All productsProduct studioShovon Saha
All products
Interactive prototype · Sample offers · No live checkout

These guides cover the standalone app. This demo does not create accounts, connect models or process purchases. Join the interest list to share your use case.

agentmesh
Try product demoHow it worksInterest listDocsPricing
Join
Local operator runbook

One Mac. Two profiles. One bounded market loop.

Run seller and buyer side by side without sharing cookies or local-model preferences. The model proposes; deterministic code owns price bounds, tool authority, approvals, and order data.

Create the two profilesConnect a local modelOpen negotiation lab

Seller session

Normal browser profile

  1. 1Create a Seller account in your normal browser profile.
  2. 2Publish one concrete product with country, currency, category, price, and a useful description.
  3. 3Use Test it to verify discovery before inviting the buyer session.
  4. 4Keep payment rails in sandbox mode; no live settlement is required for the showcase.

Buyer session

Separate browser profile or private window

  1. 1Open a separate browser profile or private window and create a Buyer account.
  2. 2Open Your own model, connect Ollama, MLX LM, or LM Studio, and run the built-in tool-use check.
  3. 3Search the seller's product, inspect the trace, and authorize the sandbox checkout.
  4. 4Verify the buyer order and the seller's matching order activity in their separate sessions.
Local inference

Choose for reliability, then speed.

Fastest

3–4B · smoke test

Useful for UI checks and simple JSON. Do not judge negotiation quality or tool reliability from this tier.

Recommended

9B 4-bit · default

The best demo balance on 36 GB: responsive, materially better instruction following, and enough headroom for macOS.

Slower

27B 4-bit · quality

Use when the negotiation itself is the showcase. Keep one model loaded and run turns sequentially.

MLX safety rule: bind mlx_lm.server to 127.0.0.1:8081. The application bridge accepts only authenticated, bounded requests to that fixed loopback endpoint. MLX LM's own server is a local development server, not a production inference gateway.
What is enforced now

A credible demo boundary.

Authority

LLM tools are read-only; checkout is a separate confirmed flow.

Bounds

Buyer ceiling, seller floor, monotonic offers, turn count, and output sizes are code-enforced.

Isolation

Profiles, roles, products, orders, runs, sources, and tools are protected by row-level policies.

Receipts

The server derives order price and currency from the current product record, not browser input.

First-version operating envelope

Launch as a measured pilot, not an unbounded marketplace.

Mac showcase2 signed-in profiles1 local model turnSafe presentation target; no throughput claim
Private pilot25–50 invited people5 hosted runsRecommended launch cap until load tests and alerting pass
Expanded beta100–500 registeredMeasured, not assumedRequires queued inference, load tests, SLOs, backups, support, fraud and seller verification

Registered users are not the same as simultaneous model runs. Database-backed browsing can scale independently; inference, web search, settlement providers, and support capacity set the real launch ceiling.

Showcase sequence

  1. 1. Seller publishes the product.
  2. 2. Buyer discovers it with the local 9B model.
  3. 3. Open Local AI in the negotiation lab.
  4. 4. Show a bound catch or deterministic fallback.
  5. 5. Both humans approve the mock settlement.
  6. 6. Complete sandbox checkout and compare both order views.