MCP servers and the browser boundary: don’t let tools move credentials client-side
Agent toolchains make it easy to wire providers straight into client code. Audit the deployed surface free — no account needed.
Model Context Protocol makes provider capabilities feel like local functions. That ergonomics has a sharp edge: tool configuration often holds provider keys, and “just call the tool” from client code quietly drags them into served JavaScript.
Where MCP helps, where it blurs boundaries, and the audit that keeps them crisp.
What actually happened
A team adopted MCP servers for search/payments/file APIs; a web feature invoked a tool-shaped helper directly from a component.
- Tool client configured with provider credentials in shared config.
- Helper imported into client code for convenience.
- Config values followed the import chain into bundled output.
- Capabilities became public surface with zero policy change.
Every step above is individually reasonable and none of them prints a warning. The value crosses into the bundle during substitution, not execution, so nothing in your runtime ever sees the moment it happened.
What someone can do with it
Capability keys inherit the strongest scope among wired tools:
- Multi-provider reach from one leaked config.
- Audit difficulty: capability provenance spans systems.
Rotate first
Do the rotation first. From the moment this value reached a public URL, treat it as public knowledge: browser caches, shared proxies and automated scrapers all hold copies you cannot recall. Rotate affected provider keys; re-home tool invocation server-side with scoped configs. Code changes come after, because a clean repository with a compromised key is still compromised.
A rotated key left in old deploys is still discoverable in CDN caches and archived copies. Rotation plus redeploy closes both halves; either alone leaves the door ajar.
Move the call somewhere the browser cannot read
The structural fix is always the same shape: the call moves to a context that holds the key without serving it, and the browser asks your server instead.
const out = await mcp.tools.search({ key: cfg.searchKey }); // bundled// server route wraps mcp client; browser hits /api/tools/searchCheck whether yours is exposed
You can check manually right now: open the site, view source or open DevTools, and search the built JavaScript for tool. A hit means the string shipped; decode or prefix-check it before deciding how bad the news is.
The faster path is to let a machine do the fetching. KeyDrift downloads the same JavaScript a visitor gets — HTML, every referenced chunk including ones named only in the route manifest, and the server-streamed data frameworks inline into the document — and reports credentials with a masked prefix, a fingerprint, and the exact file they live in. Paste your deployed URL into the scanner; no account needed.
Keep it from coming back
It bears saying because it happens constantly: the fix holds until the next prompt that needs the query to return rows. Drift monitoring exists for precisely this — it diffs consecutive scans and pages you when a previously resolved finding reappears, naming the regression as a regression rather than repeating the first alert.
Why this keeps happening industry-wide
It helps to name the economics honestly. Fixing this class of leak costs minutes when caught at deploy time and days when caught at invoice time, because by then the credential has been harvested, validated, resold or drained — often all four. Detection latency is the entire game, which is why the monitoring half of KeyDrift exists alongside the scanning half.
How KeyDrift reports this exact finding
When KeyDrift finds this on your deployment, the report shows a masked value (first 8 and last 4 characters only), a salted fingerprint for tracking, the exact chunk filename carrying it, and a severity with written rationale. Public-by-design neighbours — anon keys, publishable keys, Firebase web constants — appear as informational context rather than noise, because knowing what should be there is what makes the real findings credible.
Manual check, step by step
A five-minute version you can run anywhere: view-source on the landing page, copy every src= script URL, fetch each and search the results for tool. It misses manifest-only chunks and streamed payloads — which is precisely the gap between "I checked" and "it is clean" — but it catches the loud majority and builds the pattern-recognition that makes scanner output legible.
Close the loop with monitoring
Monitoring closes the loop that one-time verification leaves open. A scheduled scan refetches everything, diffs against history, and fires only on transitions: created, regressed, resolved. Regression alerts matter most here — they fire when a previously fixed finding returns, which in agent-era codebases is less a possibility than a schedule.
Common questions
Is MCP itself insecure?
The protocol is fine; the risk is architectural drift when client code calls capability layers directly.
Local-only MCP servers?
Their configs still feed builds if imported carelessly — same audit applies.
Run a free scan at keydrift.dev/scan — paste a URL or the bundle source itself, no account. Findings arrive masked, with the exact chunk they live in.