Honeycomb MCP Server

The Honeycomb MCP Server connects AI assistants such as Claude Code, Cursor, Claude Desktop, Codex, or VS Code to the Honeycomb observability platform. Anyone monitoring their applications with Honeycomb can ask the assistant questions like "Why were response times elevated last night?", "Which spans cause the most errors?", or "How is our SLO doing?" without switching between the editor and the web interface. The server is operated by Honeycomb itself. According to the documentation it lets agents query, explore, and iteratively analyze telemetry data, much as people do in the interface.

What is the Honeycomb MCP Server?

Honeycomb stores telemetry as wide events and traces and makes it searchable through queries with aggregations, filters, and groupings. The MCP server exposes these capabilities as tools. According to the configuration guide the address depends on the region: https://mcp.honeycomb.io/mcp for the US and https://mcp.eu1.honeycomb.io/mcp for the EU. The guide names as prerequisites that Honeycomb Intelligence is enabled for the team and that the agent supports remote MCP servers or plugins.

Which tools are available?

The reference describes 34 tools, 25 read-only and 9 that write. Exploration covers the workspace with environments and datasets, the column schema, the search for columns, and the registry of semantic conventions for attributes. For the analysis itself there is run_query for time-series aggregations, run_bubbleup, which examines how a conspicuous subset differs from the baseline, and tools for traces, span names, and span details. You can also read boards, triggers, SLOs, notification recipients, AI conversations, and the Refinery documentation. Writing tools can create or change boards, triggers, SLOs, and recipients, talk to the Canvas agent, and send feedback to the Honeycomb team. Besides tools the server provides prompts, prebuilt multi-step workflows that you deliberately start as a slash command.

What is it particularly useful for?

The value lies in troubleshooting with production context. After a deployment an assistant can investigate latency and error spikes, find outliers, use BubbleUp to isolate the conspicuous attributes, and follow individual traces as a waterfall. According to the documentation, existing dashboards and alerts can also be translated into Honeycomb's query language, which eases migrations. Compared with a generic database server, this server knows the platform's concepts, such as traces, spans, calculated fields, and SLOs, and returns answers that fit the monitoring environment directly.

Setup

Signing in via OAuth is recommended: on the first tool call the browser opens for sign-in and consent. For unattended agents a management API key with the scopes "Model Context Protocol" and "Environments" is possible, which only team owners can create. For Claude Code the documentation describes a plugin with the /honeycomb-setup command as well as manual registration with claude mcp add using the HTTP transport; for Cursor and VS Code there are one-click links, for Claude Desktop and ChatGPT directory entries. With the X-MCP-Toolset header you can limit the offering to query or canvas; the header grants no additional rights.

Security and limits

Most tools are read-only. According to the documentation, writing tools require the scope mcp:write, which you grant explicitly during OAuth consent or API key configuration. Grant it only if the agent should actually change boards, triggers, or SLOs. The activity log environment is exposed through MCP only to team owners. Telemetry the assistant retrieves enters the AI client context and therefore reaches the model provider; span attributes can contain personal data. The page gives no information on rate limits or retention periods.

Costs

According to the pricing page, Honeycomb Intelligence and the MCP server are included in all three plans, including the permanently free entry plan. Above it sit a paid Pro plan with more triggers, SLOs, single sign-on, and support, and an Enterprise tier offered individually. Individual tools such as the service map are reserved for the Enterprise plan according to the reference. Which terms apply to you is shown on Honeycomb's official pricing page.

Who is the Honeycomb MCP Server for?

It is worthwhile for development, SRE, and DevOps teams that already use Honeycomb in production and want to bring production data into their development workflow. Anyone without Honeycomb data gains nothing from it: the server is not a standalone monitoring system but an access path to the platform.

A Honeycomb team with Honeycomb Intelligence enabled and an MCP client that supports remote servers, such as Claude Code, Cursor, Claude Desktop, or VS Code.

With the Honeycomb MCP Server, an AI assistant queries traces, errors, and SLOs of your Honeycomb team. This guide leads from the prerequisites to the first query.

Step 1: Check the prerequisites

Honeycomb Intelligence must be enabled for the team; according to the documentation a team owner can do this in the settings. You also need a client that supports remote MCP servers or plugins.

Choose the region

The address depends on where your team is located: https://mcp.honeycomb.io/mcp for the US and https://mcp.eu1.honeycomb.io/mcp for the EU. With the wrong region the server will not find your data.

Step 2: Set up in Claude Code

Run claude mcp add honeycomb --transport http https://mcp.honeycomb.io/mcp, using the EU address in the EU. Alternatively the documentation describes a plugin with the /honeycomb-setup command. On the first tool call the browser opens for the OAuth sign-in.

claude mcp add honeycomb --transport http https://mcp.honeycomb.io/mcp

Other clients

For Cursor and VS Code the documentation offers one-click links; alternatively enter the regional URL in the mcp.json file. Claude Desktop and ChatGPT add Honeycomb through their directories.

API key for unattended agents

If an agent runs without a browser, a team owner creates a management API key with the scopes "Model Context Protocol" and "Environments". It is passed as a bearer token in the format KEY_ID:SECRET_KEY and does not belong in version control.

Limit tools and write access

With the X-MCP-Toolset header you load only query or canvas. Writing tools that change boards, triggers, or SLOs require the scope mcp:write. Grant it only if the agent should really make changes.

Step 3: First queries

Start with "Which environments and datasets exist?" and then, for example, "Show the slowest spans of the last hour". Check generated queries before drawing conclusions from them.

OAuth via the browser (recommended) or a management API key for unattended agents, which only team owners can create.

Read access to the team's telemetry, boards, triggers, and SLOs; writing tools only with the explicitly granted scope mcp:write. The activity log is accessible only to team owners.

The server runs hosted by Honeycomb in the chosen region. Retrieved telemetry enters the AI client context and therefore reaches the model provider.

Exposure of span attributes containing personal data to the model provider, an unnecessarily granted write scope mcp:write, and an API key committed to version control.