Private AI search · runs on your computer
Ask your data. Get the answer and its exact source.
Connect folders, Drive, Jira or Linear and ask in plain language. AnchorCore finds the decision, who made it and why.
- Your index stays on your machine
- Every answer cites file, section and lines
- macOS, Windows, Linux · no Docker, no account
User: Who owns the billing provider migration, and how do we roll it back?
Tool call: anchorcore search (MCP), query: "billing provider migration rollback plan owner", k: 8
Result: 8 results
Answer: Sarah owns the billing provider migration [S1]. Rollback is a feature flag that toggles a tenant back to the legacy provider [S1], and the runbook says it can flip any tenant back within five minutes [S2]. The parallel run is write-only, with no customer impact [S2], and Sarah monitors it daily and can escalate via the rollback flag [S3].
[S1] decisions/d2-billing-provider-approval.md:1-18:"Rollback plan: feature flag toggles back to the legacy provider per tenant." "Owner: Sarah."
[S2] specs/billing-migration-runbook.md:6-20:"mirror invoices to the new provider (write-only; no customer impact)" "the flag can flip any tenant back to legacy within five minutes"
[S3] meetings/m2-q3-planning.md:23-27:"Sarah will monitor daily and escalate via the rollback flag if needed."
search ran on your machine · the answer is written by Claude Code from the results
POST /qa/search
{"query": "billing provider migration rollback", "limit": 2}
200 OK {"hits": [ {"path": "specs/billing-migration-runbook.md", "breadcrumb": "Runbook: Billing provider migration › Steps", "line_start": 6, "line_end": 20, "provenance": [{"leg": "keyword", "rank": 1}, {"leg": "vector", "rank": 1}]}, {"path": "decisions/d2-billing-provider-approval.md", "line_start": 1, "line_end": 18, …} ]}
Answer
According to [S1], the billing provider migration is owned by Sarah.
To roll back the migration, the feature flag toggles can be flipped back to the legacy provider per tenant, as stated in [S2]. This process assumes that the rollback works until the legacy provider is decommissioned.
Sources · 3
- S1Decision: Billing provider migration approveddecisions/d2-billing-provider-approval.md · lines 1–18
“Rollback plan: feature flag toggles back to the legacy provider per tenant.”
- S2Runbook: Billing provider migration › Stepsspecs/billing-migration-runbook.md §2 · lines 6–20
“the flag can flip any tenant back to legacy within five minutes”
- S3Runbook: Billing provider migrationspecs/billing-migration-runbook.md §1 · lines 1–4
“Approved in decisions/d2. Owner: Sarah.”
How it works
Connect what you already have
A folder or Obsidian vault, Google Drive, Jira, Linear, Miro, or any JSON API. PDFs, Word, Excel, Markdown and diagrams are read as they are. It re-syncs on its own; nothing moves.
Indexed by section
Documents are split at their headings, and decision records keep their owner, status and what they replace. All of it sits in one database on your disk.
Answers point at the passage
Ask in the browser or from your coding agent. Every answer lists its sources down to the section and line range, so you can open the page and check.
Sections tree§3 · §4 · 3.2
- section§3 Billing migration3 passages · webhooks
- section§4 Instant payments2 passages · payouts
- under §33.2 Parallel-run planpath: §3 Billing › 3.2 Parallel-run plan
How a search runs
topics, then section summaries, then passages
~7 ms per search on a 30k-chunk benchmark
Headings become a tree, so a citation can say §3.2 of a 300-page PDF.
Who writes the answer
A local model, in the app
Download, open it, add a folder, ask. A local model writes the answer, or switch to agent mode and the app shows the cited passages your agent gets. It runs at 127.0.0.1:8000 with no account and no Docker. Try it on the sample company first.
Your coding agent
Claude Code, Codex, opencode and Cursor search your files over MCP or the local REST API and write the answer themselves, so the app needs no chat model. The setup skill installs the app, connects folders and checks the first cited answer.
npx skills add Atsirkunov/anchorcore --skill anchorcore-setup
What leaves your machine
Your index stays on your machine, and answers are generated locally unless you connect a cloud model. Below is everything the app can talk to, read from the source. Full security notes.
~/.anchorcore)
Your folders
Local Ollama at localhost:11434
- defaultNothing else. No connections until you enable a connector or add a cloud key. Indexing folders, search and local answers need no network. Enabled Jira sources re-sync on a timer (default every 15 minutes), and the embedding backfill (at startup and every 15 minutes) calls only the embedding endpoint you configured.
