Agent-Skills.md

Agent Skills: Build a New Fivetran Connector

>-

dataID: kilo-org/kilo-marketplace/build-connector

Repository

kilo-org/kilo-marketplace
Kilo-OrgLicense: Apache-2.0
14971

Install this agent skill to your local

pnpm dlx add-skill https://github.com/Kilo-Org/kilo-marketplace/tree/HEAD/skills/build-connector

Skill Files

Browse the full folder contents for build-connector.

Download Skill

Loading file tree…

skills/build-connector/SKILL.md

Skill Metadata

Name
build-connector
Description
>-

Context: This plugin is for the Fivetran Connector SDK (CSDK). "CSDK" is shorthand for "Connector SDK".

Build a New Fivetran Connector

FIRST: Read sdk-reference.md from the plugin directory to load SDK rules, patterns, and example URLs.

You are building a complete Fivetran connector from the user's description. This skill orchestrates the build; detailed per-phase logic lives in the plugin's workflow files.

Phase 0: Determine the Best Build Approach

Before doing any API research or code generation, check whether a simpler path exists. A custom CSDK connector means the user writes and maintains code; other paths may let Fivetran manage the connector for them. Always surface better options to the user before building custom.

Run all three checks in order and present findings to the user. The user decides — if they prefer to build custom even when a managed option exists, proceed to Phase 1.

Check 1: Is there already a Fivetran connector for this source?

Fetch all five category pages in parallel using WebFetch and scan each for the source name the user mentioned (case-insensitive fuzzy match):

  • https://fivetran.com/docs/connectors/databases
  • https://fivetran.com/docs/connectors/events
  • https://fivetran.com/docs/connectors/files
  • https://fivetran.com/docs/connectors/functions
  • https://fivetran.com/docs/connectors/applications

Ask each fetch: "List all connector names on this page, one per line."

  • If you find a match or a very close match in any page, tell the user:

    "Would you prefer to use our {Match} managed native connector or need to customize it for your use case? The connector is maintained by Fivetran — no code to write or maintain. You can configure it directly in the Fivetran dashboard."

    Wait for the user's answer. If they choose the managed connector, stop and direct them to the Fivetran dashboard. If they want custom CSDK anyway, go straight to Phase 1.

  • If there's no match, the applications page may be incomplete (it is a large list and can be truncated). As a fallback, probe a few direct URL slug variations using WebFetch — try the most likely slugs derived from the source name (lowercase, spaces to hyphens, common abbreviations) across all categories:

    https://fivetran.com/docs/connectors/applications/{slug}
https://fivetran.com/docs/connectors/databases/{slug}
https://fivetran.com/docs/connectors/files/{slug}
https://fivetran.com/docs/connectors/events/{slug}

    For example, for "Google Analytics 4" try: google-analytics-4, google-analytics. For "Amazon RDS for MySQL" try: amazon-rds-mysql, mysql-rds. Fetch 2–3 variations — if a page returns a valid connector description, a match exists.

    If the slug probes also return no match, tell the user:

    "I didn't find {Source} in Fivetran's connector catalog, but the catalog may be incomplete. Please quickly verify at https://fivetran.com/integrations before we proceed."

    Proceed to Check 2 once the user confirms.

Check 2: Is this a good fit for a Lite Connector (AI builder)?

Lite Connectors are built with Fivetran's AI builder and managed by Fivetran — the customer doesn't write or maintain any code. Docs: https://fivetran.com/docs/connectors/applications/lite-connectors

Criteria for recommending Lite:

  • Source is a SaaS application (not a database, file system, event stream, or in-house system)
  • Exposes a REST API with JSON responses
  • Uses standard authentication (API key, Bearer token, or OAuth 2)
  • No complex stateful logic or heavy transformations needed

If the user's source fits these criteria, tell them:

"Based on what you described, {Source} looks like a good fit for a Fivetran Lite Connector. Lite Connectors are built with Fivetran's AI builder and are managed by Fivetran — you won't have to write or maintain any connector code. Fivetran handles API changes, bug fixes, and upgrades. Want to try the Lite path first? It's typically faster to set up and lower ongoing maintenance than a custom CSDK connector."

Wait for the user's answer. If they choose the Lite path, stop and direct them to the Lite Connector builder. If they want custom CSDK, continue to Phase 1.

Check 3: Proceed with custom CSDK

If neither Check 1 nor Check 2 produced a better option, or the user explicitly chose custom CSDK, continue to Phase 1.

Phase 1: Research & Validate Requirements

Apply the validator workflow — read workflows/validator.md in the plugin directory (or, in plugins that support subagents, invoke the connector-validator subagent) to research the API and produce a complete specification and a discovery result: EXACT MATCH, FUZZY MATCH, or BUILD ON TEMPLATE.

Stop and wait for the user to answer any clarifying questions the validator surfaces before proceeding to Phase 2.

Phase 2: Scaffold the Project with fivetran init

fivetran init is the canonical scaffolding path — it produces a complete, runnable connector with the correct structure (validate_configuration(), docstrings, the __main__ block). Always scaffold with fivetran init; never hand-write the project from scratch. Pick the command from the Phase 1 discovery result. The project directory is the connector name (lowercase, underscores).

