Kilde Internal MCP Connector

What is this, in one line? “MCP” is just the plug that lets Claude reach Kilde’s tools. You turn on one connector called “Kilde Internal MCP” with your @kilde.sg account, and then you can ask Claude things like “show me the last messages in #general”, “how many rows are in the invoices table?”, “what PRs are open on kilde-frontend?”, “list my open Pipedrive deals”, “what VWO experiments exist?”, “what’s the GDP growth forecast for Singapore?”, “what are the unresolved Sentry errors from today?”, or “list the Xray tests in project ABC” — and it does it for you. No installing anything on your computer.

Start here — connect once ~2 min

All three tools live behind the same single connector. You only do this part once; after that, picking which tool to use is automatic — you just ask Claude in plain language.

What you need

A Claude account on the Kilde organization — i.e. you signed in to Claude with your @kilde.sg email.
Nothing to install. No coding. No setup files.

Turn on the “Kilde Internal MCP” connector

Open Claude (web, Desktop, or mobile).
Go to Settings → Connectors.
Find “Kilde Internal MCP” in the list and click Connect.
A login window opens — sign in with your @kilde.sg Google account and approve.
Done. You’ll see the connector listed with its tools.

Don’t see “Kilde Internal MCP” in the list? It’s an organization connector that an Owner switches on once for the whole company. If it’s missing, it may not have been rolled out to you yet — ask in #product-and-tech.

In each chat where you want to use it, turn it on with the + (tools) menu in the message box — toggle “Kilde Internal MCP” on. That’s it. Now pick the section below for what you want to do.

The one difference to know: MotherDuck, GitHub, VWO, Trading Economics, Sentry, and Xray work the moment you connect — nothing else to set up. Mattermost and Pipedrive each need one extra one-time step (a personal token) so that Claude acts as you — seeing exactly what you’re allowed to see. Both are enrolled on the same enroll page; details in their sections.

💬 Mattermost — team chat needs your token ~5 min setup

Read and post in Mattermost from Claude — as yourself, in plain English.

In one line: after connecting, you enroll your own Mattermost token once. From then on Claude can reach exactly the channels you have access to, and anything you post shows up authored by you.

Why the extra step?

Mattermost acts as you — so it can only see your channels, and your messages are genuinely from you (not a shared bot). To do that, Claude needs a personal “key” from your Mattermost account. You set this up once.

Step 1 — Create your Mattermost token ~2 min

This “Personal Access Token” (PAT) is what lets Claude act as you in Mattermost.

Open Mattermost (https://mm.kilde.sg) in a browser.
Click your profile avatar → Profile → Security → Personal Access Tokens.
Click Create Token, name it (e.g. Claude), and copy the token — you’ll paste it next.

Don’t see “Personal Access Tokens”? An admin needs to enable it (System Console → Integrations → Integration Management → Enable Personal Access Tokens). Ask in #product-and-tech.

Treat the token like a password — don’t paste it into chats, emails, or screenshots.

Step 2 — Enroll your token one-time ~2 min

You paste it once on a Kilde-internal page. It never goes through Claude or any chat.

Open https://mcp-backend.kilde.sg/enroll in a browser.
Sign in with your @kilde.sg Google account (if prompted).
Paste your Mattermost token and click Save token.
You should see ✓ Enrolled … (@your-username) — that confirms it’s valid and saved.

You only do this once. If you ever create a new token, just come back and enroll the new one.

Your token is stored encrypted on the Kilde side (Google Secret Manager) and is only used to act as you when you ask Claude to do something in Mattermost.

Step 3 — Try it ~1 min

Wait ~30 seconds after enrolling, open a new chat, make sure Kilde Internal MCP is on, and try:

“List my Mattermost channels in product-and-tech.”
“Show me the last 10 messages in #general in product-and-tech.”
“What’s been discussed in #incidents today?”
“Post to #test in product-and-tech: Hello from Claude 👋”

When you post, Claude asks you to approve first, and the message appears in Mattermost authored by you.

What you can do

You ask…	What happens
“list my channels in team”	Lists channels you belong to in that team
“read / show messages in #channel”	Returns recent messages (up to 100, default 20)
“post / send message to #channel”	Posts as you (asks for approval first)

Teams & channels

Name one of these teams (or a raw team ID) in your prompt:

Team key	Team
`deal-origination`	Deal Origination
`finance-and-ops`	Finance & Ops
`kilde-internal-chateau`	Kilde Internal Chateau
`marketing-and-sales`	Marketing & Sales
`optimworks`	Optimworks
`product-and-tech`	Product & Tech
`safebay`	Safebay

For channels, use the channel’s handle/slug (lowercase, dashes), e.g. general, product-updates — not the display name “General”.

