Treg is a tool registry and proxy for AI agents, modeled on what OpenRouter does for models. An agent gets one base URL and one token, searches a curated catalog of external endpoints by the job it needs done (SEO and backlinks, social and trends, people and company enrichment, ads, scraping, image and video generation), reads the price, and calls it. Calls are billed per call against a team's prepaid balance, so the agent needs no account with the provider. The hosted registry runs at treg.to, and the server can also be self-hosted.
The registry also holds a team's own tools: endpoints (an upstream base URL plus credential bindings), vendor CLIs such as Stripe, GitHub and Vercel run with the credential injected, and skills registered together with their secrets and tools. Teammates' agents call these without ever holding the key, and calls on a team's own key are never metered. A scan command reads a project and previews which keys, skills and CLIs it would register, and an upload command registers them, encrypted server-side.
For a cataloged call, Treg picks the credential in order: the team's own registered tool for that provider, then a stored team secret, then a verified public route that needs no provider key and is free, and finally Treg's own key billed to the balance. An endpoint with no published price is refused rather than served free. Where several providers serve one capability, search shows them side by side with prices and the caller chooses, except on opt-in routed endpoints for common jobs such as web search, page extraction and email lookup, where Treg picks the provider and names it.
The proxy relays the upstream request faithfully and injects authentication server-side. It changes only hop-by-hop transport headers, its own control headers and session cookie (stripped before the upstream sees them), and the injected credentials. Out-of-balance responses are an HTTP 402 carrying the balance, the estimated cost and a top-up link, so an agent can act on them without parsing prose.
Features
- Pay-per-call catalog: endpoints grouped by what they do, searchable by task, each with parameters, price and an example response
- Credential ladder: own tool, then stored team secret, then a free public route, then Treg's key on the prepaid balance, with own keys always winning and never metered
- Routed endpoints: opt-in capability endpoints for web search, page extraction, Google results and email or phone lookup, with optional verification of found contacts
- Team endpoints: register an upstream URL with one or more credential bindings (API key, token file, auto-refreshed OAuth, or CLI keychain material) and call it through the proxy by host
- Vendor CLIs: run CLIs such as Stripe, GitHub and Vercel with the team credential injected, locally or on the registry server so the key never reaches the caller
- Shared skills: register a skill recipe with its secrets and tools so the whole team runs the same version, and install shared skills into an agent
- Workflow skills: ready-made recipes for finding in-market buyers from public signals, making UGC videos, and giving Claude Code a persistent preference memory
- Agent surfaces: a Claude Code plugin, a skill for other agents, an MCP server, a Claude connector limited to curated endpoints, an llms.txt onboarding file and an OpenAPI-documented HTTP API
- Arenas: Enrich Arena and Web Arena compare providers' answers, cost and speed side by side or as a sequential waterfall, with a rolling leaderboard
- Teams and roles: every secret, tool and skill belongs to an org, with owner, admin, member and viewer roles
- Audit and health: call and CLI-run audit logs, plus optional per-tool health probes that validate credentials, refresh OAuth and notify owners of breakage
