Intent Surfaces™

What If Context Engineering
Was Like Code?

What If Context Engineering
Was Like Code?

Capture once. Invoke deterministically. Share across your team.

Capture once. Invoke deterministically. Share across your team.

Developer

Agent

mq

MCP

Turn your MCP exploration into reusable flows.

First run costs ~30 seconds + 8K tokens. Every run after: 2 seconds + 250 tokens.

Search before building. Export to share. Works with any agent framework.

Turn your MCP exploration into reusable flows.

First run costs ~30 seconds + 8K tokens. Every run after: 2 seconds + 250 tokens.

Search before building. Export to share. Works with any agent framework.

Example

Install

Quick Start

Download on github

02 

Problem

The Pain You Know Too Well

The Pain You Know Too Well

Pull Request

You're using Cursor's AI agent to automate your Gmail context. The agent makes a request. Gets it wrong. You add a comment:

You're using Cursor's AI agent to automate your Gmail context. The agent makes a request. Gets it wrong. You add a comment:

|

Agent ignores it. You try again:



Agent ignores it. You try again:



|

Still ignored. You resort to:

|

Maybe it works. Maybe it doesn't. You have zero control.

Cursor → Claude

Cursor → Claude

Switch from Cursor to Claude Desktop? Breaks. The agent uses different tools in a different sandbox.

Switch from Cursor to Claude Desktop? Breaks. The agent uses different tools in a different sandbox.

GPT 4 → Claude

GPT 4 → Claude

Switch from GPT-4 to Claude? Breaks. Different model, different interpretation, different execution.

Switch from GPT-4 to Claude? Breaks. Different model, different interpretation, different execution.

You're fighting stochastic execution with natural language.

  • The agent framework controls the tools.

  • LLM models control the interpretation (you don't)

  • Prompts are your only lever (unreliable)

This is the Control Problem.

Result: Every execution is a gamble.

Switch models? Breaks. Switch agents? Breaks. Switch frameworks? Breaks.

You need determinism. You're getting probability.

“ This is the Control Problem. mq solves it. ”

03 

Solution

What if your agent's MCP exploration could be observed, guided, and reified into deterministic flows?

What if your agent's MCP exploration could be observed, guided, and reified into deterministic flows?

Not by writing code. By doing what you already do: exploring, querying, iterating.

Context is engineered by the agent executing the task through mq.

You do what you already do: express intent in natural language.

You do what you already do: express intent in natural language.

mq does the heavy lifting: assesses call flow, guides execution, converts successful patterns into scripts for future use.

Needs

Then, when you (or an agent) needs that context again:

No retries

No mistakes

No rediscovery

Deterministic invocation from natural language

Context

Your interactive MCP session becomes a first-class artifact: cached, compiled, discoverable, shareable.

Context engineering becomes code without writing it.

Context engineering becomes code without writing it.

The Complete Solution:
One Example Shows Everything

mq is a self-learning lightweight layer. No LLM. No prompts. No neural networks. Just old-fashioned intelligence: observation, feedback, adaptation the way humans learn, form habits, and build procedural memory.

Monday

Jill Explores

Tuesday

Bob Reuses

Wednesday

Nancy

Monday: First Jill Explores Gmail Context

When working with mq, the agent follows three rules:

  1. mq guidance modiqo essential (read completely in order)

  2. Discovery FIRST: mq flow search before building anything

  3. After success: ask "export? (yes/no)" before exporting

Jill says: "Fetch my most recent emails"

Agent (guided by mq):

|

Agent made 3 query mistakes, corrected instantly from inline hints over cached response (no HTTP retries)

  • mq guided toward correct API usage through hints

  • mq detected parallelization opportunity automatically

  • Successful sequence saved, errors filtered out

  • Resulting flow is deterministic: same ops every time

Time

30 seconds

Tokens

8,400

Outcomes

Agent learned a new skill

The Complete Solution:
One Example Shows Everything

mq is a self-learning lightweight layer. No LLM. No prompts. No neural networks. Just old-fashioned intelligence: observation, feedback, adaptation the way humans learn, form habits, and build procedural memory.

Monday

Jill Explores

Tuesday

Bob Reuses

Wednesday

Nancy

Monday: First Jill Explores Gmail Context

When working with mq, the agent follows three rules:

  1. mq guidance modiqo essential (read completely in order)

  2. Discovery FIRST: mq flow search before building anything

  3. After success: ask "export? (yes/no)" before exporting

Jill says: "Fetch my most recent emails"

Agent (guided by mq):

|

Agent made 3 query mistakes, corrected instantly from inline hints over cached response (no HTTP retries)

  • mq guided toward correct API usage through hints

  • mq detected parallelization opportunity automatically

  • Successful sequence saved, errors filtered out

  • Resulting flow is deterministic: same ops every time

Time

30 seconds

Tokens

8,400

Outcomes

Agent learned a new skill

That's the difference:

First time:

15 minutes, 8,400 tokens

Learns

Every time:

2 seconds, 250 tokens

Executes

Discovery

Prevents re-learning what's known

Caching

Enables instant error correction without retries

Reification

Transforms explorations into procedures

Three Features That Make Context Engineering Like Code

Watch how all three features work together

/01

Command Language

/02

Flow Discovery

/03

Response Caching

BONUS

/04

Reification

/05

Adaptive Forgetting

/06

Self-Reflection

A Lightweight Command Language for MCP (Not Prompts)

Instead of prompting agents to use MCP, agents now have commands to orchestrate it—and learn through iteration.