Good to know

You see what you’d see in Mattermost. The tools use your token, so they only reach channels you’re a member of.
Posts are really from you. No shared bot — messages are under your name with a normal audit trail.
Access changes apply instantly. Join or leave a channel and it’s reflected on your next request — nothing to re-enroll.

🦆 MotherDuck — data warehouse read-only ~1 min

Ask questions about Kilde’s data in plain English — Claude writes and runs the query for you.

In one line: once the connector is on, just ask about the warehouse — Claude runs read-only queries. Unlike Mattermost, there’s no token to set up: data access is handled on the Kilde side.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Claude figures out the query and runs it. Try:

“List my MotherDuck databases.”
“What tables are in kilde_analytics?”
“In MotherDuck, how many rows are in kilde_finance.main.invoices?”
“In MotherDuck, how many deals closed last month?”

No approval prompt is needed — every MotherDuck tool is read-only, so it can look but never change anything.

What you can do

You ask…	What happens
“list / show the databases”	Lists all databases available
“what tables are in db”	Lists tables, optionally filtered to a database/schema
“query / count / show me data”	Runs a read-only query (up to 1000 rows, default 100)

Databases you can query

The default is kilde-prod-hashed — if you don’t name a database, queries run against it.

Database	Notes
`kilde-prod-hashed`	Default — production data with sensitive fields hashed
`kilde-prod`	Production
`kilde-nonprod`	Non-production
`kilde_analytics`	Analytics
`kilde_finance`	Finance
`kilde_ga4`	Google Analytics 4
`kilde_ga_control_source`	GA control source
`kilde_ga_fund`	GA fund

Good to know

Read-only, always. Only “look” operations are allowed. Anything that would change data is rejected — you can’t accidentally break anything.
Shared access. Everyone uses the same read-only connection, so what you see isn’t scoped to you individually — it’s the same view for all teammates.
Big results are capped. Queries return up to 1000 rows (100 by default). Ask for a count or add a filter to narrow things down.
Hyphenated names need quotes in SQL. If you paste your own SQL, wrap hyphenated database names in double quotes, e.g. "kilde-prod-hashed".main.my_table. Claude usually handles this for you.

🐙 GitHub — our code & pull requests read-only ~1 min

Browse and search Kilde’s code, and check pull requests and recent changes — all by just asking.

In one line: like MotherDuck, GitHub works as soon as the connector is on — no token to set up. Claude can read code, search it, and inspect PRs and commits, but it can never push, merge, or change anything.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Try:

“List the Kilde repos.”
“Show me src/index.ts from kilde-frontend.”
“Search the kilde repo for InvestorLoginController.”
“What pull requests are open on kilde-frontend?”
“Show me the last 10 commits on kilde-ai.”

No approval prompt is needed — every GitHub tool is read-only.

What you can do

You ask…	What happens
“list the repos”	Lists the repos this connector can see, with each one’s default branch and description
“show me file from repo”	Reads a file’s contents (or lists a folder). You can pin a branch, tag, or commit. Large files are trimmed.
“search repo for term”	Searches code across the repos and returns the matching file paths
“what PRs are open / closed on repo”	Lists pull requests (open by default), most recently updated first
“recent commits on repo”	Lists recent commits on a branch

Repositories you can query

Repo	What it is
`kilde`	The main backend repo
`kilde-frontend`	The frontend / web app
`kilde-ai`	AI / data tooling

Ask about a repo by its name above. If you ask about a different repo, GitHub will simply return “not found” — the connector is locked to these three.

Good to know

Read-only, always. Claude can read and search; it cannot push code, merge, or open/close anything.
Shared access. Everyone uses the same read-only connection, so the view is the same for all teammates — it isn’t tied to your personal GitHub account.
Three repos only. Access is limited to kilde, kilde-frontend, and kilde-ai.
Big files are trimmed. Very large files are cut off — ask for a specific path or a narrower section if you need more.

🤝 Pipedrive — the CRM needs your token read-only ~3 min setup

Look up deals, contacts, organizations, and activities from Claude — as yourself, in plain English.

