NewMCP-native context for Codex & Claude Code

The context layer your AI coding agents are missing.

Mikie gives Codex, Claude Code, Cursor, and any MCP client compact, line-anchored code context — then measures the tokens, cost, and retries you save on every single turn.

Local-first · No source code stored · Works with your existing agents

mikie — context packet
0.00×
Fewer total tokens than raw on our accepted-patch suite
7.61M → 2.18M total tokens · 71.33% reduction
0 / 5
Accepted patches resolved — one more than raw
vs. raw 3 / 5 on the same 5-instance suite
0%
Lower cost per resolved patch vs raw
$0.90 vs $2.69 · GPT-5.5 standard pricing

Measured on a local 5-instance accepted-patch / SWE-style suite (2026-06-25). Promising engineering evidence, not a broad benchmark or production-ROI claim. Token savings are reported separately from quality.

Works with the agents you already use

Claude Code
Codex CLI
Codex IDE
Cursor
GitHub Copilot
Windsurf
Cline
Zed
Any MCP client
The problem

AI coding is getting expensive and opaque.

Agents are powerful, but they read too much, retry too often, and prove too little. Four signals show up in almost every codebase.

01

Context bloat

Agents stuff entire repos into the prompt. You pay for hundreds of thousands of tokens the model never needed to read.

02

Runaway token cost

Max plans, Pro plans, per-token APIs. Costs climb every month and nobody can point to where the spend actually goes.

03

Blind broad scans

Re-reading files, duplicate searches, and read-before-edit loops burn turns and tokens without moving the task forward.

04

Unverifiable savings

Every tool claims it saves money. Almost none can show you a measurement row that proves the number on a real task.

How it works

Context that closes the loop.

Mikie owns the full cycle — from ranking your repo to proving what each turn actually cost.

1

Rank

Index the repo locally and score every file against the prompt — exact paths, symbols, and tests first.

2

Pack

Build a compact, line-anchored context packet — just the spans that matter, never the whole repo.

3

Inject

Deliver the packet to Codex, Claude Code, or any MCP client before the model turn — no wasted discovery.

4

Measure

Record tokens, cost, cache hits, retries, and outcomes to the ledger — proof on every turn.

The platform

Three engines, one context OS.

A routing layer for your agents — not another wrapper. Rank and pack context, record every turn, and keep your claims honest.

Context Engine

Exact-path, line-anchored context — built locally

Rank thousands of files in milliseconds, select only the spans the task needs, and inject them before the model turn. No uploads, no whole-repo dumps.

build_context_packetrank_relevant_filesline anchorsexact-path match
Ranking
1,284 → 4 files
Packet
1,910 tokens
Latency
< 200 ms
Measurement Ledger

Every turn, recorded

Tokens, cache hits, retries, patches, and test outcomes — written to a durable ledger. Never your raw outputs.

  • input / output / cached tokens
  • tool calls & retry counts
  • patch accepted · tests passed
  • estimated cost saved
Claim Boundaries

Honesty, enforced by the system

Savings, accuracy, and retry reduction are separate gates. If the evidence isn't there, the claim stays blocked — automatically.

raw_semantic_winretry_reductionroimodel_agnostic
Waste Reports

See where the tokens go

Useful work vs. duplicate searches, broad scans avoided, and reruns caused by missing context — broken down per agent and repo.

useful tool callsduplicate searchesscans avoidedcost per task
The outcome

More patches resolved. A fraction of the tokens.

On a local 5-instance accepted-patch suite, the best Mikie v2 hybrid resolved 4/5 vs raw's 3/5 while cutting total tokens from 7.61M to 2.18M — a 71.33% reduction (3.49×). Quality is scored on a separate gate.

Total tokens to resolve the same 5 tasks↓ 71%
Raw agent — no context tool3/5 resolved0
Mikie v1 (Sprint)3/5 · 2.93× fewer0
Mikie v2 hybrid4/5 · 3.49× fewer0

Measured 2026-06-25 on a local 5-instance accepted-patch / SWE-style suite. Lower is better. Token savings reported independently of patch quality.

Why it's different

More savings and more quality per token.

Mikie doesn't trade answer quality for a smaller prompt. By delivering the right local contracts up front, the v2 hybrid lifted resolved patches per 1,000 tokens by 4.65× while cutting total spend.

Token savings vs. quality per 1k tokens

Top-right is better — high savings with high quality density.

token savings →quality / 1k →RawMikie v1Mikie v2
Raw vs Mikie v2 hybrid
MetricRawv2 hybrid
Total tokens7.61M2.18M
Commands run19539
Failed commands4911
Cost / resolved patch$2.69$0.90
Net total-token reduction
−71.33%
Trust & security

Built for teams that need proof, not promises.

Mikie is a measurement system first. The discipline that keeps our own claims honest is the same discipline we give you.

Local-first by default

Repos are indexed and packets are built on your machine. Nothing is uploaded to build context.

No raw outputs stored

The ledger records tokens, cost, retries, and outcomes — never your source code or model responses.

MCP standard

Mikie speaks the Model Context Protocol. No lock-in, works with any compatible client.

Evidence or it's blocked

Every savings claim traces to a measurement row. Unproven claims stay blocked, not inferred.

Separate quality gate

Token savings and answer quality are measured independently — we never trade one for the other.

Your data, your control

Self-serve plans, exportable reports, and role-based access. Delete your measurement history anytime.

Stop wrestling with MCP setup. Start using Mikie.

Install Mikie and connect your coding client through secure remote MCP access with a guided setup.

$39 monthly or $384 annually · Codex, Claude Code, and generic MCP