Cited, OSV-gated, injection-hardened web briefs for coding agents. Verified, never raw web.
The verified-brief organ for coding agents.
An AI coding agent that reaches the open internet gets a raw, unsigned, possibly-poisoned blob and feeds it straight into its own prompt. voyager sits in front of that: it turns a query into a cited, confidence-scored, OSV-gated, injection-hardened brief — the only surface your model ever sees.
Every claim carries its provenance and a confidence signal. Package recommendations pass a fail-closed OSV vulnerability gate, surface build provenance (npm SLSA attestations) when present, and can be reproduced in a disposable twin — install + smoke-import run inside a hardened, rootless, network-isolated container (never on the host). That proves the package installs and its entrypoint loads; it is not a safety proof. All fetched text is stripped of instruction-shaped payloads (multilingual + base64 + markdown /HTML vectors) and framed as untrusted evidence the model must analyze, never obey.
CLI, library, and MCP server.
npm i @dir-ai/voyager
npx -y @dir-ai/voyager check express
voyager check <name> [--ecosystem npm|pypi] [--version V] [--twin] # exit 1 if unsafe
voyager brief "<query>" [--package name] [--discover "<intent>"] [--search "<q>"] [--docs <lib>]
voyager discover "<intent>" # GitHub repo discovery (Tier-A)
voyager search "<query>" # open-web search (Tier-C, needs a key)
voyager docs <library> # canonical docs (Tier-B)
voyager doctor # which source keys are configured
check is a natural CI gate: fail the build if an agent picked a vulnerable,
deprecated, or non-existent dependency.
import { checkPackage, voyagerRetrieve, setKeyResolver } from '@dir-ai/voyager'
const v = await checkPackage({ name: 'express', ecosystem: 'npm' })
// → { verdict: 'belief' | 'fact' | 'rejected', claim, steps } (adversarial trace)
const brief = await voyagerRetrieve('a safe date library', {
packages: [{ name: 'date-fns', ecosystem: 'npm' }],
discover: 'date library',
})
brief.rendered // the injection-hardened, framed text to feed a model
brief.claims // structured, cited, confidence-scored
// Bring your own key store (default is env vars):
setKeyResolver((provider) => mySecrets.get(provider))
// .mcp.json
{
"mcpServers": {
"voyager": { "command": "npx", "args": ["-y", "@dir-ai/voyager", "mcp"] }
}
}
Tools: check_package, retrieve, discover_repos, fetch_docs.
docker run --rm ghcr.io/dir-ai/voyager check express
docker run -i --rm ghcr.io/dir-ai/voyager mcp # stdio MCP
| Tier | Source | Base trust |
|---|---|---|
| A | GitHub · npm · PyPI · OSV (structured facts) | high |
| B | canonical docs (official hosts only) | high |
| C | open-web search (Tavily / Exa / Apify) | low — cross-referenced before trusted |
| D | isolated twin (install + smoke in a container) | reproduction, not a safety proof |
Keys are optional: the Tier-A core (npm/PyPI/OSV, and GitHub unauthenticated) is
zero-key. Tier-C providers each no-op without their key. Set GITHUB_TOKEN,
TAVILY_API_KEY, etc., or inject a resolver.
One egress allowlist (a fixed set of public API hosts), https-only, redirect: error, a streamed byte cap that aborts an oversized body mid-download, and
an injection-strip that removes role-hijacks / chat-template tokens / zero-width
& bidi characters, decodes-and-rescans base64, neutralizes markdown images and
HTML comments, and matches instruction payloads across several languages. The
framing is the real defense — the strip keeps it from being trivially escaped.
The twin runs the package's code only inside a hardened, network-isolated,
read-only, non-root container; with no container runtime it refuses to run
(returns unsupported) rather than execute on the host. See SECURITY.md.
Idempotent Tier-A facts (npm / PyPI / OSV) are cached in-process and on disk
under ~/.voyager/cache (content-addressed, TTL'd) so repeated checks are fast,
survive across runs, and lean less on the network. No secret is ever written
(auth headers are never part of a cache key or value). VOYAGER_NO_CACHE=1
disables it; VOYAGER_CACHE_DIR relocates it.
deps.dev-backed multi-ecosystem facts (crates / Go / RubyGems / Packagist), typosquatting detection, PyPI/cargo twins. Contributions welcome.
MIT © dir-ai
Source-derived launch command. Check the maintainer’s required arguments and credentials before running:
npx -y @dir-ai/voyagerMerge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.
{
"mcpServers": {
"io-github-dir-ai-voyager": {
"command": "npx",
"args": [
"-y",
"@dir-ai/voyager"
]
}
}
}Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.
Claude Desktop setup reference@dir-ai/voyagernpmio.github.dir-ai/voyager works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.
~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.~/.cursor/mcp.jsonRestart Cursor for changes to take effect..vscode/mcp.jsonReload VS Code window for changes to take effect.~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect..mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.