Trust Security notes

Security notes: what AnchorCore sends, stores and gates

This is the long version of the trust section on the landing page. Each statement points at the file in the public repository that backs it, so you can check it rather than take our word. Paths are under rust/crates/anchorcore/src/ unless noted. Where something is a limit rather than a protection, it is listed under Known limits.

AnchorCore is a private AI search app: it indexes your files and connected sources on your machine and returns cited answers. It has not been through a SOC 2 or ISO 27001 audit or a third-party penetration test, and it has no SSO or role system. Nothing here is a certification.

Every outbound network call

With no connector and no cloud key configured, the app makes no outbound connections: indexing, keyword search and local answers need no network. The calls it can make are the ones below; every HTTP client in the source is one of these.

CallGoes toWhenWhat it carriesEvidence
Update checkapi.github.com (fixed release URL)When you click "Check for updates". A check at startup is off unless you enable it (ANCHOR_UPDATE_CHECK_ENABLED).A plain GET with the user agent anchorcore/<version>. No index content; GitHub sees your IP address and the version.system.rs, docs/updates.md
Local Ollamalocalhost:11434 (setting ollama_base_url)Embedding during sync and search, local-mode answers, health probes, and when you ask the app to pull a model. Indexing itself uses no chat model.Chunk text and queries (embed), question and passages (answer). Stays on your machine unless you point the setting at another host.ollama.rs, embedder.rs, answer.rs
Cloud answer modelYour OpenAI-compatible answer_base_urlLocal answer mode only, and only if you set a key (or point it at a local endpoint). With a non-local endpoint and no key, no request is made. In agent mode the app sends nothing here: your agent writes the answer.Your question, the retrieved passages, and up to the last six turns of the conversation. A follow-up rewrite sends the recent turns and the question. The opt-in acronym expansion (acronym_expansion) sends only the unknown acronyms from your question.answer.rs, query_expand.rs
Cloud embedderYour embed_base_url (falls back to the Ollama URL)During sync, in the background backfill (at startup and every 15 minutes), and for each question when searching by meaning.Chunk text with its heading path, and your query text.embedder.rs, scheduler.rs
Settings "Test connection"The endpoint you are testingWhen you click it.A one-word test message ("ping"), with your key as the Authorization header.settings.rs
JiraThe base_url you enterOn sync. Enabled Jira sources also poll on a timer (default every 15 minutes).Your email and API token as credentials; issues come back.connectors/jira.rs, scheduler.rs
Linearapi.linear.app/graphql, or a base_url you setOn sync.Your API key; a GraphQL query for issues.connectors/linear.rs
Google Drivewww.googleapis.com/drive/v3On sync.The access token you provide; file listing and download for the configured folder.connectors/gdrive.rs
Miroapi.miro.com, or a base_url you setOn sync.The access token you provide; the configured boards and their items, including images on the board.connectors/miro.rs
JSON API (REST)The base_url and list_path you enterOn sync.The auth you configure (none, bearer, basic or a custom header) and paging parameters. See Known limits.connectors/rest.rs
Agent bridge (MCP, CLI)The local app, 127.0.0.1:8000 by default (ANCHOR_BACKEND_URL)When a coding agent calls anchorcore mcp, or you use the CLI.The agent's query and the cited result, between two processes you run. Not a new destination unless you change the URL.mcp.rs, cli.rs

Two further notes. The web interface loads no third-party scripts, fonts or analytics; its only external addresses are links you click, such as the releases page and Ollama's download page. And Ollama's own model downloads happen inside Ollama: AnchorCore only asks your local Ollama to pull a model. The source contains no analytics, usage-tracking or crash-reporting code.

Source labels and the cloud gate

Every source carries one label: internal (the default), public or sensitive. (Older versions had a separate pii label; it now migrates to sensitive.) A provider counts as approved when it is a local endpoint (http:// to localhost, 127.0.0.1 or ::1, or the built-in embedder), when its host is listed in ANCHOR_TRUSTED_PROVIDERS, when ANCHOR_CLOUD_TRUST=1 is set, or when you switch it on in Settings → Sensitive sources. A switch is bound to the host it was given for and closes if you point that setting at a different host; every change is recorded as a system event. AI agents (MCP or the skill) are a separate destination with their own switch, and a local answer model does not open it. Evidence: sharing.rs (allowed), common.rs (provider_is_local).

LabelLocal modelApproved cloudUnapproved cloud
internal, publicYesYesYes
sensitiveYesYesHeld back
  • Answers. Sensitive hits are removed before passages go to an unapproved cloud answer model, and a system_events warning records how many were excluded. If nothing survives, the app says relevant knowledge exists but cannot be used, instead of guessing. (answer.rs: gate_answer_hits, record_gate_event.)
  • Indexing. Indexing uses no chat model at all. Sensitive content skips an unapproved cloud embedder, and the background backfill leaves their chunks unembedded. They stay keyword-searchable. (pipeline.rs, scheduler.rs.)
  • Follow-up rewrite. Skipped when the answer provider is not approved. (answer.rs.)
  • Public answers. POST /qa/public answers only from sources labelled public, and the MCP public_only option calls that path. (answer.rs, main.rs, mcp.rs.)

