Trunk MCP server: four tools for flaky tests

A narrow hosted connector for flaky-test triage, which three of its four tools Trunk marks experimental, and what it does not reach.

Last verified 27 August 2026 · from Trunk's own MCP reference documentation

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the Trunk MCP server?

It is Trunk's server for the Model Context Protocol. Trunk's Flaky Tests product tracks intermittently failing tests, so a connected agent can tell a real regression from known noise.

The workflow is short by design. search-test looks up a test's identifier, fix-flaky-test returns insights about that test, and investigate-ci-failure pulls failing test logs from a CI run.

setup-trunk-uploads is the odd one out and the only tool without an experimental label. It exists for repositories that are not sending Trunk any test results yet.

Trunk's hosted server is focused on Flaky Tests: search, investigation, a proposed fix and upload setup. That makes it useful for test failures in CI, not for runtime, model or spend questions.

For an agent, that means Trunk answers whether a failing test is already known to be flaky, and hands back the log that failed. It cannot re-run the build and it cannot show the code. Pair it with a pipeline connector and a repository connector.

What can an AI agent do with Trunk?

search-test exists because the other tools work from a test identifier rather than a name. Running it first is the whole sequence.

What connecting costs before the first question

A context window is the amount of text a model can hold at once. Four flaky-test tools cost 540 to 683 tokens on Claude. Three of the four are labeled Experimental in Trunk's own table.

The surface is small, so the experimental labels cost little to live with. Connecting to find out spends 540 to 683 tokens on Claude.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted from the 4 tool entries in the vendor's own documentation. That is the only public inventory.

The server could not be run for a live tools/list. So no input schemas are available. These figures cover each tool's name and description only. They are a floor, not the figure.

The Claude columns are measured through the API's count_tokens endpoint, with an empty schema per tool. So they cover the same content as the GPT columns. The range there is two tokenizer generations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a quarter.

Trunk MCP server limits: where answers come back incomplete

The three limits below come from a narrow surface rather than from anything failing.

How to configure Trunk MCP for agents

1. 1

   Expect experimental behavior

   Trunk marks three of its four tools Experimental, which is a statement about stability rather than about access.

2. 2

   Prefer OAuth

   It is the documented default, and clients that support the MCP authorization specification start the flow automatically.

3. 3

   Use a token for headless use

   Trunk documents an organization API token for CI and for clients without OAuth support.

4. 4

   Know the boundary

   This is the Flaky Tests product only. Trunk's linters and its merge queue, which tests pull requests before they merge, have no tools here.

Paste this into your agent’s instructions

Three tools are experimental and their output shape can change. The block makes the agent name which one produced a claim.

When you answer using Trunk MCP tools, state:
- The tool you called. Three of the four are labeled
  Experimental by Trunk.
- The repository and CI run the answer covers.
- Whether a test was identified by search-test first, since
  the other tools work from an identifier.
Never state that data does not exist. State that the tool
above returned no results for that test or run.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Not for a weekly review. Trunk's web application shows the same flakiness data and costs no context at all.

The connector earns its place mid-failure, when the agent can ask whether this test is already known to be flaky before anyone investigates the code.

For a broader test-analytics surface, BuildPulse covers flakiness, coverage and CI submissions across more tools.

Connect it during triage. Use the dashboard for periodic review of the same numbers.

Trunk MCP server setup

The server is hosted, so there is nothing to install. OAuth 2.0 with OpenID Connect is the default and most interactive clients use it.

For CI or clients without OAuth, Trunk documents an organization API token found under Settings, Organization, General, in the API section.

claude mcp add --transport http trunk \
  https://mcp.trunk.io/mcp

{"mcpServers":{"trunk":{
  "url":"https://mcp.trunk.io/mcp"}}}

[mcp_servers.trunk]
url = "https://mcp.trunk.io/mcp"

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

{"mcpServers":{"trunk":{"url":"https://mcp.trunk.io/mcp"}}}

{"mcpServers":{"trunk":{"url":"https://mcp.trunk.io/mcp"}}}

{"servers":{"trunk":{"url":"https://mcp.trunk.io/mcp"}}}

{"mcpServers":{"trunk":{"url":"https://mcp.trunk.io/mcp"}}}

{"mcpServers":{"trunk":{
  "url":"https://mcp.trunk.io/mcp",
  "headers":{"Authorization":"Bearer <your-token>"}}}}

Read-only access, permissions and security

• Every documented tool reads

  Nothing in the four creates, deletes or triggers a build.

• OAuth is the default path

  Clients implementing the MCP authorization specification start the flow without a token in a config file.

• The API token is organization-scoped

  It comes from the organization's own settings rather than from a personal profile. Anyone holding it acts as the organization, not as one user.

• Experimental labels are Trunk's own

  Three of the four tools carry them, which is worth knowing before building a workflow on their output shape.

Troubleshooting

The client cannot connect

An unauthenticated request returns 401 with a missing-authorization message. Let the client run OAuth, or supply the organization API token as a bearer header.

A tool cannot find the test

The insight tools work from a Trunk test identifier. Call search-test first to resolve a name into one.

There is no flakiness data at all

The repository may not be uploading results yet. setup-trunk-uploads creates a plan for doing that.

Output changed between sessions

Three of the four tools are labeled Experimental in Trunk's documentation, so their shape is not guaranteed stable.

Trunk MCP server: Reference

What engineers report

The documentation is short and specific, so the counts below come from it directly.

Tool inventory and authentication routes read from Trunk's own MCP reference on 24 August 2026. An unauthenticated tools/list against mcp.trunk.io/mcp returned 401 with a missing-authorization message on the same date, which is why this page is sourced from documentation rather than from the wire.

Should you connect an agent to Trunk?

Yes. Best for OAuth from an interactive client, with the organization token kept for headless use.

Strongest fit

Trunk Flaky Tests users who want a quick known-flake check without leaving the editor.

Main advantage

A tiny, focused surface with OAuth by default and a documented token route for CI.

Main weakness

Three of the four tools are experimental, and the connector covers one Trunk product.

Operational risk

None from the tools themselves; all four read and none triggers a build.

Frequently asked questions