MODELS

• GPT-4

• Claude

• Gemini

• o1

Same Language

AGENTS

• Cursor
• Claude Desktop

• Cline

• Aider
Same Command

FRAMEWORKS

• claude code

• cursor /agents

• AutoGPT

• Custom

Same Syntax

|

No CAPITALIZED comments. No hoping the agent understands. No framework lock-in.

And here's the key: Agents don't need training to use mq. They learn it through iteration and self-reflection—the same way they learned curl, jq, ls, and cp. Except this tool watches them learn and captures successful patterns as reusable procedures.

Agents don't write prompts. They execute commands and learn from results.

/01

Command Language

/02

Flow Discovery

/03

Response Caching

BONUS

/04

Reification

/05

Adaptive Forgetting

/06

Self-Reflection

A Lightweight Command Language for MCP (Not Prompts)

Instead of prompting agents to use MCP, agents now have commands to orchestrate it—and learn through iteration.

MODELS

• GPT-4

• Claude

• Gemini

• o1

Same Language

AGENTS

• Cursor
• Claude Desktop

• Cline

• Aider
Same Command

FRAMEWORKS

• claude code

• cursor /agents

• AutoGPT

• Custom

Same Syntax

|

No CAPITALIZED comments. No hoping the agent understands. No framework lock-in.

And here's the key: Agents don't need training to use mq. They learn it through iteration and self-reflection—the same way they learned curl, jq, ls, and cp. Except this tool watches them learn and captures successful patterns as reusable procedures.

Agents don't write prompts. They execute commands and learn from results.

Built for MCP Contexts

Built for MCP Contexts

Session Management

|

-s flag handles all session headers

Template Variables

|

Extract once, use everywhere

Native Query Engine

|

jq-compatible <100μs query time

OAuth Validation

|

Flows validate tokens automatically

Technical Excellence

Technical Excellence

RUST

8.1 MB binary, zero runtime dependencies

8.1 MB binary, zero runtime dependencies

SPEED

Query: <100μs

Async runtime

Deterministic

execution

Query: <100μs

Async runtime

Deterministic

execution

Smart

DAG scheduling

Auto depends

detection

DAG scheduling

Auto depends

detection

MEMORY

ABinary metrics <500 bytes
Drift detect

ABinary metrics <500 bytes
Drift detect

Installation

Installation

Copy

Copied

Copy

Copied

curl -fsSL https://github.com/modiqo/mq-releases/releases/latest/download/install.sh | bash

Linux (x86_64/ARM64/musl)

macOS (Intel/Apple Silicon)

Windows

Quick Start

1

Search procedural memory first

|

2

If not found, agent explores (mq observes and guides)

|

3

Successful sequence becomes memory

|

4

Future invocations = habit execution

|

Why This Matters

What if context engineering was like code?

What if context engineering was like code?

HUMANS

Express intent in natural language

Express intent in natural language

AGENTS

Learn through structured exploration

Form habits

Share memory

Learn through structured exploration

Form habits

Share memory

MQ

Observes

Guides

Captures

Detects when to re-learn

Observes

Guides

Captures

Detects when to re-learn

mq demonstrates that MCP contexts can be first-class artifacts: learned through iteration, cached for efficiency, reified as procedures, shared across agents, adaptively forgotten when stale.

Context engineering is code. Agents write it through exploration. mq captures it.

For MCP Server Providers: Bootstrap Your Adoption

Here's a strategy that changes the game:

If you provide an MCP server, create common flows for your API and distribute them with your server.

|

Why this matters:

When customers install your MCP server, their agents start with procedural memory. No exploration phase. No trial-and-error. No token waste discovering your API patterns.

First interaction:

|

Your customers get instant value. Their agents invoke your API correctly from day one. You control the best practices through distributed procedural memory.

This is how MCP servers should ship: with habits, not just capabilities.

Open Source & MIT Licensed

Release

Download

Github

Star Repository

Paper

Download & Read

Discord

Join Now

Technical Deep Dive

Technical Deep Dive

Version 0.4.3

MIT Licensed • Pure Rust • No Neural Nets • Self-Learning Execution Layer

Made with love for MCP, curl, and jq by Modiqo