Part of AI Tools & Automation — automating real work with local AI models.
LLM Vault is a free, open-source password manager with a built-in LLM gateway. It lets AI agents, VS Code Copilot, Claude Code and your own scripts use the models you pay for, while the API keys stay in an encrypted KeePass-compatible vault. On this page you can download LLM Vault for Windows, see what it does in two minutes and pick the setup guide for your tool.
- LLM Vault replaces
get_secret()withinvoke_model(): clients get a scoped token, the provider key never leaves the vault. - The gateway speaks the OpenAI and Anthropic APIs, so Python SDKs, Claude Code and GitHub Copilot Chat work without code changes.
- The vault is a standard KDBX 4.1 file (AES-256, Argon2id 64 MiB), so KeePass and KeePassXC open it too.
- In my tests, 24 probes for secrets returned nothing, and the gateway added about 7 ms to a local model call.
- The Windows installer is a 13 MB per-user setup with no admin rights; a portable ZIP is available too.
Download LLM Vault
LLM Vault 0.1.0 is the first public release. Both Windows packages contain the same two programs: the desktop app llmvault.exe and the headless gateway llmvault-server.exe. Choose the installer for everyday use, or the portable ZIP if you want everything in one folder.
| Package | For | Size | Download |
|---|---|---|---|
| Windows installer | Windows 10 1809+ / 11, x64, per-user, no admin rights | 13 MB | LLM-Vault-0.1.0-windows-x64-setup.zip |
| Windows portable | USB stick or a locked-down PC, data stays in .data | 19 MB | LLM-Vault-0.1.0-windows-x64-portable.zip |
| Example scripts from the videos | Python, PowerShell and the VS Code snippet | 17 KB | llm-vault-examples.zip |
| Checksums | SHA-256 of every file above | 1 KB | LLM-Vault-0.1.0-SHA256SUMS.txt |
System requirements
First, you need Windows 10 version 1809 or newer, or Windows 11, on a 64-bit PC. The installed app takes about 50 MB of disk space. In addition, it needs no administrator rights, because it installs into your user profile. LLM Vault itself does not need a GPU. However, local models need their own engine, such as Ollama or llama.cpp, plus enough memory for the model. For cloud providers, you only need an account and an API key, and that key goes straight into the vault.
Before you run the installer, compare its checksum with the published list. In PowerShell, one command prints it:
Get-FileHash .LLM-Vault-0.1.0-windows-x64-setup.exe -Algorithm SHA256
What LLM Vault does
LLM Vault solves one problem: an AI agent that can read your files can also read your API keys. Coding agents run shell commands, open .env files and process untrusted text. Therefore, a single prompt injection in a README can send your key to a stranger. I describe that risk in detail in the AI agent API key security guide.
LLM Vault changes who holds the key. The provider key lives in the encrypted vault, and the vault makes the model call itself. Meanwhile, each client gets a client token, which is a narrow permission rather than a credential. The token lists the models it may call, and it carries an expiry, a budget and a rate limit. Moreover, you can revoke it with one click.

Who it is for
- Developers who run coding agents such as Claude Code, Copilot agent mode or custom agents in
.github/agents. - Teams that share one Azure AI Foundry, OpenAI or Anthropic account and want a revocable token per person.
- Anyone with local models in Ollama, LM Studio, llama.cpp or vLLM who wants one endpoint and one audit log.
- Script authors who need a hard daily budget, so a runaway loop stops at HTTP 429 instead of a big invoice.
In short, any tool that can call an OpenAI-compatible API can use LLM Vault. Moreover, you can start with a single script and add more clients later, each with its own token.
Features of LLM Vault
The app combines a full password manager with a policy-enforcing gateway. In other words, both live in one native executable written in Rust, without Docker, a browser or a Python runtime.
A KeePass-compatible password manager
The vault stores logins, passwords, URLs, notes and provider keys in groups. It uses the KDBX 4.1 format with AES-256, Argon2id at 64 MiB and HMAC block integrity. As a result, your existing KeePass database opens in LLM Vault, and its files open in KeePassXC. In addition, a password generator and clipboard auto-clear are built in. The full storage design is in the KeePass-compatible password manager deep dive.
Free ebook
Free AI Video, Generated Locally
Working scripts and measured benchmarks. Free.
No spam. Unsubscribe at any time.
An LLM gateway with 38 integrations
By default, the gateway listens on 127.0.0.1:8765 and exposes five routes only: /health, /v1/models, /v1/chat/completions, /v1/messages and /v1/embeddings. Streaming and tool calls work, so agent mode works too. Besides local engines, the catalog covers Ollama local and ollama.com, LM Studio, llama.cpp, vLLM, SGLang, OpenAI, Azure AI Foundry, Azure OpenAI, Anthropic, Gemini, OpenRouter, Groq, Mistral, DeepSeek, xAI and more.

Consent, budgets and an audit log
In addition, each token has its own consent rule: ask on every request, ask once per session, or pre-approve. Notably, the consent window arms its buttons after a short delay, so a script cannot click “Allow” for you. Furthermore, daily and monthly budgets and a requests-per-minute limit stop runaway loops. Finally, every call lands in the audit log with the model, tokens, cost and the policy decision.

