Through Amplitude’s official MCP server, on either the US or EU endpoint, Buda agents reach your project data. Agents can search dashboards and charts, run queries, read cohorts, experiments, and session replays, and prepare analysis — from either the US or EU endpoint.
The question is rarely “what does this chart say". It is "why did this number move, and does the answer hold up when you look at it three other ways".
That is what this integration opens up: agents query charts and metrics through Amplitude’s official MCP server, then run the follow-up query that tests whether the first answer holds. Agents can search existing dashboards and notebooks, query charts and metrics, pull cohorts and experiment results, read session replays and customer feedback, and assemble the analysis a person then checks.
Amplitude holds the behavioral evidence; a connected agent can interrogate it rather than just display it.
A chatbot can interpret a screenshot of a funnel. An Amplitude-connected agent can run the query itself, then run the follow-up query that tests whether the first answer was real — segmenting by cohort, checking the same window last month, and looking at whether an experiment was running at the time.
Instead of building five charts to chase one drop, you can ask:
Activation fell about eight percent last week. Query the funnel by platform and by acquisition source, check whether any experiment was live, and tell me where the drop is concentrated.
The agent runs the queries, compares the segments, and reports what the data supports and what it does not.
Your event data never leaves Amplitude; what Buda adds is somewhere an analysis gets stress-tested.
Amplitude’s MCP server exposes tools across analytics, experimentation, replay, and feedback surfaces, governed per project by RBAC.
Buda agents can retrieve supported information including:
With write permission on the project, the tool set extends to creating objects:
Access is project-scoped and inherited from the authenticated user’s Amplitude RBAC, governed by two actions: USE_MCP_READ for read access and USE_MCP_WRITE for creation and edits. Every role grants read; Member, Manager, and Admin roles also grant write by default. Organization admins can disable MCP entirely. Choose the endpoint matching your data residency — the standard endpoint for US, or the EU endpoint if your Amplitude data resides there.
One query is easy. The discipline that separates a real finding from a coincidence is running the checks that could disprove it — and that is the part people skip when time is short.
The first explanation for a metric change is usually incomplete.
Experiment readouts are where wishful thinking enters the process.
Cohort definitions encode assumptions worth stating explicitly.
Example workflowNumbers say what happened; feedback says why.
Funnel drops are often instrumentation problems.
Analysis quality is capped by instrumentation quality.
Four patterns, each ending with an analyst confirming the reading.
Trigger: Run every Monday morning.
Buda agent workflowTrigger: An experiment reaches its sample target.
Buda agent workflowTrigger: A conversion step falls below threshold.
Buda agent workflowTrigger: Thirty days after a feature launch.
Buda agent workflowGive a connected Buda agent any of these, replacing the metrics and windows with the ones you track.
Weekly active users dropped about six percent. Query it segmented by platform and acquisition source, compare with the same week last month, and tell me where the loss actually is.
Read the checkout redesign experiment. Query the primary and guardrail metrics per variant and tell me whether the result holds in every major segment.
Query the onboarding funnel for the last two weeks. Tell me which step lost conversion and whether event volumes suggest a tracking problem rather than a real drop.
Build a cohort of users who completed onboarding but did not return within seven days. Tell me how sensitive the count is to changing that window.
Compare the drop in feature usage this month against customer feedback from the same period, and tell me whether the themes line up.
Compare our tracking plan against events actually received in the last thirty days and list anything that stopped firing or is inconsistently propertied.
A saved dashboard is the right tool when you already know which numbers to watch.
For example:
Show weekly active users, activation rate, and retention on one screen every Monday.
An MCP-connected agent earns its place when the question is open-ended and the next query depends on the last answer.
Activation dropped. Work out whether it is a real behavior change, a tracking problem, or a segment mix shift — and show me the evidence for whichever it is.
| Capability | Saved Dashboards | Buda Amplitude MCP Integration |
|---|---|---|
| Monitoring known metrics | Strong | Supported as one step of a workflow |
| Natural-language instructions | Limited | Yes |
| Follow-up queries that depend on results | Manual | Agent chains queries |
| Segmenting until the cause is isolated | Manual chart building | Agent iterates and reports |
| Distinguishing tracking gaps from real drops | Requires an analyst | Agent can check event volumes |
| Cross-tool context | Requires separate integrations | Combines with other Buda tools |
| Explaining what the data does not support | Not applicable | Yes, stated explicitly |
| Draft-first analysis | Not applicable | Natural agent pattern |
| Read and write RBAC separation | Not applicable | Enforced per project |
For example:
Use dashboards for the numbers you already watch. Use a Buda agent for the investigation that starts when one of them moves.
