Claude + LaunchDarkly: What the Integration Can (and Can't) Do in 2026
LaunchDarkly ships its own official MCP server, so Claude can genuinely read flag status and manage flags in a chat — but there’s no LaunchDarkly connector in Claude’s directory, and nothing runs on a trigger. The real routes are two: LaunchDarkly’s official MCP server (hosted or run locally, works in Claude Code and Claude Desktop), and the plain LaunchDarkly REST API with an access token for anything custom. Both are excellent for “help me reason about this flag right now.” Neither reacts to a rollout going sideways on its own.
Here’s each route, when to pick it, and what to use when you want LaunchDarkly work that happens without you at the keyboard.
Route 1: LaunchDarkly’s official MCP server
This is the real one. LaunchDarkly built and maintains its own Model Context Protocol server — github.com/launchdarkly/mcp-server — and it translates natural-language requests into LaunchDarkly API calls across four product areas:
- Feature management — create and manage feature flags, read a flag’s current status and targeting, toggle it on or off per environment.
- AI Configs / AgentControl — manage the configs LaunchDarkly uses to govern AI features and model rollouts.
- Observability — query logs, traces, errors, and dashboards.
- Metrics — pull the numbers behind an experiment or rollout.
There are two ways to run it. LaunchDarkly offers a hosted MCP server you point your client at, and a local option for teams on federal or EU instances (or anyone who’d rather run it themselves). The local install is a single command:
npx -y --package @launchdarkly/mcp-server -- mcp startDrop the equivalent JSON block into Claude Desktop’s config (or add it in Claude Code), authenticate with a LaunchDarkly API access token, and Claude can answer things like:
“Is the
new-checkout-flowflag on in production, and what percentage of users is it serving?”
“Which flags did we change this week, and in which environments?”
“Create a boolean flag called
beta-dashboard, off by default, in staging only.”
The server officially supports Claude (including Claude Code), Cursor, VS Code, and Windsurf. For a developer who lives in flags all day, this collapses a lot of dashboard-clicking into a sentence.
Route 2: The REST API + an access token
If you want programmatic control — a script, a CI step, or Carly (below) — you go straight to LaunchDarkly’s REST API, which is what the MCP server calls under the hood anyway. Auth is a bearer token in an Authorization header, and LaunchDarkly gives you two kinds:
- Personal tokens — tied to your account, inheriting your permissions (and shrinking automatically if your permissions shrink). Fine for personal scripts.
- Service tokens — independent of any member, with permissions fixed at creation. This is what you use for automation, so a workflow doesn’t break when someone leaves.
Create either from the Authorization page in your LaunchDarkly settings; the token is shown exactly once. Best practice is the smallest scope that does the job — a read-only or environment-restricted custom role if the integration never touches production. This route is more work than the MCP server for interactive use, but it’s the honest foundation for anything that needs to run unattended.
The limits that actually matter
The MCP server is real and good, but Claude’s operating model still caps what any of this can do:
- No official directory connector — unlike Datadog, there’s no one-click LaunchDarkly connector in Claude’s + menu. You configure the MCP server or the API yourself.
- No triggers — nothing wakes Claude when a flag’s error rate spikes or a rollout crosses a threshold. A scheduled task can check on a timer; the spike itself can’t be what starts the check.
- Local servers are machine-bound — a local MCP server lives on your machine, so anything unattended has to run against the hosted server instead.
- claude.ai on the web can’t reach a local setup — the browser app can’t talk to an MCP server running on your desktop; that’s a Claude Code / Claude Desktop capability.
Claude with the LaunchDarkly MCP server is excellent for reasoning about flags, on demand or on a clock. What it won’t do is notice the rollout going sideways — the incident has to reach you first, and then you ask.
If you want LaunchDarkly work that runs on its own: Carly
Carly is an AI executive assistant built around triggers — event happens, workflow runs, no human kickoff — which is the shape flag operations actually want. It connects to 200+ tools, LaunchDarkly included, plus anything else with an API via your own key. Teams wire it up like this:
- Error spike after a rollout → Carly reads the flag’s status and recent metrics, writes a plain-English “should we roll this back?” brief, and emails the owner.
- Stale-flag hygiene → every Monday, a digest of flags that have been fully rolled out for 30+ days and are safe to remove.
- Change log → when a flag flips in production, Carly posts who changed it and to what, so nobody finds out from an incident.
You build these by describing them — tell Carly “email me a summary when a production flag changes” and it interviews you about the specifics before assembling the workflow. It actually sends, too, over Gmail or Outlook, not just a draft. To reach LaunchDarkly, paste a service token on the integrations page. AI agents start at $35/month, and steps in a workflow that don’t use AI run free and unlimited.
Claude vs Carly
| Claude (LaunchDarkly MCP) | Carly | |
|---|---|---|
| Read a flag’s status / targeting | Yes — via official MCP server | Yes |
| Create or toggle a flag in a session | Yes, with token permissions | Yes, as a workflow step |
| Act on a trigger (rollout error, flag change) | No — waits to be asked | Yes, trigger-based |
| Runs on a schedule (weekly stale-flag digest) | Yes — scheduled task on the hosted server | Yes |
| Runs unattended | Hosted server only; local MCP is machine-bound | Yes — cloud, 24/7 |
| Sends the summary by email | No | Yes — Gmail + Outlook |
| Pricing | Pro $20 / Max $100–$200 | AI agents from $35/mo |
Claude is the best flag-reasoning partner you’ll have in a chat; Carly is the one that watches the rollouts and emails you when something needs a decision.
Frequently Asked Questions
Does Claude work with LaunchDarkly?
Yes. LaunchDarkly maintains an official MCP server that plugs into Claude Code and Claude Desktop, letting Claude read flag status, create and manage flags, and query observability data. There’s no directory connector, so you configure the MCP server (hosted or local) or use the REST API yourself.
Is there an official LaunchDarkly MCP server?
Yes — github.com/launchdarkly/mcp-server, built by LaunchDarkly. It’s available as a hosted server and as a local npx @launchdarkly/mcp-server install, and it covers feature management, AI Configs, observability, and metrics. You authenticate with a LaunchDarkly API access token.
Can claude.ai on the web access LaunchDarkly?
Not a local MCP server — the browser app can’t reach a server running on your desktop. Use LaunchDarkly’s hosted MCP option, or the REST API, if you need access outside Claude Code / Claude Desktop.
Can Claude roll back a flag automatically when something breaks?
No. The MCP tools run when you ask or when a timer fires, so Claude can’t watch a rollout and react to it. Automatic, trigger-based flag work — read status on an error spike, email a rollback recommendation — is what Carly is for.
More: Claude connectors · Claude + Datadog · Claude + Sentry · Can Claude send emails · Claude vs Carly · Integrations
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
