In many organisations, the observability tool is restricted territory: SREs and developers use it every day, while the rest of the company goes through them for the smallest answer. A product owner wondering whether the latest release degraded the sign-up journey posts in a Teams channel; someone on call drops what they were doing, opens three apps, assembles an answer and pastes it into the conversation. Multiply that scenario by the number of business teams and you get a workload that is invisible but very real. A Dynatrace / Microsoft Teams agent is aimed squarely at that workload.
The Dynatrace MCP server offers another way. The Model Context Protocol is an open standard that lets an AI agent call external tools in a structured way. Connected to Microsoft Copilot Studio, it lets you publish a conversational agent in Teams that can query your Dynatrace SaaS environment live, instead of answering from general knowledge.
This article does not just describe the setup. It focuses on the questions a platform team will ask before opening this channel to hundreds of users: who sees what, how to frame the agent, and where to start so that the experience is useful rather than a gimmick.
What an MCP-connected Dynatrace Microsoft Teams agent changes
A general-purpose assistant can explain what 95th-percentile latency is, but it knows nothing about your payment service. An agent connected to the Dynatrace MCP server has tools that let it retrieve the real context: open problems, service health, data stored in Grail. Its answer is therefore based on your data rather than on statistical plausibility.
The chain has four links, each with a clearly defined role:
- Teams is the interface: this is where the question is asked, on desktop or mobile;
- the Copilot Studio agent interprets the request and chooses which tools to call;
- the Dynatrace MCP server exposes those tools and returns the requested context;
- the Dynatrace platform remains the source of truth for observability and business observability data.
The end user never touches Dynatrace directly. That is both the point of the approach and the reason governance needs care.
Before you start: the question of permissions
Access to data goes through the permissions and policies defined on the Dynatrace side and on the Microsoft side. In practice, the connection between Copilot Studio and the MCP server uses a Dynatrace platform token. Anything that token can read, the agent can potentially return to anyone who has access to the agent in Teams.
So you need to think of it as a shared service account:
- grant only the read permissions needed for the target use case, nothing more;
- ask whether some data (customer identifiers, application log content) should stay out of reach, and adjust the policies accordingly;
- limit the agent’s rollout in Teams to a population consistent with those permissions, at least at first;
- plan token rotation and name a clearly identified owner.
This is also the time to check internal rules on using AI assistants with production data. Better to settle these points before the demo than after. Our guide to securing an MCP server covers these principles in more detail (token scope, least privilege, logging).
In practice
Step 1 — Create the agent in Copilot Studio
Create a new agent and give it an explicit name that shows where its answers come from (for example « Observability assistant »). Then write its instructions: this is where much of the perceived quality is decided. Describe its role, its scope and the expected tone. A possible starting point:
Role: you help product and support teams understand
the state of applications using Dynatrace data.
Scope: service health, ongoing problems, impact
of recent deployments, user journeys.
Rules:
- back every statement with data retrieved through the tools;
- always state the period analysed;
- if the data does not allow a conclusion, say so;
- suggest a concrete next investigation step.
Context: our main applications are "boutique-web",
"api-commandes" and "paiement".
The more you describe your environment (application names, business processes tracked, internal vocabulary), the better the agent will be able to link a question phrased in business language to the right entities.
Step 2 — Enrich the agent with Dynatrace knowledge
The Dynatrace for AI project provides skills, prompts and instructions that guide an agent in using the platform. Including them saves every team from reinventing its own way of querying Dynatrace and makes answers more consistent from one user to the next.
Step 3 — Configure the MCP connection
Add a connection to the Dynatrace MCP server in Copilot Studio. Three things are needed: access to your Dynatrace environment, a platform token with the permissions chosen in the governance step, and the endpoint of the corresponding MCP server. Refer to the MCP server documentation for the current list of permissions required by each tool.
To check what the agent is supposed to get, it helps to run, yourself, in a Notebook, a query equivalent to what it should retrieve. For example, for a question about ongoing problems:
fetch dt.davis.problems
| filter event.status == "ACTIVE"
| fields display_id, event.name, event.start
| sort event.start desc
If the agent’s answer differs markedly from this result, the issue lies in the instructions or the permissions, not in the data.
Step 4 — Test with real questions
Pull from your Teams channel history the questions actually asked of technical teams over the last few weeks and replay them. A few typical examples: « did yesterday’s deployment change the conversion rate? », « which business processes are affected by the current incident? », « how many orders failed during the outage? ». For each one, note whether the answer is correct, sourced and actionable. Failures reveal either a gap in the instructions or context missing from Dynatrace.
Step 5 — Publish to Teams
Once the results are satisfactory, publish the agent to the Teams channel. Start with a small pilot group, collect feedback, then widen the rollout.
Choosing a first use case
It is tempting to promise an agent that answers everything. That is the best way to disappoint. Choose a single high-value, measurable scenario and make it excellent:
Incident status updates
During an incident, non-technical stakeholders want to know what is affected and since when. An agent that answers these questions frees the incident manager from repetitive requests.
Validating a release
After a deployment, the product owner wants to know whether their business indicators have moved. This is a case where Dynatrace business observability brings real value, provided business events are already being collected.
Service health questions
Support teams can check a service’s status before escalating a customer ticket, without asking for full access to the platform.
Best practices
- Least privilege: the agent’s token should read only what all its users are allowed to see.
- Dated, sourced answers: require in the instructions that the agent mentions the period and the nature of the data used.
- A reference question set: keep your test questions and replay them after every change to the instructions or permissions.
- A named owner: the agent is an internal product, with an owner, a feedback loop and a roadmap.
- Do not replace the tool: for an in-depth investigation, the agent should point to Notebooks, Dashboards or the relevant app rather than summarise everything.
Frequently asked questions
Do Teams users need a Dynatrace account?
No: they never touch Dynatrace directly. The agent queries the MCP server with a platform token, and anything that token can read can be passed on to anyone with access to the agent. Hence the importance of limiting its permissions and the agent’s rollout.
Which permissions should the agent’s token have?
Only the read permissions needed for the target use case, leaving out any data that must stay out of reach. The MCP server documentation gives the up-to-date list of permissions each tool requires.
How do you check that the agent answers correctly?
Run the DQL query equivalent to the question in a Notebook and compare. Also keep a set of real questions taken from your Teams channels and replay it after every change to the instructions or permissions.
Conclusion
Connecting Dynatrace to Teams through MCP and Copilot Studio takes only a few configuration steps. Most of the work lies elsewhere: defining permissions, writing precise instructions and choosing a first use case that speaks to business teams. Well framed, this kind of agent cuts the back-and-forth between teams and brings observability data closer to the people who make product decisions, without turning every Teams user into a Dynatrace administrator.
Further reading: How to connect Dynatrace MCP to Microsoft Teams using Copilot Studio (Dynatrace blog).
span.observer is an independent community blog, not affiliated with Dynatrace LLC. Dynatrace is a trademark of Dynatrace LLC. SaaS features evolve quickly: check their availability in your environment and in the official documentation.
Originally published on span.observer. French version: https://span.observer/fr/dynatrace-mcp-microsoft-teams-copilot-studio/.
Top comments (0)