The signing, discovery, and trust layer underneath MCP servers, A2A agents, Agent Skills, and collections. Built for developers shipping agent infrastructure.
Forge sits between the publishers who ship agent components and the agents that compose them. The registry signs, scans, and serves. The runtime verifies before it runs.
// The agentic ecosystem now mirrors early Node.js — capability is exploding, but provenance and security remain ad hoc. Skills are copy-pasted from gists and installed as system prompts, sight unseen. MCP servers are pulled from npm by name and granted file, shell, and network access. Agent cards are fetched by URL. Nothing is signed. Nothing is scanned. Nobody checks who published what.
// Marketplaces list tens of thousands of skills and servers, but listing is not vetting: no marketplace today verifies publisher identity, scans for known CVEs, or revokes a listing when it goes bad.
// Forge introduces a single forge.* extension that any MCP, A2A, or Agent Skill manifest can carry. It is interoperable by design. The registry, the verifier, and the AI-client tools are downstream of one specification.
// What Forge adds on top of existing standards "forge": { "security": { "scanned": true, "signed": true, "vulnerabilities": 0 }, "evals": { "accuracy": 0.94, "latency_p50": "8s" }, "compose": { "depends_on": ["ana/web-researcher@1.4"] } }
Forge indexes MCP servers, A2A agent cards, AI skills, and collections (bundles of skills, agents, and commands) today. Evaluators and structured prompts are on the roadmap.
Each layer ships independently. Together they form the supply chain substrate the standards assume but never define.
CVE scanning via OSV, Ed25519-signed manifests, verified publishers via GitHub OAuth, and domain verification. The supply chain layer that doesn't exist anywhere else.
Universal search across A2A Agent Cards, SKILL.md packages, and MCP manifests. Every entry gets a 0–100 trust score based on publisher identity, CVE status, and script safety.
Declare what your pipeline depends on. Forge resolves, verifies, and locks the graph. The routing substrate for interoperable AI systems.
Every package goes through the same pipeline: publisher identity, Ed25519 signature, live CVE scan, install-script analysis. Registry responses are signed, so tampering is detectable end to end.
Add Forge to your editor once. From then on your AI assistant can search the registry, vet packages, audit your configs, and install skills inline — no terminal round-trip.
Claude Code, Cursor, VS Code, Antigravity — Forge becomes a native tool your client can call.
“Find me a verified web scraping MCP” or “audit my current config.”
Your client calls forge_search, forge_get_package, and forge_audit and reports back vetted.
{
"mcpServers": {
"forge": {
"command": "npx",
"args": ["-y", "@forge-registry/cli", "mcp"]
}
}
}Hugging Face is the registry for ML weights. Smithery indexes MCP servers. Neither covers the agentic supply chain end-to-end, neither was designed for the multi-protocol agent stack, and neither offers signed provenance.
| Capability | Hugging Face | Smithery | Forge |
|---|---|---|---|
| Indexes agents + skills + tools | — | tools only | yes |
| Security scanning | partial | — | yes |
| Signed manifests | — | — | yes |
| Supply chain verification | — | — | identity + CVE today |
| Dependency resolution | — | — | v0.3 roadmap |
| Reproducible evals | benchmarks | — | v0.4 roadmap |
| Cross-vendor neutral | yes | yes | yes |
Every major protocol standardization created a discovery and trust layer. The agentic AI standards are next.
| Git | GitHub |
| Node.js | npm |
| Docker | Docker Hub |
| Terraform | Terraform Registry |
| Kubernetes | Helm |
| ML models | Hugging Face Hub |
| MCP + A2A + Agent Skills | Forge |
The forge.* extension is a minimal addition to existing agent manifests. Decide for yourself whether it's worth signing onto.
// The full schema for security, evals, and compose blocks. Security ships today; evals and compose are spec'd ahead of their releases.
Read the spec ↗// One config entry. Claude Code, Cursor, VS Code, Antigravity. Search, vet, audit, and publish without leaving the conversation.
{ "mcpServers": { "forge": {
"command": "npx",
"args": ["-y",
"@forge-registry/cli",
"mcp"] } } }Setup guide ↗// First batch ships with the v0.1 CLI, signed publishing, and verification. Limited to MCP, A2A, and Agent Skill manifests for now.