Baton — Product analytics for MCP servers

Your customer, to their agent

Claude

ChatGPT

Copilot

Session 5b917f6c · Claude Code · 1 task · 42 calls · tool call failure 14 retry loop 6 unrecovered failure +1 more satisfaction 2 of 5

First intent Build a dashboard tracking the Northwind launch for leadership

Expected result A dashboard showing each project's health, dates and progress

Written by a language model satisfaction 2 of 5 Covers 42 calls

What they came for

They wanted a leadership dashboard tracking the Northwind launch: every project's health, dates, progress, and the flagged work needing intervention.

What happened

The Northwind project set was pulled with status, progress and dates, and each project summarised in turn, including the ones marked In Trouble and At Risk.18 calls

Every attempt to add a project-level table was rejected for an invalid field path. The agent tried the owner field seven ways, ran a field search, and tried again with the path the search returned. Same rejection.19 calls

It never recovered. The last five calls were the same rejected query with the field name reworded.5 calls

Where it ended

The session stopped still retrying the same rejected field path, leaving a dashboard with counts and a chart but no project-level table.

Recurring failed intents

User intent	Sessions	Sessions with friction	Most common error
Build a leadership dashboard	34	26	invalid field path
Find a project by name	21	9	invalid field path
Send assets to the asset library	17	12	folder search failed: 400
Attach an approval workflow	11	4	missing project_id

Recurring errors

Error pattern	Tool	Sessions	Calls	Fix location
Invalid field path(s) — call get_field_paths() to retrieve valid paths.	projects_find_data	100	164	tool description
Upstream service returned HTTP {n} Bad Request	projects_get_field_paths	33	40	tool implementation

Recurring errors

Error pattern	Tool	Sessions	Calls	Fix location
▼Invalid field path(s) — call get_field_paths() to retrieve valid paths.	projects_find_data	100	164	tool description
The rule is already in the server's own docs CRITICAL — every `field_path` must be a verbatim path from `get_field_paths()`. from the projects-mcp condition guide Prompt Copy prompt You are working in the repository that contains the projects MCP server. Goal: stop a recurring agent error by putting a documented rule into the tool description, where every agent receives it automatically. The error, on the tool projects_find_data: Invalid field path(s) — call get_field_paths() to retrieve valid paths. It occurred on 164 calls across 100 sessions in the last observed window. The rule that prevents it already exists in the server's own docs: From: the projects-mcp condition guide CRITICAL — every `field_path` must be a verbatim path from `get_field_paths()`. Steps: 1. Find the tool definition for projects_find_data (name, description, input schema). 2. Append these lines to its description. Change nothing else. Copy as a prompt for your coding agent, verification steps included.
▶Upstream service returned HTTP {n} Bad Request	projects_get_field_paths	33	40	tool implementation
▶{n} validation errors for call[find_data]	projects_find_data	21	37	tool description
▶Tool (search_records) input validation failed: array found, [object, null] expected	projects_search_records	21	34	tool description
▶calling {v}: MCP error {n}: Unknown docs URI {v}	projects_read_docs	16	18	tool description
Show all 49 rows

Every session, explained

More than just tool calls and HTTP response codes. When your MCP server can't satisfy a user request, your logs often show a 200 anyway. Baton writes the plain-language account: what the customer came for, what happened, and where it ended.

Session 5b917f6c · Claude Code · 1 task · 42 calls · tool call failure 14 retry loop 6 unrecovered failure +1 more satisfaction 2 of 5

First intent Build a dashboard tracking the Northwind launch for leadership

Expected result A dashboard showing each project's health, dates and progress

Written by a language model satisfaction 2 of 5 Covers 42 calls

What they came for

They wanted a leadership dashboard tracking the Northwind launch: every project's health, dates, progress, and the flagged work needing intervention.

What happened

The Northwind project set was pulled with status, progress and dates, and each project summarised in turn, including the ones marked In Trouble and At Risk.18 calls▸

Showing the 18 of 42 calls this part was written from · in 1 of 1 taskshow all calls

+0s

projects_find_id_by_name

Find the Northwind project GUID for further data retrieval

expected: A project GUID and name matching "Northwind"

▸ request & response

1.1s · 5 KB

+0s

projects_get_current_user

Get the base URL for building deep links in the dashboard

▸ request & response

504ms · < 1 KB

+9s

projects_find_data

Get all Northwind projects with status, percent complete, and dates

expected: List of all Northwind projects with statuses and date fields

▸ request & response

2.7s · 38 KB

+10s

projects_find_id_by_name

Find any portfolio named Northwind that may contain these projects

expected: A portfolio GUID for Northwind

▸ request & response

699ms · < 1 KB

+ 14 more calls

Showing the 19 of 42 calls this part was written from · in 1 of 1 taskshow all calls

+6m 21s

projects_find_data✕

Add a project-level table grouped by owner

expected: Rows grouped by owner

Invalid field path(s) — call get_field_paths() to retrieve valid paths.

