DEV Community

Cover image for 16% Task Completion With a Fat MCP Toolset Is Not a Model Failure
Moshe Avdiel
Moshe Avdiel

Posted on Originally published at github.com

16% Task Completion With a Fat MCP Toolset Is Not a Model Failure

I have sat next to the travel-coordinator wall at 00:40 while Claude Code was this close to calling the wrong MCP tool.

Not a model failure. A posture failure.

Unacceptable in production. The fix people try is a bigger model. The fix is a smaller live set.

Someone on the floor said, out loud, wait for the next frontier, MCP will work. Claude Code had started earlier. The tool list was frozen in config.toml / .mcp.json / a skill file. The only “safe” move anyone trusted was:

  1. Kill Claude Code (or its MCP child)
  2. Edit the server list
  3. Restart the host
  4. Lose the plan the model already paid for

That is the ceremony. Tau-bench Claude 3.7 ~16% with MCP tools on airline scenarios

The Aha: Completion % tracks live roster size more than IQ. Shrink tools-allow now. Real-time nested state is now a primitive — as native as a hashmap — via Kiponos.io. The host stays up. The leaf moves. The next tool call already obeys.

The problem: Airline-booking MCP benches fail because the space is too big

Agentic programming in 2026 has a public fight: MCP vs CLI vs Skills. The likes on MCP and Grok Build posts are the tell. The industry is stuck on how to expose tools.

The missing primitive is live nested state:

Where the roster hid What you restart What you lose
MCP schema dump at session start The host, sometimes the child npx 16–72% of the window before the user speaks
~/.grok/config.toml / Cursor mcp.json Claude Code The open turn
SKILL.md / AGENTS.md The next turn, sometimes the host Context you already bought
“Just reconnect the server” The MCP session Tool-space you finally had right

Tau-bench Claude 3.7 ~16% with MCP tools on airline scenarios

That is not a protocol bug you wait out. That is a process that cannot get() a tree that moved.

What teams believe

Belief Production
Skills will progressive-disclose it Skills are files. Files do not fan out
A bigger context window fixes tool overload 1M tokens still rot; selection still drops
Put the SDK in the SPA so the tile matches Connect tokens do not belong in a browser
Feature flags cover MCP write Flags are another product, another delay

What is Kiponos here

Kiponos is a live nested config hub. Dashboard edit → WebSocket delta → in-memory tree on every connected process. Hot-path get() is a local map lookup. One WebSocket per process lifetime.

It is not an MCP server. It is not a Skill. It is the control plane those things do not ship: which tools may run, how fat a result may be, whether execute is allowed, whether a child inherits posture — this turn.

Profile shape: ['my-app']['v1.0.0']['dev']['base']. The leaf in this story is tools-allow under examples/agentic-mcp-0926-am-tau-bench/tools-allow.

Architecture

Architecture diagram

The model still calls tools. The wrapper reads tools-allow before the call. No schema dump policy in argv.

Config tree 

examples/agentic-mcp-0926-am-tau-bench:
  tools-allow: search,read
  schema-budget: "4000"
  result-cap: "2000"
  tool-search: "on"
  execute-gate: "plan"
Enter fullscreen mode Exit fullscreen mode

Eight-plus keys when you pair tools-allow with sister dials (budget, mute, pause). They move together on the travel-coordinator wall.

Integration

Java tool wrapper (Claude Code just calls it):

import io.kiponos.Kiponos;
import io.kiponos.Folder;

Folder gate = Kiponos.createForCurrentTeam()
    .getFolder("examples/agentic-mcp-0926-am-tau-bench");
String roster = String.valueOf(gate.get("tools-allow"));
// Claude Code tool: refuse the dangerous MCP call when posture moved
Enter fullscreen mode Exit fullscreen mode

Python peer:

from kiponos import Kiponos

k = Kiponos.connect(quiet=True)
try:
    posture = k.get("examples/agentic-mcp-0926-am-tau-bench/tools-allow", "live")
    if not str(posture):
        raise PermissionError("Airline-booking MCP benches fail because the space is too big gated live — host not restarted")
finally:
    k.disconnect()
Enter fullscreen mode Exit fullscreen mode

React/Angular: Node BFF @kiponos/react/server / @kiponos/angular/server. SPA never holds Connect tokens.

Runnable mesh: https://github.com/kiponos-io/kiponos-io/tree/master/examples/java/agentic-mcp-0926-am-tau-bench — ./gradlew test run.

Real scenarios

Event Without Kiponos With Kiponos
Three MCP servers at boot 72% of the window gone Live roster / schema-budget; dump shrinks
User says “no writes” mid-turn Restart Claude Code Set tools-allow; next call already denies
Second host started earlier Paste the json into chat Both get() the same leaf
Subagent spawned Frozen argv copy Child get()s inherit-posture
Tool returns 500k tokens Context rot, doom loop result-cap in the wrapper

Performance (this path)

  • Wrapper get() is in-process after bootstrap.
  • One WebSocket per Claude Code lifetime — not per MCP call.
  • A dashboard edit is a delta of tools-allow, not a config.toml reload.
  • You do not spend model tokens to “please restart Claude Code.”
  • Schema tax is optional once the live set is small.

Compare to alternatives

Approach Honest fit Why it still restarts
More MCP servers More integrations More schema, worse selection
Skills progressive disclosure Good docs, lazy load Not a multi-host bus
Claude Code tool search Better client Other hosts still dump
Redis poll in the tool Shared, extra RTT You invented a hub with worse UX
Feature-flag SaaS Product experiments Rarely session-safe for MCP
Kill -9 the npx child Janitor The turn still died

When not to use Kiponos

Situation Why
Tool schema itself changed (new argument) That is a code/Claude Code restart
Secret rotation of Connect / MCP OAuth tokens Credentials are not live knobs
One-off local script, no peers A hub is overkill
Browser-only “SDK in the SPA” Forbidden — tokens leak or defaults lie

Pair tools-allow with a sister dial

tools-allow rarely moves alone on the travel-coordinator wall. Pair it with schema-budget, result-cap, or execute-gate so you do not fix Airline-booking MCP benches fail because the space is too big by inventing a second incident.

Why Claude Code is the wrong restart target

Claude Code is good at calling tools. It is not a control plane. Killing it to flip tools-allow teaches the on-call that judgment requires a process ID. The travel-coordinator wall already disagrees.

Getting started (15 minutes)

  1. TeamPro on kiponos.io → Connect → KIPONOS_ID / KIPONOS_ACCESS / profile ['my-app']['v1.0.0']['dev']['base'].
  2. Clone github.com/kiponos-io/kiponos-io.
  3. cd examples/java/agentic-mcp-0926-am-tau-bench && cp kiponos.local.env.example kiponos.local.env
  4. ./gradlew test run — prints examples/agentic-mcp-0926-am-tau-bench/tools-allow=...
  5. In the dashboard, change tools-allow. Keep Claude Code up. No MCP reconnect.
  6. Point the tool wrapper at the same leaf. Do not ship a new server binary to flip Airline-booking MCP benches fail because the space is too big.

Further reading

The moral

If flipping Airline-booking MCP benches fail because the space is too big requires restarting Claude Code or reconnecting MCP, you do not have agentic control. You have a hope with a process ID.

MCP, Skills, and tools are how agents act. Kiponos is the live primitive they do not ship — so the travel-coordinator wall can change its mind without killing the session.

How to try: examples/java/agentic-mcp-0926-am-tau-bench and ./gradlew test.

Top comments (0)