- only when you clickapi.github.com: "Check for updates" fetches the newest release tag. Nothing about you or your knowledge is sent. An optional once-at-startup check is off by default.
- only if you connect itJira, Linear, Google Drive, Miro, a REST API: each connector sends requests to the URL you configure for it (the REST connector to any address you enter, including local ones), with credentials you provide (kept in your OS keychain on macOS and Windows, otherwise in an encrypted file in your data folder).
- only if you add a keyA cloud model you choose: an OpenAI-compatible endpoint for answers or embeddings. Your question and the retrieved passages go to that provider. With no key set, no answer request is made.
- only when you pull a modelOllama's model registry: your local Ollama downloads models itself; AnchorCore only talks to Ollama on your machine.
No telemetry
The app contains no analytics, usage tracking or crash-reporting code. The source is public, and the only outbound calls are the ones listed here. Read the source.
Works offline, within limits
With a local Ollama model and folder sources, everything runs with Wi-Fi off. Connectors, cloud models and update checks need a network.
Sensitive sources are gated
| Source label | Local model | Approved cloud | Unapproved cloud |
|---|---|---|---|
| internal (default) | ✓ | ✓ | ✓ |
| public | ✓ | ✓ | ✓ |
| sensitive | ✓ | ✓ | ✗ held back, audited |
"Approved" means a local endpoint, one you list in ANCHOR_TRUSTED_PROVIDERS, ANCHOR_CLOUD_TRUST=1, or one you switch on in Settings → Sensitive sources. Sensitive sources also skip an unapproved cloud embedder, and AI agents reading through MCP or the skill need their own switch. A single document can be marked Sensitive too; the mark gates it like the label. Public sharing and agent public_only mode see only sources you marked public, minus anything Sensitive. Nothing is detected as personal data automatically, and there is no SSO or role system in the local app.
Why I built it
Every team I worked on lost the why behind its biggest decisions: in Slack threads, old tickets, and the heads of people who left. AnchorCore is the tool I wanted: ask in plain language, get the answer, and see the exact source. It runs on your machine, because the data I worked with couldn't go anywhere else.
Guides
The sample company
Ten files about a fictional company. Connect them and ask your first cited question in a minute.
Your Obsidian vault
Point AnchorCore at the vault folder. Your notes stay where they are.
Any JSON API
Tickets, a CMS, your own backend: map its fields, preview five records, sync.
All guides, including the AI startup sample and backup and migration.
FAQ
Do I need Ollama?
No. Without a model you still get full-text search, and in agent mode your coding agent writes the answer from AnchorCore's cited results. Add Ollama (one install, free models) to get answers inside the app.
How is this different from NotebookLM or ChatGPT file search?
Those are cloud uploads. AnchorCore runs on your machine, connects to folders, Drive, Jira, Linear, Miro and APIs, and keeps sources you marked sensitive away from AI you haven't approved. Full comparison with NotebookLM.
Does it stay fast as my library grows?
It narrows thousands of documents to the few relevant sections via tags and summaries, then searches only those. Measured at ~7ms per search on a 30k-chunk benchmark (results, script); larger libraries are projected, not yet measured.
Can I move to a new computer?
Yes. System → Backup & restore exports your whole index into one portable .anchorcore-backup file. Credentials stay behind unless you opt into the passphrase-encrypted bundle. Backup guide.
Is it open source?
It is source-available: the repository is public, so you can read what the app does or build it yourself. Personal use is free; team and commercial use need a license. See LICENSE.md.
Pricing
Personal
The local app on your machine, with every connector and cited answers.
Team, self-hosted
One Docker container in your VPC, the same cited answers for the whole team. Request a quote.
Download
macOS, Windows and Linux. One download, no Docker, no account.
First launch
macOS: open the DMG, drag AnchorCore onto Applications and double-click it there. The app is signed and notarized by Apple, so it opens like any other Mac app. Windows: the build is not code-signed, so SmartScreen asks once: More info → Run anyway.
Verify the file
Each release has a SHA256SUMS file next to the downloads. Put it in the same folder as your download and run shasum -a 256 -c SHA256SUMS --ignore-missing (on Windows, compare Get-FileHash with the matching line). Release notes.
We only use your email for AnchorCore updates; unsubscribe anytime. Privacy notice.