Modern application engineering
Electron, React 18, TypeScript, Vite, Tailwind, Zustand, Zod, Vitest and Playwright/Electron in one working desktop product.
MADS
Multi-Agent Developer Sandbox
A local-first Electron application where AI agents can propose code and commands, while typed contracts, policy checks, human approval, reversible ChangeSets and durable audit records keep the human in control.
What the product demonstrates
MADS turns agent output into a controlled desktop workflow rather than giving an AI unrestricted access to the host.
Human-in-the-loop change control
File mutations become ChangeSets with a diff, risk signals, explicit approval, atomic application and guarded revert.
Agent safety and accountability
Typed actions, path policy, a bounded command runner, main-process provider secrets and a durable JSONL audit trail.
Agent proposal
Zod schema validation
Policy and risk checks
Human approval
Controlled executor
Durable audit
Synthetic ChangeSet demonstration
The scenario below is fictional and runs only in this browser page. It demonstrates the product contract without exposing a private repository, real prompt, filesystem path, provider response or audit log.
Synthetic fixture: `demo-calculator` and every event below were written specifically for public demonstration. The buttons do not modify any files.
ChangeSet
CS-DEMO-0042 · Guard division by zero
src/calculator.ts1 file · 4 changed lines
- export function divide(a: number, b: number) { - return a / b; - } + export function divide(a: number, b: number) { + if (b === 0) throw new Error("Division by zero"); + return a / b; + }
Intent
Modify existing file
Policy result
Workspace-confined · approval required
Evidence record
125
Vitest unit and integration tests recorded by the implementation commit
13
Playwright/Electron E2E scenarios recorded by the implementation commit
v1.2
Delivered product increment: ChangeSet hardening
| Field | Value |
|---|---|
| Private source repository | fbratten/multi-agent-developer-sandbox |
| Current source pin | c84698d91d20bde4771a813029c52e2dfa356827 |
| Implementation-bearing verification pin | dbd87c532bb01931b980ba2660a7f749b2e8a239 |
| Verification date for this public package | 2026-08-02 |
| Recorded verification gate | typecheck, lint, 125 Vitest tests, production build and 13 Electron E2E scenarios |
| Public data | Entirely synthetic fixture and browser-only state transitions |
| Source availability | Private; this page is a bounded public evidence surface, not the source tree |
Evidence precision: the verification counts come from the implementation-bearing commit. The later source merge updated documentation and did not claim a new test execution.
Implemented v1.2 safeguards
Create-over-existing detection
A create action targeting an existing path is classified as an intent mismatch and escalated instead of appearing as a low-risk new-file action.
Newline-only diff visibility
A trailing-newline-only change is shown explicitly rather than being rendered as an empty or invisible diff.
Revert-drift refusal
Revert is refused when subsequent workspace edits would be destroyed by restoring the older state.
Durable ChangeSet audit
ChangeSet lifecycle events are written to the durable audit log, and persistence failure is surfaced rather than represented as success.
Boundaries and non-claims
- MADS is a local, single-operator development environment, not a hosted multi-tenant IDE.
- Application-level policy, approval and workspace confinement are not a kernel or container isolation boundary.
- The command runner is bounded by implemented policy and path controls. It is not presented as a general hostile-code sandbox.
- The current provider integration is Anthropic-focused.
- No real private patch, repository name, prompt, provider output, audit event, credential or filesystem path appears on this page.
- The demo illustrates the contract and state transitions. It is not a browser port of the private Electron application.
- Deferred reference material in the private repository is not claimed as implemented.
Why this matters for applied AI engineering
Calling an LLM is the easy part. MADS demonstrates the surrounding engineering needed to turn agent output into a reviewable product workflow: typed contracts, user experience, secret boundaries, controlled execution, reversible change and verifiable state.