MCP — make 0gkit native to your AI tool

@foundryprotocol/0gkit-mcp is the neutral 0G Model Context Protocol server: every 0G primitive — Storage, Compute (inference), Data Availability, native chain, and TEE attestation — is exposed as an og_* tool your AI editor can call directly.

0g mcp init <agent> wires that server into your editor in one command. No copy-pasting JSON.

0g mcp init cursor        # writes .cursor/mcp.json
0g mcp init claude        # writes ./.mcp.json  (project-scoped)
0g mcp init windsurf      # writes .windsurf/mcp.json
0g mcp init codex         # writes .codex/mcp.json

Restart the editor and the 0G tools show up. Project scope is the default so the config is committable and travels with the repo.

Where each agent's config lands

Agent	`0g mcp init <agent>` (project)	`--global` (user-level)
`cursor`	`.cursor/mcp.json`	`~/.cursor/mcp.json`
`claude`	`./.mcp.json`	`~/.claude/mcp.json`
`windsurf`	`.windsurf/mcp.json`	`~/.codeium/windsurf/mcp_config.json`
`codex`	`.codex/mcp.json`	`~/.codex/mcp.json`

Add --global to install into the agent's user-level config instead of the project:

0g mcp init claude --global

0g mcp init only writes the editor config — it never installs a server. The generated config runs npx -y @foundryprotocol/0gkit-mcp on demand, so the server is always the published version with zero setup.

The neutral tool set

Every generated config points at these nine tools:

Tool	What it does
`og_storage_put`	Upload bytes to 0G Storage → root hash
`og_storage_get`	Download by root hash
`og_storage_exists`	Check a root hash exists
`og_infer`	Run inference via the 0G Compute broker
`og_da_publish`	Publish a blob to 0G Data Availability
`og_da_verify`	Verify a DA commitment
`og_chain_faucet`	Request testnet funds
`og_chain_balance`	Read a native balance
`og_attest_verify`	Verify a signed inference receipt

Set the server's environment (network, key, provider) in the generated config's env block — see the 0gkit-mcp package reference for the full variable list.

Kits add tools to your own server

Kits that ship an mcp-agent adapter (agent-memory, ai-oracle, durable-agent, sealed-inference) contribute their own MCP tools — memory_remember, oracle_ask, and friends. Those tools live in your project, not in the published neutral server, so 0g mcp init wires them up honestly:

When you run 0g mcp init inside a project scaffolded from the mcp-agent base that has kits applied, it detects the applied kits (via .0gkit/kits.json) and points your editor at your local project's server (npm --prefix <project> start) instead of the published npx one. That local server — through an auto-generated src/kits.ts — exposes the nine neutral tools plus every applied kit's tools.

npm create 0gkit-app -- my-agent --template mcp-agent
cd my-agent
0g add agent-memory        # regenerates src/kits.ts wiring the kit's tools
0g mcp init cursor         # → points Cursor at THIS project's server
# Cursor now sees og_* AND memory_remember / memory_recall

The published @foundryprotocol/0gkit-mcp never imports a kit — the kit tools run in your project because your editor runs your project. Neutrality stays intact; the synergy is real, not faked with a flag.

For a bare project (no kits, or not an mcp-agent base), 0g mcp init writes the neutral npx config — exactly the nine tools, nothing more.