Experimentation market report
PostHog is the broadest documented experimentation MCP for one AI connection across analytics, feature flags, and experiments. Statsig is the clearest experiment-first choice for project data and administration. LaunchDarkly is strongest when experimentation sits inside a feature-flag and observability workflow. This is an action-level comparison, not a claim that one statistics engine is universally better.
Pick the MCP by the action you need. PostHog documents read and write access across products, including flags and experiment management. Statsig documents authenticated access to experiments, metrics, gates, and project objects, with read-only users separated from write-capable keys. LaunchDarkly documents a hosted server for flags, AgentControl configs, observability, and metrics, while its experiment guide still describes setup as a sequence around flags, metrics, audiences, and iterations.
These ratings measure the official MCP surface and workflow fit documented on September 21, 2026. They are not a statistical-quality score, security score, or benchmark of live accounts. An API may support actions that the MCP does not expose.
The table compares the documented surface, not every REST, GraphQL, SDK, or console feature. Confirm the exact tool, role, region, plan, and client approval behavior before production use.
| Platform | What the official MCP can do today | Action boundary | Rating |
|---|---|---|---|
| PostHog | Hosted account MCP across analytics, flags, experiments, SQL, and more. Broadest documented product loop. | Read and write are documented, including flag and experiment setup. Confirm exact tool and role before live changes. | 9.5/10 |
| Statsig | Authenticated project MCP for experiments, metrics, gates, configs, segments, and audit logs. Clearest experiment-first administration. | Create and update experiments, gates, and configs with a write-permission API key; read-only users stay read-only. | 9.5/10 |
| LaunchDarkly | Hosted account MCP for feature management, AgentControl, observability, and metrics. Best flag-centric operating model. | Manage flags and configs; verify the exact experiment iteration actions and regional availability. | 8.5/10 |
| Statsig Docs MCP | Public read-only documentation endpoint at /api/mcp. Developer and docs context, not an account connection. | No project data or experiment actions. | Not rated |
| Generic automation proxy | A connector such as Zapier MCP calling connected apps. Useful bridge, not a first-party experimentation MCP. | Depends on the selected app action, permissions, and proxy. | Not rated |
PostHog, Statsig, and LaunchDarkly each solve a different part of the experimentation loop. Choose by the context the assistant must read and the change it may prepare or execute.
9.5/10
PostHog's official MCP guide describes a free hosted endpoint for MCP-compatible clients. Its documented examples cover analytics queries, feature-flag shipping, error debugging, and experiment management. The breadth is useful when a test is one part of a product investigation, but it also increases governance complexity. Scope connected features and require approval for flag changes, experiment creation, and actions that affect production traffic.
9.5/10
Statsig documents tools for audit logs, dynamic configs, experiments, autotunes, gates, layers, metrics, parameter stores, segments, reviews, and observability. Its Codex guide documents listing, creating, and updating experiments plus managing gates and configs. Read-only users can remain read-only, while write tools require a write-permission API key. Treat that key as a production capability and test bulk or cleanup operations in a safe project.
8.5/10
LaunchDarkly documents a hosted MCP server for feature flags, AgentControl configs, observability data, and metrics, with examples for creating a flag, changing targeting rules, and enabling a flag across environments. It is compelling for flag-led rollout and measurement, but the published MCP page does not promise every experiment lifecycle action as a dedicated tool. Verify create, start, stop, iteration, and regional availability.
Official support is not full parity with a console or API. Statsig's Docs MCP is public, read-only, and serves English documentation. It does not expose experiments, gates, or metrics. An automation proxy can make an API action reachable, but the vendor's account permissions, audit semantics, and support commitments do not automatically transfer.
Docs MCP
Use a documentation MCP to answer implementation questions, discover tool names, and understand configuration. It is valuable context, but it is not permission to inspect an account or change an experiment.
Account MCP
An authenticated account MCP reads or changes experiments, flags, metrics, and other project objects. Check the role, API key, plan, region, and exact exposed tools before treating a write as safe.
Proxy
A proxy can route selected app actions through another service. Record the connected account, action, field mapping, credential scope, and audit path. Do not score it as a first-party experimentation MCP.
1. Read first. Retrieve active experiments, metrics, flag targeting, exposure counts, and recent result changes without writing.
2. Model the decision. Capture the hypothesis, primary metric, expected effect, audience, allocation, stop rule, and rollback path in a Wireslate Whiteboard. Add the economic effect of a win, loss, or inconclusive result.
3. Prepare the smallest change. Let the specialist MCP create a draft experiment, gate, metric, or targeting rule only when the exact operation is documented. Use a read-only key for analysis and a separate scoped write credential for approved configuration work.
4. Approve and measure. A human decides whether traffic should start or a production flag should change. Return lift, uncertainty, guardrails, and the decision to the Wireslate board after the run.
Wireslate MCP Server
Connect ChatGPT, Claude, Codex, or another AI tool to a Wireslate Whiteboard. Map the hypothesis and economics, review the test with your team, then send approved configuration to the experimentation platform and bring measured results back.
The MCP specification calls for explicit consent, understandable authorization, and access controls. MCP does not make experiment design correct by itself. It cannot choose a valid randomization unit, detect every sample-ratio mismatch, or replace statistical review. Result freshness may also differ between stored reports, event queries, and metric surfaces.
The practical workflow is simple: read the current state, record assumptions and economics, prepare the smallest scoped change, approve live traffic separately, and measure against the agreed definitions. Keep a change record and verify any flag or config that can affect production users.
Methodology: researched September 21, 2026 using first-party vendor documentation and the official MCP specification. Ratings reflect documented action breadth and governance fit, not live testing, statistical methodology, pricing, uptime, or security posture. Recheck current tools, roles, plan, region, and limits in a non-production project before implementation.
Experimentation MCP coverage changes as vendors add tools, plans, regions, and permissions. Run a non-production test for the exact action that matters.
PostHog is the broadest documented option for analytics, flags, and experiments in one connection. Statsig is the clearest experiment-first option, while LaunchDarkly is strongest for flag-led rollout and measurement. Confirm the exact tool, role, region, plan, and approval path before choosing.
Sometimes, but do not infer launch rights from read access or from the existence of a server. Statsig documents experiment creation and updates with write permissions. PostHog documents read and write across products. LaunchDarkly documents flag and config management, while its experiment docs describe additional iteration steps that must be verified in the current tool list.
No. A docs MCP retrieves product documentation. Statsig's Docs MCP is explicitly public and read-only; its authenticated Statsig MCP is the connection for project data such as experiments, gates, and metrics.
Start read-only. Ask the assistant to summarize the hypothesis, exposure, metric definitions, result window, lift, uncertainty, and guardrails. Put the decision model in Wireslate, then use a scoped write connection only for an approved configuration change.
Wireslate is the shared visual and economic plan. The specialist experimentation MCP implements the approved flag or test and returns measurements. Wireslate holds the pre-test assumptions, human review, post-test results, and decision record.
Claims were checked against public product and help material. A red X means the feature was not identified in the reviewed material; a yellow ~ means the tool has a related capability but not the same workflow. These are not claims about every plan or future release.
Map the hypothesis, economics, guardrails, and rollback in a Wireslate Whiteboard, review it with the team, then hand approved configuration to the specialist MCP.