Repository navigation
Releases
v1.7.0
AnchorCore 1.7.0
The app has a new layout — Library, Agents, and a Settings rail with
everything reachable by link — and a whole-app reliability pass means
operations that used to report success while doing nothing now fail
loudly instead.
Highlights
- New layout: Ask · Library · Records · Agents · Settings. The
Library (was Sources) lists your sources first, each with its own
page showing documents and recent jobs. The Agents tab gathers
everything about connecting a coding agent: one-click MCP setup,
the skill files for when MCP is blocked, and the switch for whether
agents may read Sensitive content. Settings is now a rail of
sections — Answers, Search, Sensitive sharing, Glossary, Backup,
Updates, Status & logs, Appearance, Advanced — with the old System
tab folded into Status & logs. - Every tab has an address. Tabs live in the URL hash, so reload,
Back, bookmarks, and links from warnings all work. Old#/sources
and#/systemlinks redirect to their new homes. - Ask keeps its conversation. Switching tabs no longer erases your
questions and answers, and the project scope moved from the header
into Ask itself, next to a plain "Public sources only" option. - No answer model, no panic. Without a working answer model, Ask
now shows the matching passages as evidence cards with one calm
setup hint — instead of four warnings and raw markdown cut off
mid-word. - Calmer look. Dark mode is a neutral cool black instead of warm
brown, links are ink with oxblood reserved for citations ("this is
sourced"), and small caps labels are plain text. The website now
opens in Light with the theme switch in its footer. - Starter questions from your own records. The sample questions
only show for the bundled sample corpus; on your own library Ask
suggests questions built from your newest decision and newest open
action — and clicking one runs it. - Quieter failures, fixed. A full review of error handling: source
delete and update are now all-or-nothing (a failed write answers
500 instead of 200 with half the change applied), marking a
document Sensitive, renaming records, and saving settings all
report real failures, and the job queue no longer loses finishes.
The Sensitive gate now fails closed: if a label can't be read, the
content is treated as Sensitive. - RTF files read correctly, including non-English text — and any
RTF indexed before this release is automatically re-read once on
its next sync, so garbled text repairs itself.
Fixed
- RTFs with accented or non-Latin text came out garbled
("Café" indexed as "Café"); underline anduccontrol words also
misread. Non-English RTFs now index as written. - Ask timed out after 30 seconds while the server was still
working (it allows two minutes or more). Ask now waits up to three
minutes and only says "cancelled" when you actually cancel. - Every write failed with 403 until reload after a server
restart. The app now refreshes its expired token and retries
once, on its own. - Marking a document Sensitive could answer 200 while leaving it
unprotected. It now succeeds or fails as one unit, loudly. - Stale claims corrected. The website no longer advertises the
oldpiilabel (there is one Sensitive concept plus a
per-document mark), the MCP card lists all the tools, backups no
longer mention the removed review queue, and footers no longer say
"v1 local".
Upgrade notes
- One new database step on first start: "adding extractor
generations". It stamps each document with the version of the
reader that extracted it. Instant; nothing in your library changes.
Afterwards, RTFs indexed by older versions re-extract once on
their next sync — only the file read is repeated, and files whose
text is unchanged are left alone. - The
/review/*API is removed. The review queue was dropped
earlier; record disputes and restores live in the Records tab and
the/decisions+/actionsendpoints. Agents still calling
/review/*need to move. GET /sourcesnow includes each source's config (secrets
masked), so clients no longer need one request per source.
v1.6.6
AnchorCore 1.6.6
Large libraries upgrading from before 1.5.0 start much faster, and the
search skill offers a lasting fix for the sandbox question.
Highlights
- The search skill offers to stop asking. The first time a sandbox
blocks the skill, it asks once whether you want to allow AnchorCore's
address (for example127.0.0.1:8000) in your sandbox settings. In
Claude Code you do that with/sandbox. The skill never runs it for
you: say yes and it tells you the exact address to allow, say no and it
won't bring it up again this session.
Fixed
- The first start after upgrading a large, older library no longer
sits on "cleaning up legacy entities" for minutes. The one-time
cleanup of old extracted entities (libraries last opened with a version before 1.5.0) looked
up related rows one table scan at a time. It now uses indexes. In our
test it went from about three minutes to well under a second, with the
same rows removed.
