Security and data

Where is data processed, and what is sent?

The Linklbee engine runs on your server. Requests go from there straight to your LLM vendor; they never pass through ours.

Scanner

One file, standard library only, zero lines of network code. No AI API calls and no automatic usage reporting occur during the scan. Logs that keep the vendor's raw responses work too. The output file keeps only numbers, model names, times and hashes (--hashify). If the hashed output would still contain original text, the file is not saved. The --share summary holds numbers only: no text, file names, or keys.

Linklbee engine

Runs on your server. Requests go from there straight to the vendor. Past answers, and the requests that produced them, are stored on your own server, separated per customer. Document IDs and item types are stored as hashes.

Within a tenant, answers are kept per document key. If one account serves many end customers, per-customer separation must be configured at integration: put the end-customer ID in X-Linklbee-Key (e.g. {customer_ID}:{document_ID}). Without an end-customer identifier, an earlier response within the same account can be reused as the draft for a different end customer’s request. Reuse only happens within the same customer’s responses; each customer’s store is separate, so another customer’s responses are never mixed in.

Keys

In Air-gapped, Linklbee keys are issued on your server and only their SHA-256 is kept. Vendor keys are not stored; they go to the vendor with the request as they came. Linklbee API keys are not sent to the AI provider.

Air-gappedCloud-managed
Where the engine runsYour serversYour servers
Where you manage itCommand line on that serverLinklbee dashboard
What goes to LinklbeeA monthly usage report file you review and exportUsage statistics and operating metadata (see items below)
Key issuingOn your serverBy Linklbee

Data sent in Cloud-managed

There is one path. The Linklbee dashboard server reads the engine’s usage response once a day (the engine does not push anything). The engine reads only its usage ledger, never the stored request and response text files. “Stored” below means stored on the Linklbee dashboard side.

ItemWhat it containsExampleWhen it is sentStored?
Query windowStart and end dates · price-table date · your account identifier (a string ID)2026-10-01 to 10-28 · price table 2026-10-25Once a day + when the settings screen opensDates only
Usage totalsModel name · route (reuse of a previous answer / exact match / changed part only / re-call / full / not in scope) · arm (Linklbee / baseline / not in scope) · request count · input/output tokens · cache read/write tokens · audio and image tokens · answer tokens · cost (from the price table)gpt-4o-mini · changed part only · Linklbee · 1,204 requests · $18.42Once a day + screenStored (daily summary, one row per model, route and arm)
Per-document statisticsHash of the document key (12 characters, not the original) · arm · model · request count · tokens · cost · answer tokens3f9a1c0e…Once a day + screenHashes are not stored — only the document count and sums used for range estimates
Cost-label totalsLabel names you attached to keys or requests · model · request count · tokens · cost"billing-team"Once a day + screenNot stored (shown on screen only)
Baseline settingsBaseline share in effect · whether it is a tuned value or the default · the next candidate · a one-line reason · baseline request count in the window0.10 · defaultOnce a day + screenShare only
Saving-setting statusOn/off per automatic saving setting · decision history · estimated amount each setting saved · on/off of the switch that turns off rewriting only (no full rewrite) · proposal cards (setting name · amount · your answer) · auto-adjustment summaryprefix normalisation on · padding proposal $12.3Once a day + screenNot stored
Additional instructionsThe cache-minimum-length padding text you typed into the settings screen (only if present)(text you wrote)Once a day + screenStored on the engine server (your account’s settings file) · the dashboard only displays it — the one exception where text travels
Waste totalsCount and cost per scan item (retries · length-cap stops · unstable prefixes · cache minimum length · cache hit rate · shared prefix · Flex)retries 12 · $0.41Once a day + screenNot stored
Cache hit-rate summary · paid empty answersRate · judged row count · count · cost23.4%Once a day + screenNot stored
NotesFixed English sentences the engine writes beside numbers (basis of calculation · caveats) — no customer data"cache-read input tokens / input tokens …"Once a day + screenNot stored
Read logTime of reading · success or failure · row count · failure reason2026-10-28 00:10 UTC · successEach readRecorded by the dashboard

Numbers and names only, plus notes the engine writes itself.

When it is sent: usage data is sent once a day. Opening the settings screen triggers an additional read; what is read then is not stored.

What is not sent: request and response text · prompts · API keys (not even hashed) · end-customer IDs · your company name · document text and document names · request timestamps (no per-row times, only window totals).
One exception: additional instructions you type into the settings screen are stored in the engine via the Linklbee management screen and shown again there when settings are read. What you write in the “fixed text prepended to prompts” field under dashboard settings is stored only in your engine’s settings file; it is not stored in the dashboard database.

In Air-gapped, all of the above stays on your servers. The only thing that comes to Linklbee is the monthly report file you check and export yourself (human-readable JSON). That file holds the period · licence reference · engine version · price-table date · scale (eligible documents · baseline documents and share · request counts) · cost per token for both arms · requests, cost and savings per route · the billing basis (savings and the 95% lower bound) · per-model totals for reconciling with your bill · integrity data (ledger line count · hash chain · list of files read) and a signature. No text, document names, keys, request timestamps or company name.

Check it yourself

You can verify the local processing design and the public source.

# count lines in the scanner that mention networking grep -cE "ht[t]p|sock[e]t|url[l]ib" linklbee_scan.py 0

FAQ

What people ask first about anything in the request path.

Does it add latency?

🧪 Linklbee’s own processing overhead, measured against a mock AI provider (medians): +0.33 ms when passing a request through, +4.8 ms when regenerating in full and +20.8 ms with selective regeneration. This is the time the engine adds, not the full API response time. When it groups requests without a document ID, the check adds a median 3–9 ms per request, p99 at most 23 ms (stores of 1,000 to 6,000 past answers).

Streaming on eligible calls goes out after the answer is assembled, so the first token arrives later.

When an edit fails, the vendor is called once more. That request takes one more vendor call longer.

Do our requests pass through Linklbee’s servers?

No. Linklbee runs on your server and requests go from there straight to the vendor. You can also point it at an internal LLM gateway upstream. It listens on 127.0.0.1 by default and has no TLS of its own; to accept from other machines, put a reverse proxy in front.

Can we turn logging off?

Selective regeneration uses previous responses, so response storage is required. Storage is one store directory on your disk; delete it and it is gone. There is currently no setting that disables only the request-path records.

Does it run in our infrastructure?

Yes. Python 3.9+, standard library only, no pip install. In Air-gapped, issuing Linklbee keys, loading past logs and savings reports are all done from the command line on your server.

We already use prompt caching.

Even with caching on, the scanner has things to check: requests too short to be cached, prefixes that change every time and break the cache, and your real hit rate.

Selective regeneration cuts output, and at the rates used in this experiment output tokens cost more than input tokens: in the price table we measured with (claude-sonnet-5-5, official prices checked 2026-10-08) output costs 5× input, and the cache discount applies to input only.

How do we remove it?

Restore the engine address (base_url) and the Linklbee request header to their original settings to disconnect. The request format is not changed.