In one line: like Mattermost, you enroll your own Pipedrive token once. From then on Claude sees exactly the deals and contacts your Pipedrive account can see — and it can only read, never change anything.

Why the extra step?

Pipedrive acts as you, so it respects your CRM visibility — you only see what you’d see when logged in to Pipedrive yourself. There’s no shared all-seeing key. To do that, Claude needs a personal API token from your Pipedrive account. You set this up once.

Step 1 — Get your Pipedrive API token ~1 min

Open Pipedrive in a browser.
Go to Settings → Personal preferences → API.
Copy your personal API token — you’ll paste it next.

Treat the token like a password — don’t paste it into chats, emails, or screenshots.

Step 2 — Enroll your token one-time ~1 min

This is the same enroll page as Mattermost — it just has a separate Pipedrive box.

Open https://mcp-backend.kilde.sg/enroll in a browser.
Sign in with your @kilde.sg Google account (if prompted).
In the Pipedrive box, paste your token and click Save Pipedrive token.
You should see ✓ Enrolled … (Your Name (you@kilde.sg)) — that confirms it’s valid and saved.

You only do this once. If you ever regenerate your token, come back and enroll the new one.

Your token is stored encrypted on the Kilde side (Google Secret Manager) and is only used to act as you when you ask Claude something about Pipedrive. It never passes through Claude or any chat.

Step 3 — Try it ~1 min

Wait ~30 seconds after enrolling, open a new chat, make sure Kilde Internal MCP is on, and try:

“List my open Pipedrive deals.”
“Search Pipedrive deals for Skybridge.”
“Find the Pipedrive contact for jane@acme.com.”
“What Pipedrive activities are still pending?”

No approval prompt is needed — every Pipedrive tool is read-only.

What you can do

You ask…	What happens
“list my deals” (open / won / lost)	Lists deals you can see, newest first; can filter by status or stage
“get / show deal 123”	Full details of one deal by its ID
“search deals for term”	Full-text search across deals
“find the contact / person name or email”	Searches people — name, email, phone, linked org
“search organizations for company”	Searches companies
“what activities are pending / done”	Lists activities (calls, meetings, tasks)
“list the pipeline stages”	Shows pipelines + stages, so a deal’s stage reads as a name

Good to know

You see what you’d see in Pipedrive. The tools use your token, so visibility matches your own CRM access.
Read-only, always. Claude can look up deals and contacts but can never create, edit, or delete anything in the CRM.
Access changes apply instantly. Nothing to re-enroll when your CRM permissions change.

🧪 VWO — A/B experiments read-only ~1 min

Look up Kilde’s VWO account and its experiments (tests) from Claude — just ask.

In one line: like MotherDuck and GitHub, VWO works as soon as the connector is on — no token to set up. Claude can read the account info and campaign configuration, but can’t change anything.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Try:

“Show me the VWO account info.”
“What VWO experiments / campaigns exist?”
“Get the config for VWO campaign 242.”

No approval prompt is needed — every VWO tool is read-only.

What you can do

You ask…	What happens
“VWO account info”	Account summary — name, timezone, company, industry
“list VWO campaigns / experiments”	Lists campaigns newest-first — status, type, created date/author
“get / show VWO campaign id”	Full config of one campaign — status, traffic split, variations, goals

Good to know

Configuration, not results (yet). These tools show campaign setup. Conversion/uplift reporting isn’t available through the connector yet — check the VWO dashboard for results.
Read-only, always. Claude can look but never edit or start/stop a campaign.
VWO rate-limits. A rapid burst of questions can return a “too many requests” message — wait a few seconds and ask again.
Shared access. Everyone uses the same read-only connection; the view is the same for all teammates.

📈 Trading Economics — global macro data read-only ~1 min

Ask about economic indicators, forecasts, historical series, the economic calendar, and market quotes for countries around the world — just ask.

In one line: like MotherDuck, GitHub, and VWO, Trading Economics works as soon as the connector is on — no token to set up. Claude can pull macro indicators (GDP, inflation, rates, unemployment…), forecasts, historical data, the economic calendar, and market quotes. It only reads — nothing to change.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Try:

“What’s Singapore’s current inflation rate in Trading Economics?”
“Show me the GDP growth forecast for the United States.”
“Give me the historical unemployment rate for Germany.”
“What’s on the economic calendar this week?”
“Show me the key indicators for Indonesia.”

No approval prompt is needed — every Trading Economics tool is read-only.