Upgrade notes
- One new database step on first start: "indexing records". It adds
indexes so that deleting records (during that cleanup and when a
document is reclassified) no longer scans the whole search index. It
normally takes seconds. Nothing in your library changes. - Refresh the installed skill to get the new sandbox offer. In the
System tab, click Install… on the Ask docs skill and confirm.
v1.6.5
AnchorCore 1.6.5
The Entities tab is now Records, a plain list of your decisions,
actions and glossary. The search skill also gets better at telling "the
app is closed" from "my sandbox is blocking it".
Highlights
- Records replaces Entities. The tab opens on Decisions and also shows
Actions and the Glossary. You can Dispute… a decision or action that
is wrong (a reason is required, and the history is kept) and Restore
it later. A Disputed filter shows the ones you've set aside. A
disputed record drops out of the Decisions and Actions lists and out of
what agents get from thedecisionsandactionstools.
Fixed
- The search skill no longer tells you to open an app that is already
open. When an agent runs in a sandbox (Claude Code, for example), a
blocked call to AnchorCore looks just like a closed app, so the skill
always said "open AnchorCore". It now checks whether the app is
actually running. If it is, the skill asks once: "OK to run ask-docs
outside the sandbox for this session?" It only leaves the sandbox if you
say yes. If you say no, it gives you the command to run yourself. - The skill speaks in its own name. Its messages use the name it was
installed under (for example "kb is running on this computer, but my
sandbox blocks local network calls"), not "AnchorCore". When your company
hosts the server, a failed call points you to your network or to
whoever runs it, not to an app on your computer.
Upgrade notes
- "Looks right" is gone. Marking a record as looking right never
changed any answer: records are used unless they are stale or
disputed. The review queue, its confidence badges and its "New / In
answers" labels went with it. Nothing in your library changes. The
entity and review endpoints stay available for REST and MCP callers. - Refresh the installed skill to get the new sandbox step. The skill
is a file in your agent's skills folder, so updating the app doesn't
change it. In the System tab, click Install… on the Ask docs skill
and confirm, keeping the same name or choosing a new one.
v1.6.4
AnchorCore 1.6.4
A fix for large libraries: AnchorCore 1.6.3 could sit on "Checking the
index" for an hour or more every time it started. This release makes that
step take seconds.
Fixed
- Large libraries start again in seconds. On a library of a few
gigabytes, 1.6.3 stopped on the upgrade screen's last step ("Checking
the index") for a very long time, and did so on every start, not only
after upgrading. AnchorCore looked up each document's search chunks
without an index, reading the whole chunk table once per document. It
now keeps an index for that. In our test with a 4.2 GB library
(15,000 documents), 1.6.3 was still not ready after 10 minutes; 1.6.4
was ready in 8 seconds on the first start and in 1 second after that. - Removing or rebuilding documents is faster too. The same missing
index slowed every operation that replaces a document's chunks, such as
re-syncing a changed file or deleting a source, on big libraries.
Also
- The upgrade screen shows how long a step has been running ("This
step has been running for 3 min 12 s") once a step takes longer than
10 seconds, so a slow step no longer looks frozen. The last step now
says "Looking for documents to rebuild" while it checks which documents
need new chunks, instead of staying on "Checking the index".
Upgrade notes
- The first start after upgrading adds the index ("Migration 31:
indexing chunks by document"). That takes about a minute on a large
library; later starts are not affected. - If 1.6.3 is stuck on "Checking the index" right now: quit
AnchorCore and install 1.6.4. It is safe to quit at that step: every
earlier upgrade step is already saved, and nothing in your library
changes while it checks.
v1.6.3
AnchorCore 1.6.3
The Mac app is now notarized by Apple, so it opens with a plain
double-click. Upgrades show what they are doing instead of going quiet,
semantic search is several times faster on big libraries, and you decide
where Sensitive content may go, one switch per destination.
