Skip to content

What Prism stores and sends

Prism runs on your machine. Requests go from your clients through Prism to the model providers you connected. Nothing goes to the Prism project unless you turn it on, with the exceptions listed under Requests to the Prism project.

Everything lives in the storage directory ([storage] dir; the installer uses ~/.prism/data, the Debian package /var/lib/prism).

What Where Contents
Request journal routing.db One record per model request: time, model requested and used, provider, strategy, tokens, cost, latency, status, the X-Prism-App and X-Prism-Session values you send, the client family inferred from the request. No request or response text.
Usage totals usage-cubes/ The console’s day, week, month and all-time usage, counted from the request journal and kept so a restart does not count it again: per person, team and time slot, the request, error and token counts, cost and latency, by model, application and provider. No request or response text. Deleted files are counted again from the journal.
Routing decisions routing.db Per request: the candidate models, which one was chosen and why, what the alternative policies would have chosen. No text.
Outcome verdicts routing.db Per chat turn: worked, not worked or unknown, judged from the next turn. Prism compares turns by hashes held in memory for 10 minutes; no text is written.
Observed applications observed-apps.json Applications recognized from traffic: client family, device name, request count, recent model names, and truncated SHA-256 hashes of the system-prompt start, tool names and the client’s session marker. No prompt or tool definition in clear.
Learned model capabilities observed-capabilities.json Per provider connection and model: whether tools, JSON output or image input worked or was refused, and a context limit the provider named. No request or answer text. An entry expires after 30 days; the store keeps at most 4096 models
Provider keys added in the console provider-secrets/ Encrypted with the key in provider-secrets/master.key
Console accounts and connections accounts.db Users, email addresses, passkeys, provider connections
API keys issued to clients and agents routing.db, accounts.db A SHA-256 digest of each key, its policy, label and expiry; never the key itself
Notifications notifications.db Provider and model events, read markers; kept 30 days, at most 4096 events
Learned parameter fixes learned-param-rules.json Per provider and model, which request parameters to drop or rename, with a short excerpt of the provider’s error message; each rule expires after 7 days
Instance identity instance-identity.json, instance-identity.record The instance’s Ed25519 key pair and its fingerprint (used only for a paid subscription)
Model quality prior quality-prior.json, quality-priors/ The public prior downloaded from the Prism project, and the active prior paired with your connected models. Public data, nothing about your traffic

The request journal, routing decisions and outcome verdicts are deleted after [storage] log_retention_days (default 90). Prism’s own log lines go to standard output (journald under systemd); they carry metadata, not request or response text.

Do not put personal data in X-Prism-App or X-Prism-Session: Prism stores those values as sent.

A sign-in request from an app is kept in memory for ten minutes and is not written to disk: the name and client string the app sent, where the request came from, and a SHA-256 digest of its device code. An allowed sign-in creates a key for the person who allowed it. Like every key, it is stored only as a digest, here with the name the app gave itself, its client string and the address the request came from; see Sign in from an app.

These go to your providers, with your keys, at your cost.

Request When Turn off
Model discovery: GET /v1/models At startup, on Refresh, and when a provider answers that a requested model does not exist (at most every 10 minutes per provider) —
Health checks: GET /v1/models, GET /health Every 60 seconds ([health] check_interval_seconds) —
Funding check for a provider that answered “payment required” At most once per 900 seconds per account —
Role target check: one request with the message ping and max_tokens 1 to a target of a [routing.role_aliases] role that no live request has answered in the last six hours, typically a backup target At most every six hours per target, the first one six hours after start; a failed check is repeated once ten minutes later to confirm it Off unless [health.role_check] enabled = true
Capability probe: one small request per capability (tools, JSON output, image input) to a model whose provider, your settings and your traffic have not said whether it has that capability; what the answer shows is kept for 30 days In the background, up to 10 models every 10 minutes, newest connections first; a model connected while Prism runs is checked 30 seconds later [catalog] capability_probes = "off"; "api" (the default) checks API-key and local connections, "all" also subscription (OAuth) accounts, whose quota the checks use
Prompt cache keep-warm for Claude conversations (Anthropic, Claude subscription, Claude on OpenRouter): a one-token request that re-reads the cached conversation prefix shortly before the one-hour cache expires For 180 minutes after a conversation’s last turn ([harness] prompt_cache_keep_warm_minutes) prompt_cache_keep_warm_minutes = 0

Role target checks are not written to the request journal and are not counted in Prism’s spend reports or budget limits; the provider still bills them, and /prism/status shows their spend for the day under role_state.check. Each role’s state (normal, on a backup target, or dead) is shown under role_state and in the console under Diagnostics whether or not the check is on; it is computed from what Prism already knows. See the health configuration reference.

Request Address When Turn off
Model price catalog (LiteLLM’s public price file) raw.githubusercontent.com At startup and every 24 hours ([pricing] refresh_interval) [pricing] auto_refresh = false; Prism then uses the catalog embedded in its build, and models released after the build have no price

The download sends nothing about your installation. When it fails (for example behind an outbound allowlist), Prism logs a warning and keeps the last downloaded or the embedded catalog.

Request Sent to When Contents
Update check releases.prism.tm When the console or prism update check looks for a new release A plain download; no identifier, account count or statistics
Model class catalog and subscription plan prices releases.prism.tm Once a day A plain download of two public JSON files
Model quality prior (Prism Model Exam, CC BY 4.0) releases.prism.tm Once a day; [catalog] quality_prior_url = "off" turns it off A plain download of a public JSON file and its signature; no identifier
Subscription checks (Pro, Teams, Enterprise) license.prism.tm Only after the owner binds a subscription; see Statistics and subscription checks Instance fingerprint, Prism version, number of active human accounts and teams
Anonymous usage statistics license.prism.tm Off by default; at most once a day when on Version, OS, architecture, user-count range
Experiment metrics eu.i.posthog.com (PostHog EU) Off by default; only for people the owner admitted who consented themselves; see Experiments and experiment metrics Experiment results (outcome, model, latency, cost) under a pseudonym; no text, names, emails or IDs
Install signal prism.tm Once, from the installer, only when the install link carried a token; see Install signal Token, event, version, OS, architecture, install mode

Fleet telemetry is a separate opt-in export to a receiver you name; see Fleet telemetry.

  • Stop Prism and delete the storage directory to remove everything.
  • prism backup create makes a copy first; see Backup and migration.
  • Turning a feature off stops new collection; it does not delete what is already stored.