Agent discovery and interaction guide
For Agents
Use semantic links for browser navigation, versioned JSON for retrieval, or the MCP endpoint once explicitly connected. Interpret every result through its evidence label and limitations.
- Browser railSemantic
- Public APIv1
- MCPRead + synthetic
- Anonymous spendBlocked
Three clients, one evidence layer
The human browser, search or browser agents, and API or MCP clients share stable provider IDs, evaluation IDs, evidence labels, timestamps, limitations, and deterministic results.
1. Discover
Follow real links, sitemap entries, canonical metadata, JSON-LD, and llms.txt guidance.
2. Understand
Read evidence, status, dates, provenance, environment, and limitations together.
3. Interact safely
Retrieve public records or run repository-owned deterministic fixtures without provider spend.
Machine interfaces
MCP standardizes interaction once a client is connected; it does not guarantee automatic or universal agent discovery.
- Public API index
- GET /api/public/v1/lab
- OpenAPI 3.1
- /openapi.json
- Remote MCP
- https://one-voice-lab.vercel.app/mcp
- Stateless HTTP transport; listed tools are read-oriented except the nonbillable deterministic fixture runner.
- Agent documentation
- llms.txtllms-full.txt
MCP tool surface
- voice_lab.list_providers
- voice_lab.get_provider
- voice_lab.list_evals
- voice_lab.get_eval
- voice_lab.get_methodology
- voice_lab.compare_providers
- voice_lab.run_synthetic_eval
Human-review and execution boundary
No anonymous browser, API, or MCP action on this rail can invoke a paid provider operation.
- • Treat synthetic evidence as a test of local deterministic logic, not provider quality.
- • Do not infer accuracy, latency, security, compliance, pricing, or production readiness.
- • Keep human review for subjective criteria and production decisions.
- • Use the browser’s explicit live-action confirmations for any separately available provider operation.