AI-Agent Framework

Cortex Harness

Pick a vertical. Edit one file. Ship a memory-backed agent.

Fork it, build it, run it — on your own machine.

  harness.config.yaml   ←  the one file you edit
        ↓
   [ COMPILER ]  ←  pick a vertical + tool loader
        ↓
   build/<you>/
     brain/        rendered lanes (identity, rules, tools)
     agentb.yaml   memory backend config (persona = rule-set)
     mcp.json      MCP registration with your enabled tools
     start.sh      launch the memory backend
        ↓
   Mnemo Cortex  — persistent memory · facts · brain lanes

Mnemo Cortex gives an agent memory. Cortex Harness is the layer on top that turns that memory into a product you can ship for a specific business — without rewiring anything. Set your business name, pick a vertical, and build. You get an agent with persistent memory, a business-specific rule-set, and a tool-set scoped to what you turned on.

View on GitHub Install

The one file you edit

Name it. Pick a vertical. Build.

Everything else is derived. Blank fields fall back to the vertical’s recommended defaults.

business:
  name: "Acme Co."
  vertical: small-business      # python -m harness list
  agent_id: acme-assistant
tools:
  profile: standard             # minimal | standard | full
  enable: []
  disable: []
rules:
  persona: ""                   # blank = the vertical's recommended rule-set
memory:
  isolation: isolated           # isolated | shared | hybrid

Then the build-and-run loop:

python -m harness list      # see the verticals
python -m harness doctor    # check your environment
python -m harness plan      # preview enabled vs. dropped tools (no writes)
python -m harness build     # compile to build/<you>/
python -m harness run       # build, then launch the memory backend

Verticals shipped

Five businesses in the box. Adding your own is a folder.

A vertical is just a folder: a rule-set, brain templates, and a recommended persona. Everything in _base/ is inherited, so a new one only states what’s different.

Small Business

Customers, quotes, scheduling, follow-ups. Persona: operator — concise, fact-checked, approval-gated.

Ecommerce

Catalog, inventory, orders, support. Persona: merchant — SKUs, prices, and stock as hard facts.

Art Studio

Commissions, briefs, inventory of works — plus a creative muse. Persona: studio — speculative muse, exact business.

Writing Partner

Drafting with a story bible and continuity tracking. Persona: editor — generative, but ruthless on canon.

Day-Trading Desk

A trading journal & research desk. Persona: risk_desk — analysis only, enforces your own rules.

Analysis only · never executes

The day-trading desk never places orders, moves money, or gives financial advice — enforced in its rule-set and brain lanes, and structurally: there is no execution tool anywhere in the catalog. Its job is to hold the trader to their own written rules and keep an honest record.

How it works

Three things every business customizes — made first-class.

Without the harness you wire these by hand for every business. With it, you edit one file and build.

🧠

Brain templates

Identity, playbooks, and reference lanes — markdown with {{placeholders}} filled in from your config.

📏

Rule sets

How cautious the agent is, what it may speculate on, what needs sign-off. Compiles into a Mnemo Cortex persona.

🔧

Tool loader

Pick a profile, then enable/disable. A tool whose prerequisite is missing is dropped — with a reason — not silently exposed.

Built on Mnemo Cortex

The harness compiles. Mnemo remembers.

Cortex Harness isn’t a memory engine — it’s the framework that configures one. Mnemo Cortex is a cloned dependency (kept out of the repo, not vendored); a build compiles its backend config and harness run launches it. Local SQLite by default, your data on your machine.

About Mnemo Cortex →

Install

Fork, clone the memory layer, build.

git clone https://github.com/GuyMannDude/cortex-harness.git
cd cortex-harness
git clone https://github.com/GuyMannDude/mnemo-cortex.git ./mnemo-cortex
pip install -r requirements.txt        # PyYAML only

python -m harness build                # compile your business
python -m harness run                  # build + launch the backend

Edit harness.config.yaml (business name + vertical), then build. Point your MCP host at the generated build/<you>/mcp.json. Requires Python 3.10+ and Node 18+ (for the Mnemo Cortex bridge a build points at).

Works with

🤖 Any MCP Host

Claude Desktop, Claude Code, OpenClaw, Hermes Agent, LM Studio, AnythingLLM, Open WebUI — point it at the generated mcp.json.