  • EXACT MATCH / FUZZY MATCH — start from the community connector: 
    printf '\n' | fivetran init "<connector_dir>" --template connectors/<name> --force
  • BUILD ON TEMPLATE — start from the default template: 
    printf '\n' | fivetran init "<connector_dir>" --force

Windows PowerShell: replace printf '\n' | with "" |.

Why the piped newline and --force: fivetran init always runs an interactive "which coding agent shall we install the plugin for?" prompt, and there is no flag to skip it. --force auto-confirms project creation and file overwrites; the piped empty line answers the agent prompt with an invalid choice, so it logs invalid choice; skipping agent setup (this is expected and benign — the plugin is already installed) and continues.

Verify success by checking that <connector_dir>/connector.py exists, not by the exit code — an exhausted input pipe can make init exit non-zero even after the files download correctly. connectors/<name> resolves to the community_connectors repo; examples/<path> resolves to connector_sdk.

Phase 3: Customize the Scaffolded Files

Apply the generator workflow — read workflows/generator.md (or invoke the connector-generator subagent) to adapt the already-scaffolded files to the Phase 1 specification using Edit (do not rewrite from scratch). For an EXACT MATCH, the connector may need little or no change beyond configuration; for FUZZY/BUILD ON TEMPLATE, study 2–4 relevant SDK examples and adapt connector.py, configuration.json, and README.md to the spec. Preserve the template structure (validate_configuration(), docstrings, global connector = Connector(...), the __main__ block).

Phase 4: Setup Environment

fivetran init already created requirements.txt (or pyproject.toml). Set up the virtual environment and install dependencies with commands appropriate for the user's OS:

macOS/Linux:

cd "<project_directory>"
uv venv .venv
uv pip install --python .venv/bin/python -r requirements.txt fivetran_connector_sdk

Windows PowerShell:

cd "<project_directory>"
uv venv .venv
uv pip install --python .\.venv\Scripts\python.exe -r requirements.txt fivetran_connector_sdk

Phase 5: Enter Configuration & Test

After scaffolding and customizing the files (or finding existing files with placeholder values), configuration values must be entered via the encryption script. This is not negotiable and not a user choice — it is the only supported configuration-entry flow.

HARD RULES — violating any of these is a failure:

  • DO NOT use AskUserQuestion (or any choice-menu / multi-option UI) to ask how the user wants to enter configuration values. There is exactly one way.
  • DO NOT present "Paste in chat", "Edit the file yourself", "Use a public repo", or any other option as a configuration-entry choice.
  • DO NOT tell the user to paste configuration values in chat.
  • DO NOT accept values pasted in chat.
  • DO NOT run enter_configuration.py yourself. The user must run it in their own separate terminal.
  • DO NOT inspect or print configuration values.

THE ONLY ACCEPTABLE FLOW. Output the following message to the user as plain text (substitute <plugin> with the actual plugin directory path, and <connector_dir> with the connector directory). Use one fenced command block: bash on macOS/Linux, powershell on Windows. Quote both paths. Do not insert a line break inside the python command.

I've generated the connector files (or the files already exist). To fill in configuration values securely, open a separate terminal, then run:

```bash
cd "<connector_dir>"
python "<plugin>/tools/enter_configuration.py" "configuration.json"
```

The script will prompt you for each configuration field and encrypt values in place. I never see plaintext configuration values. Let me know when it's done and I'll run the test.
If the local encryption secret file does not exist yet, the script creates it first.

After the user confirms configuration values are entered, run the connector via the secure runner:

python <plugin>/tools/run_connector.py <connector_dir>

This decrypts encrypted configuration values in memory, passes user-chosen plaintext values through, and supplies the runtime config via named pipe.

Check results:

macOS/Linux:

.venv/bin/python -c "
import duckdb
conn = duckdb.connect('files/warehouse.db')
tables = conn.execute(\"SELECT table_name FROM information_schema.tables WHERE table_schema = 'tester'\").fetchall()
print(f'Tables synced: {len(tables)}')
for (t,) in tables:
    count = conn.execute(f'SELECT COUNT(*) FROM tester.{t}').fetchone()[0]
    print(f'  tester.{t}: {count} rows')"

Windows PowerShell:

.\.venv\Scripts\python.exe -c 'import duckdb; conn = duckdb.connect("files/warehouse.db"); tables = conn.execute("SELECT table_name FROM information_schema.tables WHERE table_schema = ''tester''").fetchall(); print("Tables synced:", len(tables)); [print("  tester." + t + ": " + str(conn.execute("SELECT COUNT(*) FROM tester." + t).fetchone()[0]) + " rows") for (t,) in tables]; conn.close()'

Report: tables synced, row counts, any errors.

Phase 6: Auto-Fix on Failure

If the test fails:

  1. Read the error output carefully.
  2. Classify the error:
    • INFRA error (network, JVM, SDK internal): explain the infrastructure issue. Do NOT change code.
    • FIRST_RUN error (connector has never succeeded — likely credentials/config): guide the user to verify config. Do NOT change code.
    • CODE error (syntax, logic, SDK misuse): apply the fixer workflow — read workflows/fixer.md (or invoke the connector-fixer subagent). Re-test after fixing.
    • TOOL error (run_connector.py fails): report to the user. Do NOT modify plugin tools.

IMPORTANT: Never modify plugin tools. Only fix the user's connector code.