MCP Setup & Usage Guide

1. Go to Settings > Connectors and click Browse connectors.

2. Find Dewey Data, click Connect, and sign in with your Dewey account.

3. In a chat, make sure Dewey is turned on in the + menu.

• Team or Enterprise plan: an Owner must enable Dewey for your organization first.

• Mobile: connect on web or desktop; Dewey then works in the mobile app.

• Claude Code and Cowork: sign in with the same Claude account and Dewey is already there. In Claude Code, type /mcp to authenticate it if prompted.

Source: Anthropic help center

1. Open the Dewey Data app, or go to Settings > Apps.

2. Click Connect and sign in with your Dewey account.

3. In a chat, type @ and pick Dewey, or choose it from the + menu.

• Business, Enterprise, or Edu workspace: your admin controls which apps are on. Apps are off by default on Enterprise and Edu, so ask your admin to enable Dewey.

• Codex: sign in to Codex with the same ChatGPT account and Dewey is available there too.

Source: OpenAI help center

• Option A: run the command or edit the config yourself.

• Option B: paste a prompt and let the AI tool set itself up.

Option A. In your terminal, then restart Gemini CLI and type /mcp:

gemini mcp add --transport http --scope user \
  --header 'X-API-KEY: $DEWEY_API_KEY' \
  dewey https://api.deweydata.io/mcp

Single quotes keep your key out of the file; Gemini CLI reads DEWEY_API_KEY each time it starts. On Windows, use the settings.json option below instead.

Or add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "dewey": {
      "httpUrl": "https://api.deweydata.io/mcp",
      "headers": { "X-API-KEY": "$DEWEY_API_KEY" }
    }
  }
}

Option B. Paste into Gemini CLI:

Add the Dewey MCP server to ~/.gemini/settings.json. Create the file if needed, keep existing settings, and add this under "mcpServers". My API key is in the DEWEY_API_KEY environment variable; don't ask me for it or write it into the file.

"dewey": {
  "httpUrl": "https://api.deweydata.io/mcp",
  "headers": { "X-API-KEY": "$DEWEY_API_KEY" }
}

Tool only supports local servers? Use mcp-remote as a bridge (requires Node.js):

{
  "mcpServers": {
    "dewey": {
      "command": "npx",
      "args": ["mcp-remote", "https://api.deweydata.io/mcp", "--header", "X-API-KEY:${DEWEY_API_KEY}"],
      "env": { "DEWEY_API_KEY": "YOUR_API_KEY" }
    }
  }
}

Note: responses are shortened, and will vary. This example is to showcase an example workflow and possible uses of the MCP connection using LinkUp Job Records.

Step 1. Discover: search by topic

• Prompt: "Find LinkUp job postings data I can use."

• Example response: LinkUp Job Records (Company Insights, global), one row per job posting, plus companion tables such as Job Descriptions, Structured Fields, Extracted Salary, and Remote Tag.

Step 2. Evaluate: review the details

• Prompt: "Tell me about Job Records: coverage, refresh schedule, and size."

• Example response: Job postings scraped from employer websites worldwide, one row per posting. Refreshed monthly on the 6th. About 362 million rows (52.21 GB, Parquet).

Step 3. Evaluate: review the schema

• Prompt: "Which columns hold location and date?"

• Example response: Location: CITY, STATE, ZIP, and COUNTRY; UNMAPPED_LOCATION flags postings whose location couldn't be identified. Date: CREATED, the first time LinkUp saw the posting, which the data is partitioned on.

Step 4. Evaluate: preview the data

• Prompt: "Show 4 sample rows with title, city, state, country, and date."

• Example response: Cities are in title case, US states are two-letter codes, and countries are three-letter codes:

Step 5. Evaluate: check data quality

• Prompt: "Anything I should watch for before counting postings?"

• Example response: The data is global, so filter COUNTRY = 'USA'. A posting with several locations becomes several rows (same BASE_HASH, different JOB_HASH), so count BASE_HASH for unique postings. Missing states can appear as the text "None". And CREATED is when LinkUp first saw the posting; the employer's posted date is in Structured Fields.

Step 6. Evaluate: confirm access

• Prompt: "Can I download it?"

• Example response: Yes. It's included in your subscription and doesn't require customization, but filtering first keeps the download far smaller than 52 GB.

Step 7. Plan: join tables

• Prompt: "What companion tables come with Job Records, and how do they join?"

• Example response: Eight, including Job Descriptions, Extracted Salary, Remote Tag, Skills, and ONET Taxonomy. Job Descriptions and Extracted Salary join on JOB_HASH; company reference data joins on COMPANY_ID.

Step 8. Plan: write download code

• Prompt: "Write the code to get February 2026 job postings for Houston, TX."

• Example response: three steps.

1. In Dewey, add Job Records to a Project. To shrink the download, customize it first (country USA, state TX, city Houston). City filters match any value containing your text, so step 3 narrows to exact matches.

2. Download February 2026 only (the Dewey Client reads DEWEY_API_KEY automatically):

python -m deweypy speedy-download prj_YOUR_PROJECT_ID \
  --partition-key-after 2026-02-01 --partition-key-before 2026-02-28

3. Filter to Houston with DuckDB:

SELECT *
FROM read_parquet('dewey-downloads/**/*.parquet')
WHERE COUNTRY = 'USA'
  AND STATE = 'TX'
  AND CITY = 'Houston'
  AND CREATED >= TIMESTAMP '2026-02-01'
  AND CREATED <  TIMESTAMP '2026-03-01';

Step 9. Plan: cite the data

• Prompt: "How should I cite this dataset?"

• Example response: LinkUp. (2025). Job Records (Version 1.3) [Data set]. Dewey Data. https://doi.org/10.82551/D6Q4-3W96

Note: Before you run AI-written code

• Dewey doesn't run or check this code. Review it and test on a small download first.

• Ask the AI to check value formats in a sample before filtering (here, cities are title case and countries are three-letter codes).

• Ask what a column means before relying on it. Here, CREATED is when LinkUp first saw the posting, not the employer's posted date.

• Decide your geography: the city "Houston" is not the same as the Houston metro area.