Highlights
- Opens like any other Mac app. The macOS download is now signed with
a Developer ID and notarized by Apple: open the DMG, drag AnchorCore to
Applications, double-click. No more trip to System Settings → Privacy &
Security → Open Anyway. - Database upgrades you can watch. When a new version needs to update
your library, AnchorCore now opens straight away and shows an upgrade
screen with the current step and a progress bar, and the menu-bar icon
says it is upgrading. Before, a large library could sit silent for
minutes with no window and no icon. Before it starts, AnchorCore checks
there is enough free disk space and refuses, without touching anything,
if there is not. Agents asking during an upgrade are told to wait and
retry rather than seeing "not running". - If an upgrade fails, you see why. AnchorCore keeps running and shows
the error in plain words, what state your data is in, where the log is,
and a copy-for-support button. Each upgrade step is saved on its own, so
the next start resumes where it stopped. - Faster semantic search on big libraries. The vector search reads far
less data per query: about 4× faster on a 50,000-chunk library
(340 → 88 ms in our measurement). Results are identical. - Choose where Sensitive content may go. Settings → Sensitive
sources has one switch for each place it could leave this computer: a
cloud embedding provider, a cloud answer provider, and AI agents
connected over MCP or the skill. All are off by default, turning one on
is recorded, and each source row says whether its content stays "local
only" or is "also sent to" a named host. Local providers (the built-in
embedder, Ollama on this machine) need no switch. If you change a
provider's address, its switch turns off until you confirm again.
Upgrade notes
- AI agents and Sensitive sources: check one switch. Agents used to be
allowed to read Sensitive content automatically whenever your answer
model was a local Ollama. They no longer are: an agent that reads
Sensitive documents now gets "gated" until you turn on Settings →
Sensitive sources → AI agents. The AnchorCore app itself always shows
everything. - Admins:
ANCHOR_CLOUD_TRUSTandANCHOR_TRUSTED_PROVIDERSstill
work and show in the card as "set by environment". - No database changes in this release; it starts as fast as 1.6.2.
- Older downloads (1.6.2 and earlier) were not notarized; if you open
one, macOS still asks for Open Anyway once.
v1.6.2
AnchorCore 1.6.2
Your agent now understands questions about your documents in your own words,
the search skill has a shorter name you can change, and the System tab's
errors list can finally be cleared and filtered.
Highlights
- Ask in your own words. Agents pick AnchorCore up for questions like
"what did we decide about refunds?", "who owns the payout migration?" or
"where is onboarding documented?" — no need to say "AnchorCore". This
works over MCP (the server now tells agents what it is for) and through
the skill. Decision questions go to the decision log, owner and open-work
questions to the action tracker, everything else to search, and the agent
reads the passage before citing it. When nothing indexed matches, it says
so instead of answering from general knowledge. - The search skill is now
ask-docs, and you choose its name. Call it
with/ask-docs what did we decide about refunds?. When you install it
(System tab, the welcome wizard, or the setup skill), you can pick a
different handle, such as/kb. Installing under a new name retires the
old copy, so your agent never runs two. - Errors & warnings you can clear. In the System tab:
- Dismiss one row, or clear everything shown, and bring it back any time
under "Show dismissed". - Filter by level, component and source.
- Repeats of the same problem collapse into one row with a count and when
it was first and last seen, and long details expand on click. - Info events (mostly the log of agent calls) are hidden by default.
- Dismiss one row, or clear everything shown, and bring it back any time
Fixed
- The component filter on the errors list missed most components. It
offered a fixed list; it now lists the components that actually have
events. - The skill could ask you to loosen your sandbox when AnchorCore simply
wasn't open. It now tells the two apart: "open AnchorCore" when nothing
is listening, and one question asking for access to just the app's local
address when your agent's sandbox blocks it. It never bypasses the
sandbox without your yes.
Upgrade notes
- Reinstall the search skill to get the new name and behavior: System
tab → Agent skills → Install… (or ask your agent to run the setup skill).
Your existinganchorcore-searchcopy keeps working until then; the
reinstall retires it (renamed toSKILL.md.anchorcore.bak, not deleted).
