Start with a safe, useful first step.

A local model does the AI thinking on your own device. A tool-capable application connects it to Vaultfire. You do not need a cloud AI provider, a wallet or a payment key for the five public read-only MCP tools.

1. Load a local model: use LM Studio with a model that supports tool use. Model downloads and hardware needs vary; choose a local model rather than a cloud-backed one.
2. Add the connection: in LM Studio, open Program, then Install > Edit mcp.json. Merge the server entry below into your existing mcpServers object; keep any unrelated entries.
3. Try a safe question: ask your model to use vaultfire_list_operations, then vaultfire_get_operation for /api/x402/trust/contracts, and explain what is available without paying or signing.
4. Look for a real tool result: discovery should describe 76 operations, including two free ones. Reading the description of a paid operation does not execute or pay for it. A model answer without a visible tool call is not proof that a request ran.

{
  "mcpServers": {
    "vaultfire-read-only": {
      "url": "https://theloopbreaker.com/api/mcp"
    }
  }
}

Local inference is not the same as offline operation. This connection sends tool requests to the public Vaultfire server, not to a cloud model provider. Tool arguments and normal request metadata leave your device. Never put private keys, wallet seeds or confidential conversation text in those arguments.

The configuration grants no signing, payment or filesystem tools. Standard MCP transport tests do not certify every local model or app version. See LM Studio's setup instructions and verify the result on your own device.

This path uses a terminal, which is the application where you type commands for your computer to run. If that is unfamiliar, you can use the browser check above first, or share these steps with a developer. Do not type commands into a wallet or a chat window and assume they ran.

1. Prepare: use the included Linux development container, or install the exact versions listed below. Installation can take longer than the example itself. Other host setups are not implied to be tested.
2. Download and verify: obtain the selected source and its checksums. A matching checksum means the bytes match the published package, not that the software is independently audited.
3. Run locally: install both packages and execute the existing Quickstart. It signs only with disposable public test keys on a local test chain, never your personal wallet or a public network.
4. Check success: the final report must have top-level "status": "passed", show a consent record being registered and revoked, and show the local receipt reaching finalized. Keep the report and source identity.
5. Build your integration: follow the versioned SDK and API operation contracts. Check which operations are supported before adding authorization, retries or status polling. Do not treat a schema or simulated catalog as live execution.

Register consent, revoke it, and inspect a local receipt. This example runs on your computer with disposable local services. It is not a Base mainnet or Solana deployment, and it needs no funded wallet, private key or payment.

Not ready to install development tools? Start with either path below. A terminal is the window where these commands run on your own computer.

1. Get the current source. Use the current download and follow its checksum and extraction checks before running it. Do not substitute an older bundle if retrieval fails.
2. Install the pinned environment. Use the partner devcontainer included at full-v3/partner/.devcontainer, or install the exact versions below. Dependency installation and native tool installation are separate setup costs, not part of a prepared quickstart timing.
3. Run only locally. From the verified extracted root, install the SDK dependencies as well as the sandbox, then run the existing quickstart. No wallet, private key, seed phrase, payment header or public RPC is required or accepted.

# From the authenticated extracted source root, with pinned tools installed:
npm --prefix full-v3/sdk ci --ignore-scripts
npm --prefix full-v3/partner ci --ignore-scripts
npm --prefix full-v3/partner run quickstart

What success looks like: the final report has top-level "status": "passed", confirms a local consent record was registered and revoked, and shows the local receipt reaching finalized. Keep the output. A pass proves this local example ran; it does not activate V3 or qualify partner funds.

If it fails: keep the error and stop at the failing step. Missing tools require the exact pinned environment, not looser checks. A busy local port must not be replaced with a public network. Package-download failures require restoring the documented download path, not sharing credentials. Recovery details are below.

export VAULTFIRE_TWIGPINE_TOOLS="$(mktemp -d)"
npm install --prefix "$VAULTFIRE_TWIGPINE_TOOLS" --no-save --ignore-scripts=false @gitlawb/gl@0.7.1
export PATH="$VAULTFIRE_TWIGPINE_TOOLS/node_modules/@gitlawb/gl/bin:$PATH"
test "$(git-remote-gitlawb --version)" = "git-remote-gitlawb 0.7.1"
test "$(uname -s)-$(uname -m)" != "Linux-x86_64" || test "$(sha256sum "$(readlink -f "$(command -v git-remote-gitlawb)")" | cut -d' ' -f1)" = "af5da2da412dfe24fff57b1d8a0834da1fa5d971897423a19ded454a0f732e74"
export GITLAWB_NODE=https://node.gitlawb.com
git clone --recurse-submodules gitlawb://did:key:z6MkryiNsYdFEMHv95wxzSQ1vtFXPyVQQrxXdPw3tE5HpfxV/vaultfire
cd vaultfire
git checkout --detach 385415d13aca6f8bd97143cdc63dd2a70f27fa02
test "$(git rev-parse HEAD)" = "385415d13aca6f8bd97143cdc63dd2a70f27fa02"
npm --prefix full-v3/sdk ci --ignore-scripts
npm --prefix full-v3/partner ci --ignore-scripts
npm --prefix full-v3/partner run quickstart

Pinned environment: Node 20.20.1, npm 10.8.2, SQLite 3.46.1, Foundry 1.3.1, and project Solc 0.8.24. This is local software verification only. It does not deploy V3 to Base or Solana, enable payments, verify a real ML-DSA proof, or authorize partner funds.