Laya for agents
Agents make many small decisions: which tool to call, whether an input is safe, which element to click. Laya answers each one as a typed probability in one forward pass, locally, so an agent can branch on a number instead of parsing prose from an LLM.
0 projects
MCP servers
The core package includes an optional MCP stdio server with laya_predict, laya_route, laya_preset and laya_status tools:
pip install "laya[mcp]"
laya-mcp-server
{
"mcpServers": {
"laya": {
"command": "laya-mcp-server",
"env": { "LAYA_DEVICE": "cpu" }
}
}
}
Other options:
- laya-mcp (NVentimiglia):
pip install laya-mcp, with guardrail, routing, PR-diff and moderation tools. - ollaya:
claude mcp add ollaya -- ollaya mcp. - typesafe-mcp: built for Jev, and can point at a local Laya server through
TYPESAFE_BASE_URL.
Browser automation
- laya-browser-agent: Playwright/CDP browser decisions with Laya, speaking TypeSafe's
/v1/systemonedialect. It also publishes a fine-tuned browser checkpoint. - laya-ultrafast: a port of browser-use/jev-ultrafast that runs Laya through laya-mlx (Apple Silicon only).
- The core repository documents a browser-agent fine-tune on one 16 GB GPU. Element top-1 accuracy went from 0.10 zero-shot to 0.66, and task success from 0% to 62%.
Agent frameworks and evals
- system1-agents: prebuilt browser-use, computer-use, robotics and game agents on Jev, Laya or Cua-S1.
- jevals: agent evals and guardrails as typed decisions (
pip install jevals), with Laya as a local option. - FluidUse: on-device computer use on macOS, with Laya over Core ML.
Caveats
laya-ultrafast states it plainly: Laya answers narrow questions well (which field is the destination, whether two dates match) but does not reliably answer "what should the browser do next?". Those projects combine narrow Laya questions with rules. The zero-shot numbers above also show that fine-tuning is where most of the gain comes from.
More ways to use Laya