Ensemble mcp

Twelve chapters, one story.

From the harness concept to live installation.

1. 01The Problem — why AI coding tools are wasteful todayContext
2. 02What is an Agent Harness? — the model + harness equationFraming
3. 03What is Ensemble mcp? — the harness infrastructure layerThesis
4. 0419 Harness Tools — memory, drift, routing, indexing, skills, compressionSurface
5. 05Zero-LLM Architecture — how it's efficientUnder
6. 06Token & Cost Reduction — five mechanisms that save moneyEconomics
7. 07Live Workflow — what happens behind the scenesBehavior
8. 08Technical Architecture — under the hoodSystem
9. 09Dashboard — observability & visualizationUI
10. 10Installation — one command to get startedOnboarding
11. 11Future Roadmap — what's nextHorizon
12. 12Key Takeaways — the harness pillarsClose

Every session starts from zero.

A harness is everything around the model.

The harness stack.

Agent harnesses have layers. Different providers handle different layers.

Nineteen primitives, eight categories, one envelope.

Pattern memory.

Agents remember what worked before — the harness enables continual learning.

Drift detection.

A harness primitive that catches agents going off-task before damage is done.

Task:    "Add user authentication"
Changes: auth controllers, login views

→ score:   0.12
→ verdict: "aligned"

Task:    "Add user authentication"
Changes: blog system, payment gateway

→ score:   0.78
→ verdict: "significant_drift"
→ Agent warned before continuing

Smart model routing.

Stop paying premium prices for simple tasks.

Codebase indexing.

Stop re-exploring the same files on every single run.

Skill intelligence — progressive disclosure.

A harness primitive that prevents context rot by loading only what's needed.

Resume, and say less.

Everything runs locally.

ensemble-mcp makes zero external API calls. All intelligence runs locally on your machine.

The economics of running nothing.

Pattern memory & indexing.

Every session:
  Agent reads 30 pattern entries
  → ~8,000 tokens consumed

Every session:
  Agent queries top-3 patterns
  → ~800 tokens consumed

Savings: ~90% → ~$8.10/dev/mo

Agent explores:
  glob("**/*.php")
  grep("class.*Controller")
  read file by file…

→ ~4-6K tokens per cycle

Agent queries:
  project_query(
    query="TodoController",
    file_types=["php"]
  )

→ ~700 tokens · ~$4-6/mo

Tier the task,
save the budget.

~30–60% cost savings by using the right tier for each task.

Every task → Claude Opus
  $15/M input, $75/M output

Including:
  ✗ Typo fixes   → Opus
  ✗ Test runs    → Opus
  ✗ Git commits  → Opus

Complex tasks → Opus   $15/M
Simple tasks  → Sonnet $3/M   · 80% cheaper
Trivial tasks → Haiku  $0.25/M · 98% cheaper

Compression & caching.

The cumulative effect.

The flywheel in motion.

Four layers. One contract.

Six harnesses. One intelligence layer.

One envelope. Every tool.

{
  "ok": true,
  "data": { "...payload..." },
  "error": null,
  "meta": {
    "duration_ms": 12,
    "source": "sqlite",
    "confidence": "exact"
  }
}

Security.

Secrets never leave the machine. Boundaries are explicit.

• Secret redaction

  9 regex patterns scan all text before storage — AWS keys, Bearer tokens, API keys, GitHub tokens, passwords — all replaced with [REDACTED].

• Trust boundaries

  Data is classified by source — local_state (trusted) · client_input (validated) · filesystem_scan (read-only).

• Local-only dashboard

  Binds to 127.0.0.1, never exposed to the network.

• DOMPurify sanitization

  All rendered markdown is XSS-sanitized before it ever hits the DOM.

• Destructive operations

  Require an explicit confirm=true flag. No accidents.

A window into the living memory.

Install · register · launch.

Roadmap.

Right-sized for today.

With a clear map of where it grows next.

Memory · Self-Verification · Orchestration · Context Rot · Progressive Disclosure.