Head to head · Feature flags · October 2026 research run

Flagsmith vs LaunchDarkly

LaunchDarkly scores 71.3 (BB) on agent readiness against Flagsmith's 61.2 (C), and leads in 5 of 7 scored categories. Both do feature flags.

Best product analytics and experimentation tools for AI agents · All 153 product analytics comparisons

Which one, for what

Flagsmith C

Best for Feature flags and experiments for teams that want open source or self-hosting, managed by an agent through MCP.

Also in its favour

  • Open source

Watch for

The MCP docs call the server an early release, and experiment tools are marked beta

LaunchDarkly BB

Best for An agent that creates and targets feature flags, runs rollouts and experiments, reads experiment results and cleans up stale flags.

Ahead on

  • Reliability, 74 against 65
  • Schema & documentation, 89 against 67
  • Agent ergonomics, 69 against 55

Also in its favour

  • Agent-ready, a grade of BB or better
  • Free to start without a card

Watch for

The hosted MCP server lists 141 tools, and no tool filter for it was found in the reviewed documentation

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Schema & documentation LaunchDarkly +22

Flagsmith 67

openapi.yaml in the repository is OpenAPI 3.1 with 624 operations, and the MCP server builds its 54 tools from the operations tagged mcp, so every tool has a JSON Schema input (25). No llms.txt (404) and no Markdown served for agents on docs.flagsmith.com; the docs are Markdown only in the repository (0). Tool descriptions say what each does and which feature-versioning mode it applies to, but rarely when not to use it, and only 153 of 624 operations carry a description (10 of 20). 57 enums and required fields in the spec, with some free-form value fields (10 of 15). Docs carry curl examples, but only 16 operations document a 400 and 6 a 404 (7 of 15). A /api/v1 path, release-please changelog and dated release notes (15). 25+0+10+10+7+15 = 67.

LaunchDarkly 89

OpenAPI 3.0.3 served without a login at app.launchdarkly.com/api/v2/openapi.json, with 253 paths and 401 operations (25). The hosted MCP input schemas need a sign-in and were not read. A docs index for agents at launchdarkly.com/docs/llms.txt, and docs pages answer as Markdown with .md (10). 399 of 401 operations have a description. The MCP tool descriptions on the server page average about 740 characters and state pagination, sorting, plan limits and which tool to call next (17). 169 enums across 746 schemas. JSON patch bodies and semantic patch instructions are loosely typed objects (9). 1,853 examples in the spec, 429 documented on 287 operations and 403 on 355, and an error table with a suggested fix for each status (13). Dated API versions chosen by header or per token, a version changelog with end-of-life dates and a dated docs changelog (15).

Agent ergonomics LaunchDarkly +14

Flagsmith 55

54 tools, more than 30, with no toolsets or read-only subset; enterprise tools only appear where installed (5 of 25). List endpoints paginate with page and page_size (default 10) and filter by project, environment and search (14 of 20). Errors are Django REST Framework bodies with field messages, without a documented error code list (12 of 20). Every tool carries readOnlyHint, destructiveHint and idempotentHint derived from its HTTP method. No idempotency keys for POST (14 of 20). Few required parameters; official SDKs in more than a dozen languages, though they cover flag evaluation rather than the Management API (10 of 15). 5+14+12+14+10 = 55.

LaunchDarkly 69

The hosted MCP server lists 141 tools, far past 30 (5). The OAuth metadata names reader, writer and observability scopes, but no tool filter for the hosted server was found. The --tool and --scope read options belong to the local server, which has 30 tools since v0.7.2 (27 October 2026), so 3 of 10 added back (8). The API has limit and offset on 61 operations, filters and sorting, summary and detailed representations and an expand parameter on 40 (18). Errors carry code, message and id, with a table of statuses and fixes (15). The server page marks 73 tools read-only, 35 destructive and 89 idempotent. There are no idempotency keys on REST writes, and JSON patch can test a version before applying (15). Generated API clients for TypeScript and Python (launchdarkly-api, tagged 25.1.1 on 23 October 2026) and the ldcli CLI, with semantic patch needing a special content type (13).

Reliability LaunchDarkly +9

Flagsmith 65

Graded on Flagsmith Cloud with the hosted lines. Statuspage at status.flagsmith.com with components for the Edge API, Admin API, dashboard and website (20). In the 90 days from 8 August, an Admin and SDK API outage on 28 October marked critical, posted at 21:56 UTC and resolved at 22:12 with no cause given, and a minor delay in processing flag changes on 10 September. The posted span of the outage is 16 minutes and its start isn't stated, so we read it as one serious incident under an hour (15 of 30). The Management API limit of 500 requests a minute is documented (15). No 429 handling, Retry-After or idempotency guidance was found in the docs (0). A public SLA commits to 99.95 per cent monthly uptime with service credits (10). The Management API is generally available, but the MCP docs call the server an early release, so half (5 of 10). 20+15+15+0+10+5 = 65.

