AgentStatus raises a new roundChatGPT MCP events are a real checkiMessage is a host now

Agents inside Claude CodeCursorCodexuse your MCP tools.

When your tool breaks, the agent picks another one and moves on. Now, you find out before this ever happens.

Cover your blind spots

44K tools

Agents don't tell you the tool is broken.They pick another one.

You cannot prioritize the work.

Your tickets and your 200s do not say the agent finished the job. We tell you which tools did the work on hosts you do not run, so the roadmap is not a guess.

What to ship

fetch

Page title and first paragraph came back.

Ship this
ClaudeCursorVS Code

search

Came back. Not the work.

Hold
ClaudeCursorVS Code

write

No run.

Don't ship
ClaudeCursorVS Code
Cover your blind spots

Just the MCP URL. No SDK needed.

Paste the server URL and auth. No SDK and no code change. We start the hosts from there.

Connect an MCP server

URL, auth, and the hosts we will connect as

Ready to connect

URL

https://mcp.example.com

Check every

15 minutes

Hosts we start

Claude CodeClaude Code
CursorCursor
CodexCodex
Muse CodeMuse Code
GooseGoose
ClineCline
OpenCodeOpenCode
DevinDevin
iMessageiMessage
GitHub CopilotGitHub CopilotSoon
MistralMistralSoon
Qwen CodeQwen CodeSoon
AgAntigravitySoon
KiKiroSoon
GkGrokSoon

Auth

Bearer token

Know what each host did with a real request.

Same server. Different hosts. We start each one, send the same requests, and keep what it called.

Your MCP Tools

Claude Code

Claude Code

Called a tool

Version

v2.1

Called

search_docs

Cursor

Cursor

Called a tool

Version

v1.7

Called

search_docs

Codex

Codex

Called a tool

Version

v0.98

Called

list_invoices

Muse Code

Muse Code

Did not finish

Version

v1.4.1

Called

No call

Know when agents can't use your tool because the guide is wrong.

tools/list is the guide the agent follows. If the names, types, or fields are wrong, the agent never completes the call.

The published guide

list-records

What the agent could use

search-records

The published guide

list-inboxes

What the agent could use

empty

The published guide

email

What the agent could use

string

The published guide

mail_header_id

What the agent could use

mail_id

The published guide

optional

What the agent could use

required

Prove which tools did the job.

Listed is not called. Called is not the work. We send the requests on those same hosts and score whether the tool did the job.

Per tool

8:029:1410:0211:4012:171:052:223:084:41
fetch
search
write
Did the jobCame back, not provenNo run

fetch

Did the job

Answer came back

No hidden error inside a 200

No auth or permission errors

Non-trivial payload

The page title and the first paragraph came back

Catch the step that killed the path.

Search, then fetch, then write. We walk that chain on the same MCP and stop where it breaks.

Do multi-step tool chains hold?

Connected workflow evidence

Enforcement open

  1. 1

    Detected dependency

    SEsearch→needs id→FEfetch→needs id→WRwrite
  2. 2

    Executed workflow

    Failed
    SEsearchpassed→FEfetchfailed→WRwritenot reached
  3. 3

    Recommended for improved performance

    Install

    pip install agentstatus-mcpjobgraph

    docs

    Use this path as the first graph to enforce.

Pin what broke after you shipped.

Compare catalogs, initialize results, and call verdicts across hosts, runs, and releases.

Day-by-day · Last 9 days (UTC)

8/218/228/238/248/258/268/278/288/29
Overall pass rate99%98%97%100%100%99%99%100%100%
Reach OK99%98%97%100%100%99%99%100%100%
Response time2.0s967ms1.1s–1.8s4.2s2.9s3.4s–
Job completed100%–98%100%–97%100%100%–
Policy followed100%100%–100%100%92%–100%100%
Tool execution–99%100%––100%98%–100%
Required action100%–97%100%–100%99%–100%

Cooler = better for that row · Hotter = worse · Empty = no data that day

CoolerHotter

Your biggest blind spots sit outside your logs,inside the host that called your tool.

We start the host, send the request a person would type, and keep the version and the tool it called.

Cover your blind spots