A work-specific exobrain — a locally-sovereign AI assistant scoped entirely to a single employer relationship. Built on the Local Mind Foundation architecture.
When the job changes, the number increments: Prosper0 → Prosper1 → Prosper2. The architecture carries forward. The vault stays archived.
Most AI productivity tools are designed for individual optimization inside a trust relationship that doesn't exist: the employer trusts the AI, the employee trusts the vendor, and no one has a clear answer for where the data goes.
Prosper0 takes a different position:
- The operator owns the hardware. The stack runs locally — no cloud, no vendor data custody.
- The employer can verify everything. Every AI action is logged. Every data transfer is emailed with the employer CC'd. The employer has a paper trail they didn't have to ask for.
- The AI cannot expand its own permissions. Tool access is controlled by a config file the AI cannot write. Any capability expansion requires the operator to edit the config manually.
This isn't a productivity tool. It's a trust architecture.
Six layers, each independently testable:
┌─────────────────────────────────────────────────────────┐
│ Layer 6: Portable Deployment │
│ Encrypted USB drive · Hardware-drive pairing · │
│ Single desktop shortcut = entire setup │
├─────────────────────────────────────────────────────────┤
│ Layer 5: Testing Infrastructure │
│ Sample data generator · Model version comparison · │
│ Bridge integration tests · Data boundary tests │
├─────────────────────────────────────────────────────────┤
│ Layer 4: Employer Transparency │
│ AI-immutable tool config · Audit trail · │
│ Data transfer manifest · Transparency reports │
├─────────────────────────────────────────────────────────┤
│ Layer 3: Prospero Bridge │
│ Context switching · Shared TTF calendar · │
│ Mode propagation · Email-CC'd data transfer │
├─────────────────────────────────────────────────────────┤
│ Layer 2: Prosper0 Vault │
│ Work-scoped flat-file vault · Surfacing engine · │
│ Webhook handler · Work modes │
├─────────────────────────────────────────────────────────┤
│ Layer 1: LLM Stack (von Prosper0) │
│ Local inference · Model-agnostic interface · │
│ MCP wiring · Orchestrator · Tool config │
└─────────────────────────────────────────────────────────┘
The AI brain for the work instance. Local inference only — no cloud API. The inference layer is model-agnostic: swap the model without touching the orchestrator.
stack/tools.config.yaml— operator-controlled tool permissions; the AI cannot write this file- Model selection is an open deliverable; the interface contract is designed first
- Version testing harness: run the same prompt battery across model versions and diff the output
See: stack/README.md
A work-scoped flat-file vault (markdown). Fully separate from any personal exobrain — no shared directories, no passive data flow. Mirrors Marlin patterns but scoped to work context.
prosper0.py— surfacing engine (one work task at a time, operator-declared mode)- Work-specific modes:
available·in-meeting·deep-work·off-hours
See: vault/README.md
The controlled interface between the personal exobrain (Marlin) and the work instance (Prosper0). No data flows automatically. Every crossing is explicit, operator-initiated, and logged.
- Context switch signal (which instance is active)
- Shared TTF calendar view (both vaults push tasks with a
source:tag; calendar renders both without cross-contaminating vault content) - Mode propagation (Marlin mode change → Prosper0 behavioral response)
- Data transfer: operator initiates → system drafts email → operator sends with employer CC'd → system logs
See: bridge/README.md
The structural guarantee that Prosper0 is trustworthy to an employer.
tools.config.yamldefines all AI tool access; human-editable only- Every AI tool invocation is logged before and after execution
- Every approved data transfer is logged with content hash, destination, and email message ID
- Transparency report generator produces human-readable summaries for any time range
Every layer is independently testable. No real employer data is ever used in tests.
- Sample data generator: realistic synthetic vault content
- Model version comparison harness: capability regression testing across model upgrades
- Bridge integration tests: end-to-end context switching and transfer flows
- Data boundary tests: assertion-based proof that vault isolation holds
See: tests/README.md
The full stack runs from a USB drive. The reference deployment story: hand someone the drive, ask to put a shortcut on their desktop, done. The drive is encrypted and hardware-paired — physically removing it ends the session cleanly.
Deployment configurations are flexible: vault on host, model on host, or everything on the drive. Size is tracked and optimized as a project statistic.
See: deploy/README.md
See spec/ for architecture decision records:
spec/lmf-adr-003.md— LMF Virtualization Layer (how Prosper0 relates to the broader architecture)spec/prosper0-adr-001.md— Project Charter (founding principles)spec/architecture.md— layer-by-layer design reference
See spec/ for architecture decision records:
spec/lmf-adr-003.md— LMF Virtualization Layerspec/prosper0-adr-001.md— Project Charterspec/prosper0-adr-002-enforcement-middleware-chain.md— Enforcement middleware chainspec/prosper0-adr-003-ed25519-employer-signed-config.md— Ed25519 config signingspec/prosper0-adr-004-self-certification-transfers.md— Self-certification for data transfersspec/prosper0-adr-005-fail-closed-enforcement.md— Fail-closed enforcementspec/prosper0-adr-006-model-selection.md— Qwen2.5 7B selectedspec/prosper0-adr-007-inference-runtime.md— Ollama in Dockerspec/prosper0-adr-008-orchestrator-design.md— Orchestrator loop designspec/architecture.md— Layer-by-layer design referencespec/mvp-roadmap.md— MVP milestone map
cd deploy/
docker compose up -d
./pull-model.sh # first run only — pulls qwen2.5:7b
docker attach prosper0-orchestrator| Layer | Status |
|---|---|
| Layer 1: LLM Stack | ✅ M0–M3 complete — Ariel can start and take input |
| Layer 4: Employer Transparency | ✅ Enforcement chain complete — 33 tests passing |
| Layers 2, 3, 5, 6 | 🔲 Not started |
Current: Merging M0–M3 to main. Next: live smoke test (docker compose up).
AI: Ariel von Prosper0 — Qwen2.5 7B via Ollama.
Built by Jared Allison. Part of the Local Mind Foundation project.