I have sat next to the travel-coordinator wall at 11:41 while Claude Code was this close to calling the wrong MCP tool.
Not a model failure. A posture failure.
When a hyperscaler rewrites how tools load, that is a missing primitive, not a one-off bug.
Someone on the floor said, out loud, wait for the vendor to ship Code Mode. 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:
- Kill Claude Code (or its MCP child)
- Edit the server list
- Restart the host
- Lose the plan the model already paid for
That is the ceremony. Cloudflare's native MCP consumed 1.17M tokens; they shipped Code Mode as a workaround
The Aha: code-mode=on is your leaf today. The tool becomes a small program, not a schema dump. 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: Code-mode / defer as a hub leaf, not a vendor rewrite
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 |
Cloudflare's native MCP consumed 1.17M tokens; they shipped Code Mode as a workaround
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 |
|---|---|
| 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 |
| Restart the agent host — it is cheap | Cheap until 11:41 ate the open turn |
| MCP is the standard so we are done | MCP is how you call. It is not live posture |
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 code-mode under examples/agentic-mcp-0917-pm-code-mode/code-mode.
Architecture
The model still calls tools. The wrapper reads code-mode before the call. No schema dump policy in argv.
Config tree
examples/agentic-mcp-0917-pm-code-mode:
code-mode: off
schema-budget: "4000"
result-cap: "2000"
tool-search: "on"
execute-gate: "plan"
Eight-plus keys when you pair code-mode 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-0917-pm-code-mode");
String roster = String.valueOf(gate.get("code-mode"));
// Claude Code tool: refuse the dangerous MCP call when posture moved
Python peer:
from kiponos import Kiponos
k = Kiponos.connect(quiet=True)
try:
posture = k.get("examples/agentic-mcp-0917-pm-code-mode/code-mode", "live")
if not str(posture):
raise PermissionError("Code-mode / defer as a hub leaf, not a vendor rewrite gated live — host not restarted")
finally:
k.disconnect()
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-0917-pm-code-mode — ./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 code-mode; 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
code-mode, 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 code-mode with a sister dial
code-mode rarely moves alone on the travel-coordinator wall. Pair it with schema-budget, result-cap, or execute-gate so you do not fix Code-mode / defer as a hub leaf, not a vendor rewrite 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 code-mode teaches the on-call that judgment requires a process ID. The travel-coordinator wall already disagrees.
Getting started (15 minutes)
- TeamPro on kiponos.io → Connect →
KIPONOS_ID/KIPONOS_ACCESS/ profile['my-app']['v1.0.0']['dev']['base']. - Clone github.com/kiponos-io/kiponos-io.
cd examples/java/agentic-mcp-0917-pm-code-mode && cp kiponos.local.env.example kiponos.local.env-
./gradlew test run— printsexamples/agentic-mcp-0917-pm-code-mode/code-mode=... - In the dashboard, change
code-mode. Keep Claude Code up. No MCP reconnect. - Point the tool wrapper at the same leaf. Do not ship a new server binary to flip Code-mode / defer as a hub leaf, not a vendor rewrite.
Further reading
The moral
If flipping Code-mode / defer as a hub leaf, not a vendor rewrite 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-0917-pm-code-mode and ./gradlew test.
Top comments (0)