls
    get_designexplain_design

  3. Decide

    Choose a direction without overstating the available evidence.

    Calls
    recommend_designsdesign_consensus

  4. Create and review

    Produce an implementation brief or critique design documentation.

    Calls
    design_synthesisdesign_critique

Connect in minutes

Give your agent a better design vocabulary.

The hosted server uses Streamable HTTP and requires no credentials, local process, or repository clone.

Hosted MCP endpoint https://open-design-intelligence.vercel.app/mcp
Local stdio with npx npx -y open-design-intelligence@1.0.0

Requires Node.js 22+. The process waits for an MCP client, so a silent terminal is expected.

Hosted setup

Connect Codex

  1. 01

    Open ~/.codex/config.toml or the project-level config.

  2. 02

    Add the remote server entry, then restart the client.

Codex config.toml

[mcp_servers.open-design-intelligence]
url = "https://open-design-intelligence.vercel.app/mcp"

Verify the connection

Ask one real design question.

Your agent should recommend references, explain its reasoning, and state what the corpus does not establish.

Using the Open Design Intelligence MCP, shape the design direction for an online store used by everyday shoppers.

The main journey is finding a product, comparing options, adding it to a cart, checking out, and tracking delivery. The experience should feel clear and trustworthy, work well on phones and desktop, and make prices, stock, errors, and order status understandable without relying on color alone.

Recommend 2–3 relevant references. Explain the specific decisions worth adapting and what should not be copied. Then produce an implementation brief covering layout, hierarchy, typography, interaction states, responsive behavior, and accessibility.

Separate corpus-backed guidance from product assumptions and identify anything the references do not establish.

Common questions

What to know before you connect.

Does ODI generate the interface?

Your coding agent still designs and implements the interface. ODI gives it stronger references, explicit constraints, and a way to inspect what is known or missing before it builds.

Does the hosted MCP require an account?

No. The public Streamable HTTP endpoint needs no credentials, package installation, repository clone, or local process.

What is the corpus based on?

ODI searches reviewed, canonical DESIGN.md records and their classification metadata. It separates direct evidence, derived observations, and assumptions, and treats missing information as unknown.

How does ODI choose relevant references?

ODI ranks the reviewed corpus using deterministic lexical, trait, quality, taxonomy, and local similarity signals. It does not call an embedding or model API at query time.

Can I run it locally?

Yes. With Node.js 22 or newer, run npx -y open-design-intelligence@1.0.0 as a local stdio server. Pinning the package version also pins the bundled corpus.