A metric change is usually caused by something that happened in another system.
A release shipped, a campaign started, a bug was filed, and a customer complained — each in a different tool. An agent that reads those systems together can test whether the timing actually lines up.
Check whether a behavioral change begins at a specific deployment.
Trace an adoption shift back to the pull requests that changed the flow.
Attach usage evidence to the issue proposing a change.
Compare product engagement against pipeline and account health.
Test whether the assumptions written into a product spec survived contact with real usage.
Lining those systems up is how you test whether your explanation for a number is actually supported.
Access runs through Amplitude’s own server, on whichever regional endpoint matches your data residency.
Authentication uses OAuth 2.0 and the connection respects your existing Amplitude permissions. Agents can only reach projects and data the authenticated user could already view. Two RBAC actions govern the boundary per project: USE_MCP_READ for reading and USE_MCP_WRITE for creating and editing. Every role grants read by default; Member, Manager, and Admin roles also grant write. Organization admins can disable MCP entirely.
Teams with EU data residency should connect the EU endpoint rather than the standard one, so queries stay within the intended region.
USE_MCP_READ only to begin with.USE_MCP_WRITE where saved objects need to be created.An agent can run more queries than an analyst has time for. That makes it a strong research assistant and a poor decision-maker, because the judgment that matters is knowing which result to trust.
Buda is built the same way: agents query and compare, and an analyst decides which result is worth trusting.
Agents can query, segment, compare, correlate, and draft.
People should keep approving anything involving:
The teams that gain most are the ones whose analysis backlog is longer than their analyst capacity.
Investigate a metric move without waiting for an analyst to be free.
Skip query assembly and spend the time on whether the result holds.
Test funnel and cohort hypotheses in one pass instead of five chart builds.
Check whether launch adoption matches what was projected.
Pair behavioral evidence with feedback and session replays on the same question.
Audit tracking coverage against what the plan says should be firing.
It is a connection that lets another application work with Amplitude project data — querying analytics and reading experiments, governed by the user’s existing role.
Through Amplitude’s official MCP server, on either the US or EU endpoint. Agents authenticate over OAuth 2.0 and operate under your project RBAC.
It is the arrangement that lets an agent turn an analytical question into the right sequence of Amplitude queries, instead of a person building each chart or API call.
A Buda agent can search dashboards, charts, and notebooks, query charts, metrics, and datasets, read experiments, feature flags, cohorts, users, session replays, taxonomy, and feedback, and — with write permission — create dashboards, experiments, and cohorts, and save chart edits.
Yes. Amplitude documents a separate EU endpoint alongside the standard US one. Connect the endpoint matching where your Amplitude data resides.
Access is project-scoped and inherits your Amplitude RBAC, governed by USE_MCP_READ and USE_MCP_WRITE. Every role grants read; Member, Manager, and Admin also grant write by default. Admins can disable MCP for the organization.
The connection uses OAuth 2.0 and respects existing Amplitude permissions. The agent can only reach projects and data the authenticating user could already view.
Amplitude’s MCP tools support creating dashboards, experiments, and cohorts, and saving chart edits, where write permission is granted. This is a metadata capability, not a data-ingestion path.
No. You describe the question in natural language. Trustworthy analysis still requires a person deciding whether the result is sound before it informs a decision.
No. Buda sits alongside it, turning behavioral data into investigation and reporting an analyst validates.
No. Amplitude AI, including its Global Agent and specialized agents, works inside Amplitude on Amplitude data. Buda sits outside Amplitude, turning behavioral data into investigation and reporting that an analyst validates.
Read access covers most of the value. Grant write only where saved cohorts, dashboards, or experiments genuinely need to be created, and keep definitions used for targeting under human review.
The hard part of product analytics is not producing a number. It is running the follow-up queries that tell you whether your first explanation survives contact with the data.
Connect Amplitude to Buda and let agents do that work, then hand an analyst a finding with the evidence and the caveats attached.