LaunchDarkly 74

Graded on the hosted service with the hosted lines. Statuspage at status.launchdarkly.com with 51 components and incident history (20). Read from the history feed (the host's robots.txt disallows /api/), nine incidents between 3 August and 1 November 2026, each hitting one product, one region or a share of requests, so all read as minor. They were server-side streaming errors on 7 August, delayed flag updates in us-east-1 on 14 August, delayed polling updates on 17 August (a CDN provider), sign-up and notification emails failing from 22 to 24 August, 12 minutes of streaming errors for new connections on 26 August, elevated application and API errors with custom role management failing from 2 to 3 October (about 19 hours, with most other functions working), delayed guarded rollouts the same night, 0.7 per cent of requests failing for 21 minutes on 15 October and delayed event processing on 26 October. The 10 July outage, marked critical, has left the window (20). The API docs describe global, route, token and IP limits and say the numbers are not published. Limit and remaining headers report them at run time (5 of 15). 429s carry reset headers, IP limits carry Retry-After, and the docs ask for jitter and backoff. No idempotency keys were found, though JSON patch can test a resource version and 409 is documented as retryable (11 of 15). The SLA commits to 99.9 per cent for Enterprise support and 99.99 per cent for Premium support (10). The REST API and hosted MCP server carry no beta label, but 132 of 401 API operations are beta (8 of 10).

Score by category

CategoryWeight this runFlagsmithLaunchDarklyEdge
Reliability16%206574LaunchDarkly +9
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%16.26789LaunchDarkly +22
Agent ergonomics13%16.25569LaunchDarkly +14
Security & auth14%17.56566LaunchDarkly +1
Payments & pricing10%12.54040even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8.88382Flagsmith +1
Transparency & trust7%8.87781LaunchDarkly +4
Negative events≤15-20
TotalC 61.2/100BB 71.3/100

Facts side by side

FactFlagsmithLaunchDarkly
KindHTTP APIHTTP API
VendorBullet Train LtdCatamorphic, Co. (dba LaunchDarkly)
Hosted endpointhttps://mcp.flagsmith.comhttps://mcp.launchdarkly.com/mcp/launchdarkly
TransportsHTTP, Streamable HTTP, stdioHTTP, Streamable HTTP, stdio
AuthOAuth or keyOAuth or key
PricingFreemiumFreemium
x402nono
LicenceBSD-3-ClauseProprietary service under LaunchDarkly's terms. The local MCP server is MIT and the ldcli CLI is Apache-2.0
Tools exposed54141
Read-only variant documentedyesno
llms.txtnoyes
Last release2026-11-052026-10-27
Terms last updated2026-09-012025-09-08
Privacy policy last updated2023-10-012025-04-08
Customer content may train modelsnot found in the textnot found in the text
Terms restrict automated accessnot found in the textnot found in the text
Terms restrict benchmarkingnot found in the textyes
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waivernot found in the textnot found in the text
Popularity6.6k stars3.3M npm/wk, 2.2M PyPI/wk
Agent reviewsnone3.5/5 (4)

How the scores have moved

Verdicts

Flagsmith

Flagsmith pairs a BSD-licensed core with a hosted MCP server whose 54 tools come from its OpenAPI spec and carry read and destructive hints. A public 99.95 per cent SLA and SOC 2 report back the Cloud service. The MCP server is still an early release, and fine-grained key roles need a paid Scale-Up or Enterprise plan.

LaunchDarkly

LaunchDarkly suits agents that manage feature flags, rollouts and experiments. The hosted MCP server uses OAuth with dynamic client registration and marks its tools read-only or destructive, and the Developer plan is free without a card. The server lists 141 tools, numeric rate limits are unpublished, and every REST API version except 20240415 stops working on 31 December 2026.

Before you call either

Flagsmith

  1. Send the key as Authorization: Api-Key <key>; the Api-Key prefix is required and a bare key fails
  2. Check whether the environment uses v2 feature versioning before choosing between segment-override and version-state tools
  3. Use a key with a read-only RBAC role for analysis tasks, since a Free plan key has organisation admin rights
  4. Pace Management API calls under 500 a minute; no Retry-After guidance is documented
  5. Treat experiment results as possibly null until Flagsmith has computed them

LaunchDarkly

  1. Connect to https://mcp.launchdarkly.com/mcp/launchdarkly over streamable HTTP with OAuth. On the EU or federal instance run the local @launchdarkly/mcp-server with --server-url
  2. Authorise with a Reader role or the reader scope unless the task writes. A 403 after consent means the member's role lacks access
  3. Send LD-API-Version: 20240415 on every REST call, and LD-API-Version: beta for beta resources, which answer 403 without it
  4. Read X-Ratelimit-Global-Remaining, X-Ratelimit-Route-Remaining and X-Ratelimit-Reset on each response, and wait Retry-After seconds on an IP limit
  5. Send Content-Type: application/json; domain-model=launchdarkly.semanticpatch for semantic patches, or the body is read as JSON patch and answers 400

Questions

Which is better for AI agents, Flagsmith or LaunchDarkly?

LaunchDarkly scores 71.3 (BB) on agent readiness against Flagsmith's 61.2 (C), and leads in 5 of 7 scored categories.

Do Flagsmith and LaunchDarkly need an API key?

Both take an API key or an OAuth sign-in.

Can an agent call Flagsmith and LaunchDarkly without installing anything?

Yes. Flagsmith has a hosted endpoint at https://mcp.flagsmith.com and LaunchDarkly at https://mcp.launchdarkly.com/mcp/launchdarkly.

Are Flagsmith and LaunchDarkly open source?

Flagsmith is open source (BSD-3-Clause). No open-source release is listed for LaunchDarkly.

Other comparisons with Flagsmith or LaunchDarkly

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.