What you can do

You ask…	What happens
“search for indicator / country”	Finds matching indicators and countries in the Trading Economics catalog
“key indicators for country”	Snapshot of a country’s main macro indicators (GDP, inflation, rates, etc.)
“current value of indicator”	Latest value for a specific indicator
“forecast for indicator / country”	Trading Economics’ projected values
“historical indicator for country”	Time series of past values
“economic calendar”	Upcoming and recent releases (with dates, actual vs. forecast)
“market quotes for stocks / bonds / currencies / commodities”	Current market snapshots by asset class

Good to know

Read-only, always. Claude can look up data but can never change anything.
Shared access. Everyone uses the same read-only connection; the view is the same for all teammates.
Coverage depends on the subscription. Some countries or indicators may be limited by Kilde’s Trading Economics plan — if something’s missing, that’s usually why.
Name the country and indicator clearly. Results are best when you say both, e.g. “inflation rate for Vietnam” rather than just “inflation”.

🐛 Sentry — error monitoring read-only ~1 min

Look up Kilde’s Sentry projects, errors, and stack traces from Claude — just ask.

In one line: like MotherDuck, GitHub, and VWO, Sentry works as soon as the connector is on — no token to set up. Claude can list projects, browse issues (errors), and pull the latest stack trace for one — but can’t change or resolve anything.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Try:

“What Sentry projects do we have?”
“List the unresolved Sentry issues in the last 24 hours.”
“Show me the top Sentry errors in the frontend project.”
“Get the latest event for Sentry issue 1234567890.”

No approval prompt is needed — every Sentry tool is read-only.

What you can do

You ask…	What happens
“list the Sentry projects”	Lists projects in the Kilde org (slug, name, platform)
“list / show Sentry issues”	Lists issues (errors), newest first — defaults to unresolved; can scope to a project or filter with Sentry search terms
“get / show Sentry issue id”	Full detail of one issue — status, event counts, first/last seen
“latest event for issue id”	The most recent occurrence — full stack trace, tags, and context

Good to know

Read-only, always. Claude can read errors but can never resolve, assign, mute, or delete anything.
Shared access. Everyone uses the same read-only connection; the view is the same for all teammates.
Use the project slug. Ask “list the Sentry projects” first to get the exact slug the other questions expect.
Sentry rate-limits. A rapid burst can return a “too many requests” message — wait a few seconds and ask again.

✅ Xray — test management read-only ~1 min

Look up Xray tests, test executions, and test plans (the QA data in Jira) from Claude — just ask.

In one line: like MotherDuck and GitHub, Xray works as soon as the connector is on — no token to set up. Claude can read your Jira/Xray tests, executions, and plans, and summarise pass/fail rates — but can’t create or edit anything.

How to use it

Open a chat with Kilde Internal MCP on, and ask in plain English. Xray lives in Jira, so results are best when you name a project (or give a Jira JQL filter). Try:

“List the Xray tests in project ABC.”
“Show recent Xray test executions for ABC and their pass rate.”
“What Xray test plans exist in project ABC?”

No approval prompt is needed — every Xray tool is read-only.

What you can do

You ask…	What happens
“list Xray tests in project”	Lists tests matching a project/JQL — key, type, summary, status
“show Xray test executions for project”	Lists executions with a pass/fail summary of each one’s runs
“list Xray test plans in project”	Lists test plans — key, summary, and how many tests each holds

Good to know

Read-only, always. Claude can read tests and results but can never create, edit, or run anything.
Name a project. Add a project (e.g. ABC) or a Jira JQL filter — asking with no filter scans the whole Jira instance and is slow.
Scoped to one Jira account. The connector reads through a shared service account, so it sees the Jira projects that account has access to — not your personal Jira view.
Xray rate-limits. A rapid burst can return an error — wait a few seconds and ask again.

Troubleshooting

Any tool

Problem	Fix
“Kilde Internal MCP” isn’t in my Connectors list	The org Owner hasn’t rolled it out to you yet — ask in #product-and-tech.
The login window won’t let me in	You must sign in with your `@kilde.sg` Google account (not a personal Google account).
Every tool returns a generic error	Your connector session may have expired — Disconnect and Connect the connector again in Settings → Connectors.
Claude doesn’t seem to use the tool	Make sure Kilde Internal MCP is toggled on via the + menu in that chat, and say which system you mean (“in Mattermost…”, “in MotherDuck…”, “in GitHub…”, “in Pipedrive…”, “in VWO…”, “in Trading Economics…”, “in Sentry…”, “in Xray…”).

