LaunchDarkly MCP server: 20 tools, two routes

What an AI agent can change through LaunchDarkly, which of the two servers you should be running, and how many of its tools alter what users see.

Last verified 1 September 2026 · from LaunchDarkly's repository, its documentation and the server's own tools/list response

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the LaunchDarkly MCP server?

It is LaunchDarkly's server for the Model Context Protocol. LaunchDarkly controls feature flags, so the agent gains tools for reading and changing which features are on for whom.

Half the surface is AI configs, LaunchDarkly's newer product for managing model and prompt selection behind a flag. Eleven of the twenty tools name ai-config directly.

There are two servers. The hosted one is the default recommendation. The local one covered here is for EU and Federal instances, and LaunchDarkly says plainly that it lags the hosted server.

LaunchDarkly covers release in full. Code references reach part of the way into code review. AI configs reach part of the way into agent operation. Nothing reaches production observability or cost.

For an agent, that means LaunchDarkly answers what is switched on for whom, and can change it. An agent connected here can tell you which flags are live in production and where they are referenced in code. It cannot tell you whether turning a flag on made anything worse. Run this connector beside an observability connector, and give it tighter permissions than most.

What can an AI agent do with LaunchDarkly?

Ten of the twenty tools write, and on a feature-flag product a write is a production change with no deploy in between. This is the connector on which an allowlist matters most.

What connecting costs before the first question

A context window is the amount of text a model can hold at once, and connecting uses some of it before the agent asks anything. Every tool the server registers sends its name, description and input schema to the model when the session starts. The tool list is charged whether or not a single tool is called.

LaunchDarkly's cost sits mostly in its schemas rather than its prose: descriptions average 150 characters, while input schemas average 713, because flag and targeting objects are structurally large.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 21 August 2026 from tools/list on the published @launchdarkly/mcp-server package, run locally with a placeholder key.

Registration is static, so no LaunchDarkly account was involved. Figures cover name, description and input schema. The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 75%, wide because the schemas carry most of the payload. Reasoning tokens are output, spent per request. The tool list is input, charged once at connection.

The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer, so the definitions were sent to the API's count_tokens endpoint and a no-tools baseline subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end, Opus 5 and Sonnet 5 at the high end. The gap between them is about a third.

Can LaunchDarkly watch your own AI agents?

LaunchDarkly sits on the control side of this stage rather than the observing side: it decides what an agent runs on, and records nothing about what happened. Read from LaunchDarkly’s MCP documentation on 21 August 2026.

LaunchDarkly MCP server limits: where answers come back incomplete

LaunchDarkly's first constraint is specific to a connector that can change live flags.

How to configure LaunchDarkly MCP for agents

1. 1

   Use the hosted server unless you cannot

   LaunchDarkly states the hosted server is more feature complete and receives more frequent updates. This package exists for EU and Federal instances.

2. 2

   Allowlist read tools

   Ten of the twenty write, and on a feature-flag product a write reaches users without a deploy.

3. 3

   Scope the access token

   LD_ACCESS_TOKEN carries whatever the token was granted, across every project and environment it can reach.

4. 4

   Set up code references if you want them

   get-code-references returns nothing without LaunchDarkly's repository scanner, and an unscanned repository looks the same as an unused flag.

Paste this into your agent’s instructions

A flag answer is only meaningful with its environment, and an empty code-references result is easy to misread as a dead flag. The block makes the agent state both.

When you answer using LaunchDarkly MCP tools, state:
- The project and environment each flag status came from.
  A flag can be on in one environment and off in another.
- Whether code references are configured. An empty result may
  mean the repository was never scanned, not that the flag is
  unused.
- Whether you changed anything. Flag and targeting updates
  reach users without a deploy.
Never state that data does not exist. State that none was
returned under the constraints above.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Yes, and LaunchDarkly's own guidance points that way. The hosted server is the recommended surface, and the REST API covers the same operations with the same key.

The connector's advantage is that flag and targeting shapes are complicated, and the schemas describe them, so an agent can construct a valid targeting change without being handed the API reference.

That advantage cuts both ways here. A tool that makes a targeting change easy to write also makes it easy to apply by mistake.

Use the hosted server unless your instance is EU or Federal. Keep write tools off the allowlist either way.

