What it is
Laya is an open-weight 'System One' decision model from Convai Innovations: instead of generating text, it answers typed questions (pick one option, rate on a scale, or a yes/no probability) about a piece of text or JSON in a single forward pass. The Laya Python package ships an optional MCP server, installed with pip install "laya[mcp]" and started as laya-mcp-server, so an agent can call those decisions as tools.
The server exposes ten tools: laya_predict, laya_predict_batch, laya_predict_long, laya_route, laya_route_batch, laya_decide, laya_shortlist, laya_tournament, laya_preset and laya_status. laya_decide takes a JSON schema and returns values that fit it; laya_shortlist and laya_tournament narrow choices with more than about 20 options, which the base model handles poorly. It runs over stdio and loads the English (421M parameters) and multilingual (322M) checkpoints on your own CPU or GPU, or points at a shared laya-serve HTTP server through LAYA_BASE_URL.
The repository has about 34,000 GitHub stars, was created on 18 September 2026 and had shipped 46 releases by version 0.4.13 on 21 October 2026. Code and weights are Apache 2.0.
Install
pip install "laya[mcp]" && laya-mcp-serverOur review
Laya's MCP server is the only vendor-made MCP server we found for a System One model whose weights you can run yourself, and that is its whole case. Every decision stays on your machine, there is no per-token bill, and the README measures a single question at 32.8 ms on a T4 GPU, against the 236 to 276 ms that third-party benchmarks it cites recorded for TypeSafe's hosted Jev.
The tool set is more complete than any Jev wrapper we read. Batch tools group requests by checkpoint and share forward passes; the README measured one 16-ticket batch call at 2.1 to 2.3 times faster than 16 single calls on Apple GPUs, with no decision flips. Shortlist and tournament tools exist because the model's option budget collapses past about 20 labels: on Banking77 the README reports Laya at 0.425 on all 77 labels against a third-party 0.870 for Jev on 72.
The maintainers are unusually candid, and their 'Honest limits' section should be read before installing. The shipped base checkpoints score 0.362 and 0.352 on their own typed-decisions benchmark, barely above a 0.318 random baseline and below the 0.461 majority-class baseline; the 0.766 headline figure comes from a checkpoint fine-tuned on that benchmark's training split. The README calls Laya 'a fast base to specialise, not a zero-shot decision engine', and records negated cancellation requests classified as cancellations, one at 0.9998 probability. Both checkpoints are over-confident until you fit temperatures on your own data.
The practical costs are PyTorch, about 2.3 GB for all three checkpoints and stdio only. Sharing one model across sessions means running laya-serve separately, and in that mode the shortlist and tournament tools disappear. Every benchmark comparing Laya and Jev comes from the Laya repository, and its Jev numbers are third-party figures the authors did not measure themselves, so the speed gap is better evidence than the accuracy table.
Where it earns its place
A support engineering team at a European bank handling 5,000 tickets a day
Route tickets to the right queue without sending customer text to a US vendor
Fine-tune on a few thousand labelled past tickets, fit temperatures, then let the agent call laya_route_batch. At 103 to 332 questions a second batched on one T4, a single GPU covers the day's volume many times over, with no per-call fee.
A developer adding guardrails to a local coding agent on a laptop
Flag risky prompts and tool calls offline
Run laya-mcp-server with LAYA_DEVICE=cpu and use laya_preset for the prompt-guardrail workflow. Expect roughly 190 to 460 ms per question on CPU, and validate on your own examples, since base checkpoints are near chance zero-shot.
An ML lead at a 40-person SaaS company
Replace a hosted classification API bill with self-hosted decisions
Run one laya-serve on a shared GPU box and point every developer's MCP client at it with LAYA_BASE_URL and LAYA_API_KEY. The hosted bill goes away, but shortlist and tournament tools are unavailable in this mode.
Before you commit
- Setup
- pip install "laya[mcp]" then laya-mcp-server, configured as a stdio command in Claude Desktop, Cursor or OpenClaw. The first run downloads checkpoints from Hugging Face (about 2.3 GB for all three); after that, preloaded checkpoints answer immediately.
- Integrations
- Ten tools: laya_predict, laya_predict_batch, laya_predict_long, laya_route, laya_route_batch, laya_decide, laya_shortlist, laya_tournament, laya_preset and laya_status. All are read-only decisions over the text you pass in; none touches external systems.
- Data and privacy
- Everything runs on your hardware; no inputs leave the machine after the weights are downloaded. A shared laya-serve uses a bearer key (LAYA_API_KEY). Code and weights are Apache 2.0, with no revenue caps.
- Cost at scale
- No per-call fees. The README measures 32.8 ms per question on a T4 and 103 to 332 questions a second batched; on CPU expect about 193 to 464 ms per question. Fine-tuning is documented on two Kaggle T4s or Apple Silicon.
- Limits
- Base checkpoints score 0.362 and 0.352 against a 0.318 random baseline until fine-tuned; accuracy collapses past about 20 options without shortlisting; the English checkpoint takes 512 tokens and fails on non-Latin scripts. Stdio only, and shortlist and tournament tools are unavailable through a shared server.
What's good
- Open weights under Apache 2.0; runs fully offline
- Ten tools, including batch, long-document and schema-driven decisions
- About 33 ms per question on a T4, per the README
- Routes 100+ languages to a multilingual checkpoint automatically
- Very active: 46 releases in its first five weeks
What's not
- Base checkpoints are near chance on typed decisions until fine-tuned
- Accuracy drops sharply past about 20 options without shortlisting
- Needs PyTorch and about 2.3 GB of checkpoints
- All comparative benchmarks come from the project itself
Best for
Teams with labelled examples and data that must stay on their own hardware who want fast routing, triage or guard decisions inside an agent.
Alternatives
Laya MCP Server: common questions
- Is Laya MCP Server worth it?
- Typed decisions that never leave your hardware and cost nothing per call, from a model its own authors say needs fine-tuning before you trust it. We rate it 4.2 out of 5.
- How much does Laya MCP Server cost?
- Free and open source (Apache 2.0, code and weights). You pay only for the hardware it runs on.
- What are the downsides of Laya MCP Server?
- Base checkpoints are near chance on typed decisions until fine-tuned; accuracy drops sharply past about 20 options without shortlisting; needs PyTorch and about 2.3 GB of checkpoints; all comparative benchmarks come from the project itself.
- Who is Laya MCP Server best for?
- Teams with labelled examples and data that must stay on their own hardware who want fast routing, triage or guard decisions inside an agent.
- How much does Laya MCP Server cost for a team or in production?
- No per-call fees. The README measures 32.8 ms per question on a T4 and 103 to 332 questions a second batched; on CPU expect about 193 to 464 ms per question. Fine-tuning is documented on two Kaggle T4s or Apple Silicon.
- Is Laya MCP Server safe for company data?
- Everything runs on your hardware; no inputs leave the machine after the weights are downloaded. A shared laya-serve uses a bearer key (LAYA_API_KEY). Code and weights are Apache 2.0, with no revenue caps.
- What are the best Laya MCP Server alternatives?
- The closest alternatives we have reviewed are Jev MCP, System One Connector and Jev on Viorant.
Answers checked Oct 23, 2026.
Make Laya MCP Server? Show this rating on your site
