AI agents for LaunchDarkly
A Monmo agent works in LaunchDarkly the way a release manager who reads the audit log would. It finds flags that should have been removed months ago, plans percentage rollouts step by step, tells you who changed a flag and when, and explains experiment results in plain words. It uses the access token you connect, and asks before it toggles a flag, edits targeting or archives anything.
Updated 2026-11-13 · Free to start. Paid plans from $15 a month.
What agents do in LaunchDarkly
Find and remove stale flags
It lists flags marked launched or inactive for over 30 days, checks their code references in GitHub and drafts a pull request that removes each one from the code. Merging and archiving both ask first.
AskFind flags that have served one variation in production for 30 days and open clean-up pull requests for the top five.
Say who changed a flag and why
Ask why the new checkout appeared for everyone yesterday and it reads the audit log for that flag, finds the change, who made it, the comment left with it and the environment it was in.
AskWho changed the new-checkout flag in production yesterday, what exactly changed, and was there a comment?
Plan a percentage rollout
It writes a rollout plan, such as 5, 25, 50 and then 100 percent, with the Datadog error rate to check between steps. Each step is its own card, so you move forward only when the numbers look right.
AskPlan a rollout of the search-v2 flag to 100 percent over a week, checking error rates between steps.
Compare targeting across environments
It lines up a flag's rules, individual targets, segments and default variations in staging and production, and points out differences that would surprise someone testing in staging.
AskCompare the pricing-page flag between staging and production and tell me what behaves differently.
Flip the kill switch quickly
During an incident, ask it to turn off a flag and it prepares the change with the environment and the effect spelled out on one card, so you can approve it in seconds rather than hunt through the UI.
AskThe new recommendations widget is causing errors. Prepare turning its flag off in production right now.
Explain an experiment's results
For a finished experiment it reads the metric results for each variation, notes how certain the results are and how large the sample was, and writes a short summary. Picking the winner is your decision.
AskSummarise the onboarding-copy experiment in plain words. Is the difference real or too early to call?
Keep flags well described
It finds flags with no description, maintainer or tags, drafts a one-line purpose for each from the code and the Jira issue that introduced it, and proposes the edits as one batch.
AskList flags with no description or maintainer and draft both from the code and the original Jira issue.
How it connects
- 1You create an access token in LaunchDarkly and connect it once. The Reader role suits reports and audits; a Writer or a custom role limited to chosen projects and environments is needed for changes.
- 2LaunchDarkly belongs with your engineering agent. A product or data agent may read experiment results through the same connection; nobody else needs it.
- 3Reading flags, the audit log and experiment results needs no approval. Toggling a flag, editing targeting or rollout percentages, changing a segment and archiving a flag ask first. If you require approvals for production, the agent's request still goes through them.
Questions people ask
Can an AI agent change feature flags in LaunchDarkly?
Yes, one approved change at a time. A Monmo agent prepares the toggle, the new targeting rule or the next rollout step and shows the flag, environment and effect on a card. Once you allow it, the change is made through the LaunchDarkly connection, or in its own browser when the connection does not cover something.
How can AI help clean up old feature flags?
The slow part of flag clean-up is finding the code. The AI agent lists flags that have served one variation for weeks, finds their references in GitHub, drafts the pull request that removes each one and archives the flag after the merge. Merging and archiving both wait for your approval.
Is it safe to give an AI agent a LaunchDarkly token?
Give it the least it needs. A Reader token covers audits, reports and experiment summaries. If you want it to make changes, use a custom role limited to the projects and environments involved. Every change it makes still waits for your approval, and you can revoke the token at any time.
Can it run a gradual rollout for me?
It can plan the steps and carry out each one when you allow it. Between steps it checks the measures you choose, such as error rate in Datadog or conversion in Amplitude, and tells you whether they held. If something looks wrong, it recommends pausing or rolling back instead of the next step.