LaunchDarkly MCP server setup

The hosted server at mcp.launchdarkly.com/mcp/launchdarkly is the recommended route and needs no install. The local package runs with npx and takes an access token, either as --api-key or in LD_ACCESS_TOKEN.

The local package is intended for EU and Federal instances, which do not yet support the hosted server.

claude mcp add --transport http launchdarkly https://mcp.launchdarkly.com/mcp/launchdarkly

claude mcp add-json "launchdarkly" '{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","<your-token>"]}'

{"mcpServers":{"launchdarkly":{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","$LD_ACCESS_TOKEN"],
  "env":{"LD_ACCESS_TOKEN":"<your-token>"}}}}

[mcp_servers.launchdarkly]
command = "npx"
args = ["-y", "@launchdarkly/mcp-server", "start", "--api-key", "$LD_ACCESS_TOKEN"]
env = { LD_ACCESS_TOKEN = "<your-token>" }

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

{"mcpServers":{"launchdarkly":{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","$LD_ACCESS_TOKEN"],
  "env":{"LD_ACCESS_TOKEN":"<your-token>"}}}}

{"mcpServers":{"launchdarkly":{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","$LD_ACCESS_TOKEN"],
  "env":{"LD_ACCESS_TOKEN":"<your-token>"}}}}

{"servers":{"launchdarkly":{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","$LD_ACCESS_TOKEN"],
  "env":{"LD_ACCESS_TOKEN":"<your-token>"}}}}

{"mcpServers":{"launchdarkly":{"command":"npx",
  "args":["-y","@launchdarkly/mcp-server","start","--api-key","$LD_ACCESS_TOKEN"],
  "env":{"LD_ACCESS_TOKEN":"<your-token>"}}}}

LD_ACCESS_TOKEN=<your-token> \
npx -y @launchdarkly/mcp-server start \
  --server-url https://app.eu.launchdarkly.com

Read-only access, permissions and security

• Half the tools change production

  Ten of the twenty create, update or delete flags and AI configs, and a flag change takes effect without a deploy.

• There is no read-only mode

  The server offers no flag equivalent to a read-only switch, so the client's allowlist is the only gate.

• AI config targeting selects the model

  update-ai-config-targeting changes which model or prompt an agent runs on in production, which changes how the agent behaves, not just how it is configured.

• The token carries broad reach

  An access token spans the projects and environments it was granted, so scope it to the smallest set that answers your questions.

Troubleshooting

A flag looks off but users still see the feature

Flag status is per environment. Use get-flag-status-across-environments and name the environment rather than assuming production.

No code references for a flag you know is used

get-code-references reads data from LaunchDarkly's repository scanner. Without that integration the result is empty, which looks identical to an unused flag.

401 on connect

The access token is missing or scoped to another instance. For EU or Federal, set --server-url to the correct endpoint.

A capability exists in LaunchDarkly but not in the server

LaunchDarkly states the hosted server is more feature complete than this local one. Check the hosted server before concluding the product lacks it.

A flag changed that you did not change

Write tools have no server-side gate. Remove every create-, update- and delete- tool from your client's allowlist, and check get-audit-log-entries for what happened.

LaunchDarkly MCP server: Reference

What engineers report

The local server has a small public footprint, so the useful signals are LaunchDarkly's own guidance and the read-write split.

“Using an API key makes it difficult to use your MCP in an IDE because we can't push the key to our source code. ... Many MCPs in the IDE now just send you to authenticate, rather than working with a static API key.”

“The CLI defaults --api-key to an empty string, and the nullish coalescing operator in the security module doesn't treat it as nullish, so the env var is never read.”

Repository facts from the GitHub API, retrieved 21 August 2026. Tool counts and the read-write split come from the server's own tools/list response.

Should you connect an agent to LaunchDarkly?

Yes, though it is no longer actively maintained. Best for EU and Federal instances, read tools only, with flag changes left to a reviewed process.

Strongest fit

Teams on EU or Federal LaunchDarkly instances who want an agent auditing which flags are live and where they are referenced.

Main advantage

Schemas that describe flag and targeting structures well enough for an agent to work with them correctly.

Main weakness

LaunchDarkly says this server lags the hosted one, so a missing capability may be a packaging gap rather than a product gap.

Operational risk

Half the tools change live flags with no deploy in between, and there is no read-only mode.

Frequently asked questions