launchdarkly / ai-tooling

agent-graphs

Sets up a hand-off chain between several AI assistants in LaunchDarkly — one takes the question first, then passes it on to whichever one should handle it next.

aiconfig-agent-graphs

aiconfig-ai-metrics

aiconfig-create

aiconfig-custom-metrics

aiconfig-migrate

aiconfig-online-evals

aiconfig-projects

aiconfig-snippets

aiconfig-targeting

aiconfig-tools

aiconfig-update

aiconfig-variations

apply

The hands-on step of adding LaunchDarkly to an app — the library and the start-up code go in, and your secret key is kept wherever you tell it to go instead of in the code.

built-in-metrics

Hooks up an app's existing AI calls so each one reports on itself — how long it took, how much text went in and out, whether it succeeded — and the figures land on LaunchDarkly's monitoring page.

custom-metrics

Lets you measure whether an AI feature is doing its job in your own terms, rather than only how fast it answered or how many words it used.

detect

The first step of adding LaunchDarkly to a project that already exists: it reads the code to work out what it is written in, which LaunchDarkly toolkit fits, and whether someone installed one already.

launchdarkly-experiment-setup

Runs a proper A/B test inside LaunchDarkly: split the traffic between two versions, let the numbers come in, and call a winner at the end.

launchdarkly-flag-command

A quick lookup: give a flag's name and get back what that switch is doing right now — and only that, because anything beyond looking is handed to another skill.

launchdarkly-guarded-rollout

Lets a new version out slowly — one percent, then ten, then everyone — and stops it, or takes it straight back, the moment the numbers you chose to watch get worse.

mcp-configure

Hooks a coding assistant up to LaunchDarkly so it can read and change things there itself — you sign in once in the browser, with no keys to copy around.

migrate

When an app has its prompt and its model name written into the code, changing either one means shipping a release — this moves them out to LaunchDarkly, where they can be changed as it runs.

plan

The step between looking at a project and touching it: what will be installed, which files will change, and which settings get added — written out for you to approve first.

sdk-install

Gets LaunchDarkly's code into a project and running — working out what the project is, planning the change, then making it, in that order.

snippets

Text you would otherwise paste into a dozen prompts — a persona, a legal disclaimer, a set of house rules — kept in one place, so changing it once changes all of them.

launchdarkly-factory-settings

Controls Factory, LaunchDarkly's GitHub automation that can wrap new code changes in on/off switches and release them by itself, or explains, without changing anything, why it skipped a particular pull request.

onboarding

The very first setup: it puts LaunchDarkly into a project, builds one working switch you can watch flip on and off, and commits nothing until you have said yes.

launchdarkly-metric-choose

Advice, and nothing else: which numbers a release should be judged by, which of them are already being collected, and which get attached for you whether you ask or not.

launchdarkly-metric-create

Sets up a measurement — page views, button clicks, or whatever else you want counted — and, if nothing is recording it yet, puts that into the code first.

launchdarkly-metric-instrument

Nothing can be measured until the app itself says the thing happened — this adds that line to the code, in the right place, and then checks the message is arriving.

configs-create

Changing which model or prompt an AI feature uses stops being a code change: both get set up in LaunchDarkly instead, and nothing is served until you switch the new setup on.

configs-targeting

Decides who gets which version of an AI feature's setup: everyone, one user in ten, or only the enterprise customers — and pulls a change back on its own if it starts going wrong.

configs-update

Changes an AI feature that is already set up — which model it runs on, how its instructions read — and retires the ones nobody needs any more by putting them away rather than destroying them.

configs-variations

Makes a second copy of an AI setup that differs in exactly one thing — a cheaper model, a longer prompt — so that comparing the two actually tells you something.

online-evals

Has a second model read what your AI said and mark it — accurate, relevant, safe — so quality arrives as a number instead of a hunch.

projects

A LaunchDarkly project is the box everything else lives in — one team's or one service's switches and environments — and this works out how many you need and sets them up.

tools

An AI agent can only do what it has been handed — this writes down one of those abilities, such as looking something up in a database, and gives it to the agent.

launchdarkly-flag-cleanup

Once a switch has been left on for good, this takes it out of the code — keeping whichever behaviour it was actually serving — and opens a pull request with the change.

launchdarkly-flag-create

Adds a new on/off switch to a project — named and written the way that project already does it — and leaves it switched off until somebody deliberately turns it on.

launchdarkly-flag-discovery

Takes stock of every switch in a project and says which ones are still doing real work, which have been sitting untouched for months, and which can safely come out.

launchdarkly-flag-targeting

Who actually sees a feature and who does not — everybody, nobody, a quarter of your users, one named account — and copying that arrangement from staging over to production.