MACKESCODE AI Proxy
A private Claude Code launcher and Anthropic-compatible local proxy for running multiple model backends side by side — wrapped in a unified web console for persistent sessions, permission controls, cost tracking, and fleet-wide agent monitoring.
- Category
- AI infrastructure
- Primary language
- Python
- Platform
- Linux · Fedora 44
- Sheet
- E-01
One console for every model and every session.
MACKESCODE sits between your tools and the model backends, presenting a single Anthropic-compatible endpoint while a web console gives you control over cost, autonomy, and observability.
01Anthropic-compatible proxy
A drop-in local endpoint that speaks the Anthropic Messages API, letting existing clients target multiple model backends without code changes.
02Claude Code launcher
Spin up, name, and reattach to Claude Code sessions from one place — each one routed through the proxy with its own configuration.
03Persistent terminal sessions
Long-running terminal sessions survive disconnects and page reloads, so agents keep working while you step away.
04Autonomy & permission controls
Per-session guardrails decide what an agent may run unattended and what needs a human in the loop.
05Cost & budget tracking
Live token and spend accounting per session and per backend, with budget ceilings that stop runaway usage.
06Image prompt studio
A dedicated surface for composing, iterating on, and organising image-generation prompts alongside your coding sessions.
07Fleet-wide agent monitoring
A single dashboard shows every running agent across the fleet — status, activity, and resource use at a glance.
08Multiple backends, side by side
Route different sessions to different providers or local models simultaneously and compare behaviour without reconfiguring clients.
Python core, web console on top.
The proxy and orchestration layer are written in Python; the console is a browser front-end served directly from the same process.
The Anthropic-compatible proxy, session manager, budgeting, and monitoring services — the primary language of the project.
The unified console UI: terminal views, dashboards, and the image prompt studio, delivered to the browser.
The layout and styling of the console, kept lightweight and dependency-light to match the rest of the studio.
Verified against the API contract it emulates.
Because the proxy stands in for a real API, correctness means matching the contract exactly — and proving budgets and permissions hold under load.
- Contract tests assert request and response shapes match the Anthropic Messages API.
- Streaming replay exercises server-sent event streams end to end for parity with upstream.
- Session lifecycle tests cover attach, detach, reconnect, and persistence across restarts.
- Budget & limit tests confirm spend ceilings halt sessions before they overrun.
- Permission-boundary tests verify autonomy controls block disallowed actions.
- Runs in production daily on studio hardware — the same build you see in the repo.
Keys stay local. Nothing leaves the host uninvited.
MACKESCODE is designed to run entirely on hardware you control, so credentials and prompts never transit a third-party service beyond the model backend you explicitly choose.
- Local-first — the proxy runs on your own host; no cloud dependency for the control plane.
- Credential isolation — API keys are held server-side and never exposed to the browser.
- Explicit routing — traffic reaches only the backends you configure, nothing implicit.
- Auditable spend — every request is accounted for, giving a clear record of usage and cost.
- Permission model — autonomy is opt-in per session, defaulting to human-in-the-loop.
- Open source — the full implementation is published for inspection.