▸ request & response

1.3s · < 1 KB

+6m 32s

projects_get_field_paths✕

Retrieve valid field paths for the owner field

Upstream service returned HTTP 400 Bad Request

▸ request & response

757ms · < 1 KB

+6m 42s

projects_find_data✕repeat of an earlier call

Add the table with the path the search returned

Invalid field path(s) — call get_field_paths() to retrieve valid paths.

▸ request & response

1.4s · < 1 KB

+6m 50s

projects_find_data✕repeat of an earlier call

Invalid field path(s) — call get_field_paths() to retrieve valid paths.

▸ request & response

1.2s · < 1 KB

+ 15 more calls

It never recovered. The last five calls were the same rejected query with the field name reworded.5 calls▸

Showing the 5 of 42 calls this part was written from · in 1 of 1 taskshow all calls

+7m 02s

projects_find_data✕repeat of an earlier call

Retry with the field name reworded as ownerName

Invalid field path(s) — call get_field_paths() to retrieve valid paths.

▸ request & response

1.3s · < 1 KB

+7m 09s

projects_find_data✕repeat of an earlier call

Retry with the field name reworded as owner.name

Invalid field path(s) — call get_field_paths() to retrieve valid paths.

▸ request & response

1.1s · < 1 KB

+ 3 more calls

Where it ended

The session stopped still retrying the same rejected field path, leaving a dashboard with counts and a chart but no project-level table.

How capture works

Baton adds three parameters to every tool you already have: the goal, the expected result, and the task the agent is working through. The agent fills them in with each call, and Baton strips them out before your code runs. One added tool lets the agent send feedback directly: annotate, for friction as it happens. Your customers never see Baton.

What Baton adds

params

user_goal · expected_result · overall_task

Three parameters added to every tool you already have. The agent fills them in with each call; Baton strips them out before your code runs.

projects_find_data {
  entity: "task",
  fields: ["owner:name", "status"],
  user_goal: "Add a project-level table to the dashboard",
  expected_result: "Rows grouped by owner",
  overall_task: "Build a leadership dashboard"
}

tool

acme_annotate

The agent reports friction as it happens: what it tried, what it expected, what happened instead.

Choose your Baton configuration

For a single MCP server you control

import { withBaton, HttpSink } from "@goodtiming/baton-sdk";

withBaton(server, {
  vendorId: "acme",
  vendorDisplayName: "Acme",
  consentToken: process.env.BATON_CONSENT_TOKEN,
  sink: new HttpSink(BATON_CONSOLE_URL, {
    apiKey: process.env.BATON_API_KEY,
  }),
});

You control what's captured. You can audit the rest.

PII never leaves raw

Every payload is scrubbed before it leaves your process: emails, keys, tokens, card numbers, and sensitive fields, on by default.

Deploy where you're comfortable

Keep everything inside your perimeter, send events to an endpoint you run, or use Baton's cloud, hosted single-tenant. The proxy can write straight to your own S3.

Auditable end to end

Every event is logged and inspectable. You define the capture schema, so what you don't capture can never reach us.

Your telemetry, if you want it

Events can also route to your own OpenTelemetry sink. Possible, supported, and rarely the place to start.

Questions, answered

I can already see tool calls on my server. How is this different from my logs and observability products?

Your logs show the call and the status code. Baton captures what the customer was trying to do, whether they got it, and the fix. It runs alongside Sentry and LangSmith and sees the layer they don't: intent.

Do I have to rebuild anything?

No. The SDK builds in with about five lines and your tools stay exactly as they are. If your customers reach you through agent-generated code, a Skill calling your API, the SDK's Library API instruments those calls instead.

Where does my data go?

Nowhere, by default. The open-source capture writes to a local file or a sink you choose: your own OpenTelemetry endpoint or your S3. PII is scrubbed before anything leaves. You only send to the Console if you choose to.

What's free, and what's paid?

Everything is free to start, including the Console. Capture is open source, and light use of the Console costs nothing. You pay once you pass a monthly session threshold.

Do my customers' agents have to change?

No. You instrument your side, once. Whatever agent your customers use, Claude, ChatGPT, Copilot, or anything else, keeps working exactly as before.

Product analytics for MCP servers

Every session, explained

How capture works

What Baton adds

Choose your Baton configuration

You control what's captured. You can audit the rest.

PII never leaves raw

Deploy where you're comfortable

Auditable end to end

Your telemetry, if you want it

Questions, answered

I can already see tool calls on my server. How is this different from my logs and observability products?

Do I have to rebuild anything?

Where does my data go?

What's free, and what's paid?

Do my customers' agents have to change?

See your first session today