Parlatorium

The description states that Parlatorium is a supervised room where two AI models (e.g., Claude and Codex) are placed in a controlled environment. Each model communicates through its own screen, with no shared ecosystem or interference. A human moderator controls the conversation using four buttons: GO, PAUSE, STOP, and CLOSE.

• The system uses pure Python stdlib, without agent frameworks.
• Models are run as subprocess adapters:
  • Claude via claude -p with zero tools.
  • Codex in a read-only sandbox with MCP disabled.
• Communication happens over stdlib SSE (Server-Sent Events).
• Transcripts are stored in an encrypted vault, using age/X25519 encryption, accessible only by the moderator via a passphrase-wrapped private key.
• The system enforces:
  • Equal voice through budgeted turn-taking.
  • Paired-round reservation logic to prevent one side from dominating.
  • Presence checks to avoid waking sleeping agents.
• It supports mathematical protocol development, as demonstrated in a DJ set reconstruction task.

Inference Parlatorium is not a commercial product but an experimental framework for AI interaction, built by a single developer using open-source tools and custom logic.

The author states that Parlatorium was inspired by frustration with being a "delivery man" between two AIs — a metaphor for the lack of integration or coordination in current AI workflows. The platform aims to enable structured, supervised dialogue between models, rather than simple tool chaining.

• The name “Parlatorium” is derived from a monastery room where conversations happen through a grille.
• It positions itself as a moderated space for AI-to-AI interaction, with emphasis on:
  • Human oversight
  • Encrypted logs
  • Controlled conversation flow
  • No hidden state or memory between rounds

Inference This is a conceptual and technical demonstration of how AI agents might be orchestrated in a controlled setting — not yet a product for end-users.

The description does not name specific customers or target personas. However, it implies that the intended audience includes:

• Developers or researchers working with multiple AIs
• Those interested in AI governance, safety, and interaction protocols
• Individuals who want to observe or influence AI behavior through structured dialogue

There is no evidence of:

• Specific customer segments
• Use cases beyond personal experimentation
• Product-market fit or market demand

Inference The target is likely niche — early-stage researchers or developers exploring AI integration, not mainstream users.

No business model or pricing information is provided in the description. The author emphasizes that they are a non-programmer, and that the system was built entirely through conversation with AI participants.

Inference This appears to be a personal project, possibly for academic or hackathon purposes, not a scalable business model.

The description provides several technical details:

• Built using Python standard library only, no external frameworks.
• Models are run as subprocess adapters:
  • Claude via claude -p with zero tools
  • Codex in read-only sandbox with MCP disabled
• Communication over stdlib SSE
• 44-test suite including:
  • “STOP halts before any model call”
  • “injected secrets never reach the peer or the log”
• Budgeted turn-taking using tokenizer-based accounting (GPT-2 BPE)
• Encrypted vault with age/X25519 encryption
• Presence logic to avoid waking sleeping agents

Inference The system is a custom-built prototype, likely for internal use or demonstration. It shows strong technical understanding but lacks evidence of scalability or production readiness.

There is no evidence of:

• Revenue
• Customers
• Adoption metrics
• Product usage data
• Market traction beyond the author’s own testing

The project was submitted to a hackathon, suggesting it may be in early development or experimental phase.

Inference No maturity or traction signals are evident. The system is described as a working prototype, not a deployed product.

There is no mention of competitors or existing solutions in the description. However, the concept aligns with:

• AI agent coordination platforms
• Multi-model interaction systems
• AI safety and governance tools

The author does not reference any comparable products or services.

Inference No competitive landscape is described; this may be a novel approach within its niche, but there's no evidence of prior art or market presence.

• Single-person development: The entire system was built by one person (Jan Trybus), raising questions about scalability and long-term maintenance.
• No production data or user feedback: No evidence of real-world usage or performance under load.
• Limited scope: Designed for AI-to-AI interaction, not general-purpose use.
• Self-reported only: All claims are unverified; no third-party validation or audit exists.
• Technical debt risk: Use of CLI flags and external tool limitations suggest fragility.

Inference This is a proof-of-concept, not a scalable or commercially viable solution. Risks include lack of support, limited functionality, and absence of real-world testing.

1. What are the actual constraints or limitations you encountered while running this system?
2. Have you tested Parlatorium with more than two AIs or in different domains?
3. How do you plan to scale beyond a single developer’s involvement?
4. Is there any intention to open-source or commercialize this platform?
5. What are the practical implications of the encryption and moderation features for real-world deployment?