Get started with LLM Vault in five minutes
The first run takes five steps, and your scripts change only in two lines. To begin, install the app and create a profile with a user name and a login password. Next, create a vault with its own master password, or open an existing .kdbx file.
Then open Providers and connect one. For example, “Ollama (local)” needs no key, while “Azure AI Foundry” asks for one. After that, open Clients & agents and create a token for each tool. The token appears once, together with ready snippets for Python, Claude Code, curl and VS Code.

Finally, point the client at the vault. In Python, the OpenAI SDK needs only a new base URL and the token:
import os
from openai import OpenAI
# The token can call the models it was given; it cannot read any key.
client = OpenAI(base_url="http://127.0.0.1:8765/v1",
api_key=os.environ["VAULT_CLIENT_TOKEN"])
print(client.chat.completions.create(
model="qwen-local",
messages=[{"role": "user", "content": "Summarise this ticket"}]
).choices[0].message.content)
Claude Code reads three environment variables instead. The model alias must point at an Anthropic or Ollama provider in the vault:
$env:ANTHROPIC_BASE_URL = "http://127.0.0.1:8765"
$env:ANTHROPIC_AUTH_TOKEN = "lvk_..." # vault token, not an Anthropic key
$env:ANTHROPIC_MODEL = "qwen-local" # alias defined in LLM Vault
claude
For Copilot Chat, follow the VS Code Copilot custom endpoint guide. For local engines, read the local LLM gateway measurements.
How LLM Vault compares
Most setups protect the key at rest, but they still hand it over at use time. The table shows where LLM Vault differs from the usual options.
| Option | Agent can read the key | Per-client limits | Cost per call | Desktop app |
|---|---|---|---|---|
Key in .env | Yes | No | No | – |
| KeePass / KeePassXC | Yes, once copied out | No | No | Yes |
| Cloud key vault | Yes, with read rights | Partly (RBAC) | No | No |
| Self-hosted LLM proxy | No | Usually | Usually | No, a server |
| LLM Vault | No | Models, budget, rate, expiry | Yes, labelled | Yes |
Measured numbers
For these figures, I measured the release on my workstation, a Windows 11 PC with an AMD Radeon PRO W7900. First, unlocking the vault took 325–360 ms with Argon2id at 64 MiB. Second, the gateway added about 7 ms on Ollama with qwen3:0.6b and about 42 ms on llama.cpp. Furthermore, 24 probes on paths such as /v1/secrets and /admin returned no secret. A provider error that echoed the key reached the client as [REDACTED].

Security model and limits
LLM Vault is honest about what a desktop app can do. On the one hand, it stops agents from reading keys through files, responses, logs or errors. In addition, it limits what a stolen token can spend. However, malware running as your own OS user could still read process memory or automate the UI.
For that threat, run llmvault-server, the headless gateway, on a separate VM, OS account or device. As a result, the agent reaches only the HTTP endpoint. In that mode only pre-approved tokens work, because nobody is there to click “Allow”. The full threat model ships with the source code.
Open source under Apache-2.0
The complete source code, the tests and the build scripts are public. Therefore, you can audit the gateway, build the app yourself or contribute a provider preset. The code lives on GitHub at github.com/bestin-it/llm-vault. Please report vulnerabilities privately through its security advisory form.
Frequently asked questions
Is LLM Vault free?
Yes. It is free and open source under the Apache-2.0 license, including commercial use. There is no account, no telemetry and no cloud service behind it.
Does LLM Vault work offline?
Yes. The vault and the gateway run on your machine. Only calls to cloud providers leave it, and local engines such as Ollama never need the internet.
Can AI agents read my API keys through LLM Vault?
No. No network endpoint returns a key, a password or the configuration. Instead, agents can only call the models their token allows.
Which tools work with LLM Vault?
Anything that speaks the OpenAI or Anthropic API. That includes the Python and JavaScript SDKs, Claude Code, GitHub Copilot Chat, LangChain, curl and most agent frameworks.
Can I open my existing KeePass database?
Yes. LLM Vault reads and writes KDBX 4 files. In addition, a vault created by LLM Vault opens in KeePass and KeePassXC.
What happens if I forget my profile password?
There is no reset by design, because a reset would be a backdoor. You can delete the profile and create a new one, and your vault files stay protected by their own master password.
When do macOS and Linux builds arrive?
The build scripts for a DMG, an AppImage, a .deb and an .rpm are in the repository already. Signed packages will appear on this page; until then, cargo build --release builds the app on both systems.
Summary
LLM Vault gives AI agents access to models without access to keys. It stores keys in a KeePass-compatible KDBX 4.1 vault and exposes only model calls through a local gateway. On top of that, it adds consent, budgets, revocable tokens and an audit log. In my tests, 24 probes found no secret, and the gateway cost about 7 ms per call. Download the 13 MB Windows installer above, connect a provider, and give each tool its own token. Above all, start small: one provider, one token and one script are enough for the first test.
Companion code — the snippets above are the whole client-side change: a new base URL plus a vault token for the OpenAI SDK, and three environment variables for Claude Code. The example-scripts ZIP contains every script shown in the videos.
Free ebook
Free AI Video, Generated Locally
Run Wan 2.1 in ComfyUI on your own GPU — the scripts I use, measured times, sample clips. No cloud, no API keys.
No spam. Unsubscribe at any time.


[…] LLM Vault itself is free and open source. The installer, the portable version and the example scripts from the videos are on the LLM Vault download page. […]
[…] LLM Vault itself is free and open source. The installer, the portable version and the example scripts from the videos are on the LLM Vault download page. […]