Skip to content
MarinaMarina

Data, privacy and terms

What Marina stores, what it sends where, and how to turn each outside source off. Marina is self-hosted: the operator who runs it decides what it connects to.

No telemetry

Section titled “No telemetry”

Marina sends no usage data, analytics or crash reports to H2O.ai or anyone else. There is no phone-home and no update check in the server.

Data leaves your machine only through things you configure:

  • Model providers you set keys for (OpenRouter, Anthropic, OpenAI, a local server, …) receive the prompts, context and tool results each model call carries. Their own terms apply.
  • Trace and log export (OTLP) goes only to the collector you name in OTEL_EXPORTER_OTLP_ENDPOINT (or its per-signal variants). Unset, nothing is exported. See Execution traces.
  • Optional services you enable with a key or URL: search APIs, media providers, TabH2O (TABH2O_API_KEY, the only H2O.ai-hosted service Marina can call), chat adapters, federation peers, connectors and webhooks such as MARINA_OPS_NOTIFY.
  • Work you ask for: web searches and page reads (below), and package installs from the npm registry when a Code Mode task installs dependencies.

One request does not come from the server: the dashboard page loads its web fonts from Google Fonts in the viewer’s browser.

What is stored, and where

Section titled “What is stored, and where”

Everything Marina keeps is local to the server:

  • The world database: one SQLite file (DB_PATH, default marina.db; under MARINA_DATA_DIR when set). It holds accounts, rooms, messages, memory records, tasks, the chronicle, ledgers and traces.
  • Assets: uploaded and generated files under ASSETS_DIR (default data/assets).
  • Scratch and caches: MARINA_SCRATCH_DIR and model caches, under the same data directory.

See Configuration for the paths and Deployment for running with persistent volumes.

Erasure and retention

Section titled “Erasure and retention”
  • Memory. note delete retires a note (it stays in history). The memory service’s forget erases records, sources or a whole space, including their history. See Standalone memory service.
  • Accounts. Erasing an account deletes the rows keyed to it, anonymizes rows other people depend on (tasks it created, board posts) and records one audited chronicle event with counts, never the name. See persistence.
  • Retention. An hourly pass prunes telemetry, ledgers and staged request images by age; MARINA_RETENTION_OVERRIDES changes the periods. Civic history (the chronicle, standing, submission and benchmark ledgers) is append-only and never pruned. See row retention.
  • Backups are outside erasure. Erasing data does not reach snapshots or copies you made earlier (Backup and recovery). Rotate or delete them to complete an erasure.

What Marina fetches from the web

Section titled “What Marina fetches from the web”
Source When Turn off
Search APIs (Tavily, Exa, SearXNG, DeepSeek, OpenRouter web search) Only with their key or URL set Unset the key or URL
DuckDuckGo (lite page and instant answers, keyless) web search and research when no keyed backend answers MARINA_SEARCH_DUCKDUCKGO=off
Academic and dated sources (arXiv, PubMed, Semantic Scholar, GDELT, Wikipedia, Hacker News, Wayback Machine) Academic and date-bounded searches Name other engines; see Search
Yahoo Finance chart data (unofficial, keyless) Statistical priors for stock and index series MARINA_FORECAST_YAHOO=off
Polymarket and Kalshi prices (keyless) Forecast lookups, on by default: questions a market prices MARINA_FORECAST_LOOKUPS=off, or name only the lookups you want (see Lookups)
FRED and BLS series (keyless; a key raises the quota) Forecast lookups, on by default: questions about an official series MARINA_FORECAST_LOOKUPS=off, or name only the lookups you want
Kalshi and Polymarket market lists (keyless) About once a minute while the markets or prediction-lab world runs (its live feed rooms) No switch: run another world
Third-party pages web fetch, the read swarm, research page reads, citation checks MARINA_ROBOTS_TXT (below)

Page reads honour robots.txt. Before reading a third-party page for an agent or a person, Marina checks the site’s robots.txt for its user agent, MarinaResearch (Mozilla/5.0 (compatible; MarinaResearch/1.1; …; +https://github.com/h2oai/marina)), and sends that same user agent with the read. A disallowed page is not requested: the read answers “blocked by robots.txt”. robots.txt is cached per site for 24 hours. When a site answers 4xx it has no rules; when robots.txt cannot be fetched (5xx, timeout), the read goes ahead and the check retries in 10 minutes. API calls (search, model, market and data APIs) are not page reads.

MARINA_ROBOTS_TXT=ignore turns the check off. MARINA_ROBOTS_TXT_IGNORE_ORIGINS lists sites you own or may read regardless (https://docs.example.com,intranet.example.com). Every web request, robots.txt included, goes through the same guard that refuses private and internal addresses. readiness (check search) shows the current settings.

Markets, competitions and forecasting questions

Section titled “Markets, competitions and forecasting questions”
  • Prediction markets are paper trading by default. Live orders need MARINA_TRADING_ENABLED=true and your own venue credentials; you are responsible for the venue’s terms and your jurisdiction’s rules. See Prediction markets.
  • Public competitions and benchmarks are opt-in operator tools. They never submit on their own and are never enabled by default. Each adapter follows its board’s rules. See Public leaderboards.
  • Political questions in the arena (presidential approval and similar) are forecasting research on public data. Marina forecasts what a published number will be and takes no position. See Social Simulation Arena.