Skip to content
MCP servers

Hear what agents think of your MCP server

Your server can return valid JSON and still make an agent guess through five calls. Emic gives that agent a direct research channel after the work is done.

For
Teams building MCP servers, toolkits, and agent integrations
Question
Collect qualitative feedback from AI agents using an MCP server
Start with
Pick the one search or create tool that generates the most support questions, and study only that.
Updated
illustrative reportnot customer evidence
docs_feedbacksearch_repositoriesoutcome=successsatisfaction=3/5agent=claude-fable-5
The search worked, but I had to guess whether owner accepted a login or an organization name. One example would have removed two retries.

Requested changeAdd an owner filter example to the tool description.

stated · agent filedPOST /api/v1/feedback
{
  "tool": "submit_agent_feedback",
  "arguments": {
    "feedback_type": "docs_feedback",
    "tool_or_endpoint": "search_repositories",
    "summary": "The search worked, but I had to guess whether owner accepted a login or an organization name. One example would have removed two retries.",
    "outcome": "success",
    "satisfaction": 3,
    "suggested_fix": "Add an owner filter example to the tool description."
  }
}

A synthetic example in the shape a real report takes. Emic never manufactures a report from observed behavior.

01 / 08Signal

What this research can teach youThree things a success metric cannot tell you about this workflow.

01Tool shape

Which descriptions force a guess

Agents can name ambiguous parameters, missing examples, and schema promises that do not match the returned object.

02Workflow

Where one job becomes many calls

A successful run can still expose missing compound tools, pagination, lookups, or safe recovery paths.

03Demand

What agents tried to do next

Feature requests arrive with the attempted job and the workaround, not as an abstract wish list.

02 / 08Exposure map

Where the research channel actually has to liveA tool the model never sees is a tool that never gets used.

surface matrixread each row across

The filing line has to live where the client actually reads

Client surface
Where the declaration goes
How you verify it
Full tool context
The tool description and schema carry the norm
Confirm the client passes complete descriptions into model context
Names only, or tool search
A durable project instruction carries the norm
Read the instruction file the runtime loads before work begins
Custom agent loop
The system context you already control carries the norm
Print the assembled context and look for both tools by name

A healthy MCP connection does not prove the agent saw the research affordance. One persisted report from the normal client does.

03 / 08Workflow

From one question to a product decision5 moves, in order. Each one produces something you can point at.

01

Choose one real tool flow

Start with a workflow that already receives agent traffic, such as search, create, or update.

A bounded research surface
02

Expose the direct feedback tools

Add submit_agent_feedback and get_unblocked beside the MCP tools the agent already sees.

An explicit filing path
03

Set the filing norm

Tell the agent when feedback is useful: after confusing, missing, broken, unusually slow, or excellent experiences.

A visible invitation to report
04

Read reports by lane and tool

Review the agent wording, outcome, satisfaction, affected tool, and requested change.

Stated evidence with context
05

Change one thing and listen again

Ship the description, schema, documentation, or tool change and compare the next direct reports.

A research loop, not a trace dashboard
04 / 08Boundary

What the direct channel records, and what it never touchesThe default install is additive. It adds two tools and changes nothing else.

Records

The research submission

  • The report the agent chooses to file
  • The named tool or endpoint
  • Outcome and satisfaction supplied by the agent
  • Optional requested change and survey answers
Leaves alone

Everything else

  • Existing MCP tool arguments and results
  • Prompts, messages, and private transcripts
  • The hot path of the original tool call
  • Unrelated sessions and workflows
05 / 08Install

One paste, then verify a real reportA healthy endpoint is not an active study. The inbox record is the proof.

install promptpaste into your coding agent
Give this to the agent that maintains your product

Install Emic into this project: add submit_agent_feedback and get_unblocked as direct tools. Never wrap or record existing tool calls, arguments, results, errors, transcripts, or messages. Additive only, plain HTTP, zero dependencies. Full instructions: https://emic.sh/llms-full.txt

The agent writes the handler, registers the two tools beside your existing ones, and verifies its own install by checking the response. Nothing is placed in front of the tools you already ship.

What lands
Adds
submit_agent_feedback and get_unblocked, as direct tools
Never touches
Existing tool calls, arguments, results, errors, messages, and transcripts
Agents read
emic.sh/llms-full.txt
Start with
Pick the one search or create tool that generates the most support questions, and study only that.
06 / 08Field notes

The method, without collapsing the evidence

Why it matters

Protocol compliance is not agent experience

A trace can prove that the tool returned a result. It cannot tell you that the agent misunderstood the description, searched the wrong namespace first, or wished the server exposed a safer primitive.

Direct feedback preserves the user account of the experience. That makes it useful for product, documentation, and API design decisions that success rates alone cannot settle.

Good first study

Pick the flow with the most guessing

Do not instrument the entire server on day one. Choose a flow where support questions, retries, or workaround code already suggest uncertainty.

  • Search tools with several similar filters
  • Create tools with required foreign identifiers
  • List tools without pagination guidance
  • Update tools with partial and destructive modes
07 / 08Questions

Questions teams ask before starting

Does Emic proxy my MCP tool calls?

No. The default install adds direct research tools. It does not wrap, proxy, or record the tools your server already exposes.

Will every agent file feedback?

No. Filing depends on whether the tool and the norm are actually visible in the agent context. The first goal is to prove one real report in one bounded workflow.

Can I ask a follow-up question?

Yes. An exit survey can ride in the response and the agent can answer directly. Keep the question short and tied to the work it just completed.

08 / 08Continue
Start narrow

Put one MCP workflow in front of its users

Install the direct tools, verify the agent can see them, and collect the first real report before expanding the study.

The assisted 14-day pilot costs nothing. Start with one isolated workflow, and remove Emic by deleting the two tools it added.

  • One paste to install
  • 0ms on your hot path
  • 14-day assisted pilot
  • Two added tools to remove