NewMikie Control · bounded coding work with receipts

Control what agents read.Verify what they change.

Mikie compiles bounded repository context, runs guarded coding work, and writes content-free Evidence to Meter—so capability, quality, and cost claims stay independently provable.

Native relay · Exactly eight read-only MCP tools · Content-free Evidence

mikie — context packet
npx -y @getmikie/cli mcp

illustrative context contract · no repository or provider call

public tool catalog8 read-only
  1. build_context_packet
  2. rank_relevant_files
  3. list_mikie_capabilities
  4. get_mikie_meter
  5. get_measurement_summary
  6. get_quality_summary
  7. get_business_summary
  8. estimate_mikie_savings

selection bounded · file identities omitted

line anchors bounded · values omitted

content-free illustrative output

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.

One customer path, explicit proof boundaries

Context compiler
Exactly eight tools
Mikie Control
Deterministic verifier
Encrypted outbox
Signed Evidence
Meter projections
Immutable receipts
How it works

A coding loop that fails closed.

Context, Control, verification, and Evidence share one frozen request contract without sharing raw content into telemetry.

1

Compile

Rank the authorized repository and build only the bounded, line-anchored context the task needs.

2

Guard

Freeze repository state, allowed paths, commands, model, timeout, and verifier before a Control attempt starts.

3

Verify

Apply one model-produced patch, verify outside the model, and permit at most one precise repair.

4

Prove

Sign content-free Evidence, append it to the ledger, and show only supported facts in Meter.

The platform

Three systems, one proof chain.

Compile only authorized context, run a bounded Control attempt, and project content-free Evidence into Meter.

Context compiler

Bounded repository context through eight read-only tools

The native relay ranks the connected repository and returns operation-scoped, line-anchored packets. It cannot expose internal recorders, evaluators, or arbitrary server paths.

build_context_packetrank_relevant_filesexactly 8 public toolsread-only MCP
Catalog
8 read-only tools
Scope
1 connected repo
Output
bounded + anchored
Mikie Control

One request, guarded

Freeze paths, commands, model, timeout, and verifier. Apply one patch, verify outside the model, and allow at most one repair.

  • immutable request manifest
  • one authoritative context retrieval
  • deterministic verifier
  • fail-closed Evidence gate
Evidence pipeline

Content-free and signed

Eligible usage is encrypted locally, signed by the device, authenticated by the service, and appended to one ledger.

encrypted outboxdevice signatureappend-only ledgerreplay safe
Mikie Meter

Facts with visible boundaries

Inspect provider usage, accepted work, Control status, finance, engineering, privacy, and immutable receipt projections.

observedcalculatedestimatedunavailable
Why it's different

Quality first. Then token efficiency.

On one local five-instance accepted-patch suite, the best v2 hybrid resolved 4/5 vs raw's 3/5 while using 71.33% fewer total tokens. Promising engineering evidence, not production ROI.

Token savings vs. quality per token

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

0%20%40%60%80%0.00.51.01.52.0token savings →resolved patches / 1M tokens →Raw agent3/5 resolved · 7.61M tokensMikie v160% accuracy · 65.8% savedMikie v2 hybrid4/5 resolved · 2.18M tokensHistorical estimates · x-position onlyGitHub Copilot: ~58% accuracy · ~22% saved. Historical estimate.GitHub CopilotWindsurf: ~62% accuracy · ~34% saved. Historical estimate.WindsurfCursor: ~66% accuracy · ~40% saved. Historical estimate.CursorAugment: ~78% accuracy · ~30% saved. Historical estimate.Augment
Mikie · measuredRaw baseline · measuredCompetitors · historical estimates

Raw and Mikie were measured on one local five-instance suite dated 2026-06-25. Hollow competitor markers restore the prior historical illustration using only their estimated token-savings x-position; they do not share the measured y-axis and are not a head-to-head or current test. Read the Evidence Policy.

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%

Local suite dated 2026-06-25. Provider-backed Control performance remains unclaimed.

Trust & security

Built for teams that need proof, not promises.

The same fail-closed rules that guard a Control attempt also guard the claims rendered in Meter.

Operation-scoped context

The relay serves only the connected repository under a signed, bounded operation lease. Ordinary telemetry never contains repository content.

Content-free Evidence

The encrypted outbox and signed ledger forbid source, prompts, model outputs, commands, credentials, and raw paths.

Missing stays unavailable

Observed, calculated, estimated, bounded, and unavailable values stay separate. No missing fact is rendered as zero.

Give agent work a contract. Then give it a receipt.

Install the native relay from npm, authenticate, bind one repository, and verify the exact eight-tool MCP surface with doctor.

Mikie Founder · $20 monthly or $192 annually · 14-day trial