The gate keys on the source label and the per-document mark below. Both are decisions a person makes, so the gate is only as good as the labels you picked.

Marking one document Sensitive

There is no automatic detection of personal data. A person can mark a single document Sensitive inside a source that is not (PATCH /items/:id with sensitive_override: true; null clears it). The mark only escalates: it cannot make a document in a Sensitive source less protected. A document is Sensitive when its source is labelled sensitive or it carries the mark, and every gate above reads that one rule. The mark survives backup. (common.rs effective_sensitive, items.rs; the decision is recorded in rust/docs/decisions.md.)

Where credentials are stored

  • Connector credentials (token, API key, password, client secret) are stripped out of a source's config before it is saved, so the database row holds none, and reading the config back returns them masked. The cloud answer and embedder API keys are stored the same way. (secrets.rs, settings.rs.)
  • The store tries the operating system keychain first, which is the credential store of macOS and Windows. If that is unavailable it writes one encrypted file per key in your data folder (secrets.enc.*), with the encryption key in a file beside them (permissions 0600 on Unix). That guards against casual reading of a copied file, not against someone who can read your whole data folder.
  • We have not verified keychain behavior on Linux, so we make no claim about it; expect the encrypted-file fallback there unless you have tested otherwise.
  • The API masks stored secrets as ***set*** when a source's configuration is read back. The log is anchorcore.log in your data folder.

Who can reach the app

  • Personal. The app binds 127.0.0.1 by default and has no sign-in, so anyone using your computer account can open it. Requests are checked against a Host/Origin allowlist (localhost plus ANCHOR_CORS_ORIGINS) and carry a CSRF token. If you start the Personal app with --host 0.0.0.0 yourself, it does not stop you, and there is no password unless you set ANCHOR_AUTH_SECRET. (main.rs, security.rs, auth.rs.)
  • Team (self-hosted). The container listens on 0.0.0.0:8000 behind your own proxy; the app speaks plain HTTP, so terminate TLS in front. With ANCHOR_AUTH_SECRET set, password sign-in is on (PBKDF2 password hashes, HS256 session tokens, five failed attempts lock that key out for 60 seconds). Without it the Team edition refuses to start on a non-loopback address. There are no roles: every signed-in user sees the same index. See For teams and hosting/README.md.
  • The Team license is verified offline at startup; no license server is contacted. (license.rs.)

What is recorded

The app keeps a system_events table in its database. It receives gate refusals, sync errors, retrieval degradations and one entry for every agent ask and search call over MCP (component mcp, with the first 80 characters of the query and the hit count). The table is local and is not sent anywhere. (answer.rs, pipeline.rs, mcp.rs.) It is a log, not a tamper-evident audit trail: anyone with write access to the database can alter it.

Backups

An export is one .anchorcore-backup file. Source credentials are left out unless you opt in with a passphrase, which puts them in a separate encrypted bundle; provider API keys are never exported. Imports can be previewed with a dry run first. The archive itself is not encrypted, so store it as you would the data it contains. Details: docs/backup.md.

Verify your download

Each release carries a SHA256SUMS file next to the downloads (the macOS DMG and the zips, generated in .github/workflows/release.yml). Put it in the same folder as your download, then:

macOS / Linux:  shasum -a 256 -c SHA256SUMS --ignore-missing
Windows:        Get-FileHash .\AnchorCore-windows.zip -Algorithm SHA256
                (compare with the matching line in SHA256SUMS)

The checksum shows the file is the one the release published; it does not prove who built it. If that matters to you, build from source.

Signing status. The macOS app is signed with a Developer ID and notarized by Apple, so Gatekeeper opens it without asking. The Windows and Linux builds are not code-signed. Nothing is signed through an app store. That is why first launch needs "Run anyway" in Windows SmartScreen. Release notes: docs/release-notes.

Known limits

  • The JSON API connector has no address guard. It sends requests to whatever address you enter, including localhost and private network addresses, and a full URL in list_path is used as-is. There is no SSRF or private-address protection. Enter only addresses you trust, and treat anyone who can edit sources as able to make the app send requests from your machine or server. The Team edition has no roles to restrict who that is.
  • Personal has no sign-in; it relies on being bound to your own machine.
  • The gate works on labels and marks people set. Nothing is classified as Sensitive automatically.
  • Once passages go to a cloud model you approved, that provider's own handling applies; AnchorCore cannot see or control it.
  • The index is a SQLite file that AnchorCore does not encrypt at rest. Use full-disk encryption if the device can be lost.
  • No SOC 2, ISO 27001, SSO, role-based access control or third-party penetration test.

Reporting an issue

Open an issue on GitHub. If a statement on this page no longer matches the source, that is a bug in this page; tell us. For data this website itself collects (the waitlist and quote forms), see the privacy notice.

Download, free for personal use For teams