Mattermost

Problem	Fix
“No Mattermost token enrolled for you”	You haven’t enrolled, or it’s been <30s since you did. Enroll at `https://mcp-backend.kilde.sg/enroll` and retry.
Enroll says “Token validation failed”	The token is wrong/expired or copied incompletely. Create a fresh token and try again.
“Unknown team …”	Use a team key from the table above (or a raw team ID).
A channel isn’t found / returns nothing	Check you’re a member, and that you used the channel handle (`product-updates`), not the display name.
A 401/403 error from Mattermost	Your token was revoked or lacks permission — create and enroll a new token.

MotherDuck

Problem	Fix
Claude says the query must be read-only	You asked for something that changes data. Only reads are allowed — rephrase as a question.
“No rows” / empty result	The query ran but matched nothing. Check the database/table name and your filters.
Parse error near a hyphen in a database name	Ask Claude to “quote the database name”: `"kilde-prod-hashed".schema.table`.
“Could not connect to MotherDuck”	Usually a temporary hiccup — wait a moment and retry. If it persists, flag it in #product-and-tech.

GitHub

Problem	Fix
“GitHub is not configured”	The shared token isn’t mounted on the server — an admin needs to set it up. Flag it in #product-and-tech.
A repo returns “404 / not found”	It’s outside the three allowed repos (`kilde`, `kilde-frontend`, `kilde-ai`), or the file/path doesn’t exist on that branch.
Code search returns nothing	Try different terms, or scope to one repo. GitHub code search can miss very new or very large files.
A file looks cut off	Large files are trimmed — ask for a narrower path or a specific section.

Pipedrive

Problem	Fix
“No Pipedrive token enrolled for you”	You haven’t enrolled, or it’s been <30s since you did. Enroll the Pipedrive box at `https://mcp-backend.kilde.sg/enroll` and retry.
Enroll says “Token validation failed”	The token is wrong/incomplete. Re-copy it from Pipedrive → Settings → Personal preferences → API and try again.
A deal/contact you expect isn’t there	The tools only see what your Pipedrive account can — check your visibility/permissions in Pipedrive.
A 401/403 error from Pipedrive	Your token was regenerated or lacks access — get a fresh token and enroll it again.

VWO

Problem	Fix
“Too many requests” / 429	VWO rate-limits bursts. Wait a few seconds and ask again — no setup needed.
I want conversion / uplift numbers	Not available through the connector yet — it covers account & campaign config only. Use the VWO dashboard for results.
“VWO is not configured”	The shared token isn’t mounted on the server — an admin needs to set it up. Flag it in #product-and-tech.

Trading Economics

Problem	Fix
A country or indicator returns nothing	It may be outside Kilde’s Trading Economics plan, or the name didn’t match — try the exact country name and a clearer indicator, or “search Trading Economics for …” first.
“Too many requests” / 429	Trading Economics rate-limits bursts. Wait a few seconds and ask again.
“Trading Economics is not configured”	The shared key isn’t mounted on the server — an admin needs to set it up. Flag it in #product-and-tech.

Sentry

Problem	Fix
A project or issue returns nothing	Check the project slug (ask “list the Sentry projects” first), and widen the time window or search query.
“Too many requests” / 429	Sentry rate-limits bursts. Wait a few seconds and ask again.
“Sentry is not configured”	The shared token isn’t mounted on the server — an admin needs to set it up. Flag it in #product-and-tech.

Xray

Problem	Fix
A query is slow or returns too much	Add a project or Jira JQL filter (e.g. `project = "ABC"`) — an unfiltered query scans the whole instance.
A project’s tests aren’t showing	The connector reads through a shared Jira service account — that project may be outside its access. Flag it in #product-and-tech.
“Too many requests” / 429	Xray rate-limits bursts. Wait a few seconds and ask again.
“Xray is not configured”	The shared API key isn’t mounted on the server — an admin needs to set it up. Flag it in #product-and-tech.

Still stuck? Ask in #product-and-tech.

Admins: for the server-side setup behind this guide (per-user Mattermost and Pipedrive token enrollment, the shared MotherDuck / GitHub / VWO / Trading Economics / Sentry / Xray read-only tokens in Secret Manager, and deploy), see README.md and MATTERMOST.md in devops/MCP/.