Restart your agent afterwards. - MCP users need nothing new: restart your agent to pick up the new
routing. - One small database update runs at first launch (a column on the
events table); it takes moments, whatever the database size. - Dismissing never deletes. The events table is also the record of
which questions agents asked, so cleared rows stay stored; they are
just hidden.
v1.6.1
AnchorCore 1.6.1
1.6.0 could not start on a database created by an earlier version: the
upgrade stopped with an error and the app quit at launch. 1.6.1 fixes the
upgrade. New installs were not affected.
Fixed
- Upgrading from 1.5.x or older failed at launch. The 1.6.0 upgrade step
that turns the old per-document PII mark into the new Sensitive override
hit a database constraint (NOT NULL constraint failed: ingested_items.sensitive_override) and the app exited. The failed upgrade
rolled back completely, so your data is untouched — install 1.6.1 and
launch again. Documents you had explicitly marked keep their mark as
Sensitive; everything else inherits from its source, as 1.6.0 intended.
Upgrade notes
- The first launch on a large database takes a while. The upgrade
rewrites the chunk table once (and, coming from 1.4.x or older, also makes
a safety backup and cleans up legacy entities). On a multi-gigabyte
database expect several minutes before the window opens; the app is
working, not hung. Later launches are fast. - Coming from 1.4.x or older and you tried 1.6.0? Each failed launch
left a full copy of your database inexports/inside the data directory
(safety-backup-pre-r45.10-*.db). They are never removed automatically —
keep the newest one until 1.6.1 has started successfully, then delete the
extras to reclaim the space.
v1.6.0
AnchorCore 1.6.0
Sensitive content is now gated by default on every surface: one Sensitive
mark per source (or per document) keeps those passages away from cloud
providers and agents unless you say otherwise, while the app itself always
shows you everything. Agents also get proper read APIs — line windows,
section text, and item reads over REST and MCP — and stopping a running
job finally stops the job you meant.
Highlights
- Sensitive means Sensitive, everywhere. The old separate PII label is
gone: each source is either ordinary, public, or Sensitive, picked in the
source form, with an optional per-document mark through the API. Sensitive
passages are withheld from untrusted cloud providers and from agents
(MCP, skill, API) by default — single reads answer 403, lists and search
drop the rows — while the app UI always shows you everything. The public
scope (/qa/public,public_only) excludes Sensitive items no matter
what. Every withholding is recorded in the System tab's event log. - Agents can read, not just search. New read APIs over REST and MCP:
whole-file and line-window item reads, section text in document order,
section/tag chunk lists, and entity reads with context (list_items,
get_item,get_outline,get_entity). All of them enforce the same
Sensitive gate as search. - A setup wizard that fits your machine. The first-run wizard now
recommends a local answer model sized to your RAM, lets agent-mode users
pick the MCP or skill channel, and falls back to the built-in embedder
when Ollama is down. - Stop the job you mean. Every running or queued job has its own
labeled Stop control (kind + id + source) in the Sources row and the
System tab, plus a header badge that leads there. Stopping lands within
one provider call and reports how many documents were kept. - Faster re-syncs. Tag matching now runs over an indexed per-pass
cache (measured: the tags phase of a big re-sync went from ~73 s to
~4 s), and files with no extractable text no longer re-ingest on every
sync. - Tags you can trust. Single-word tags no longer merge into whichever
tag shares their hash bucket, so the tag list grows with your vocabulary
instead of stalling near 768, and the tag-reuse threshold finally reads
the database the app is actually running on. - Local answers without the memory spike. Chat calls to a local Ollama
now pin the context window instead of letting Ollama auto-size it, so a
3B model no longer loads gigabytes it will never use. - Bigger backup imports. The import size cap is now 2 GiB.
Fixed
- The folder picker no longer renders empty stripes in Chromium-based
browsers: rows keep their height and labels truncate inside them. - Cancelling no longer picks the wrong job when several are active, and a
second cancel (or one racing completion) answers 409 instead of
corrupting job state. - Files whose text extracts empty no longer re-ingest on every sync.
Upgrade notes
- Sources labeled
piibecomesensitiveautomatically on first
start (database migration 29). The per-document mark moved to
PATCH /items/:id {"sensitive_override": true}(nullclears it);
the oldis_piifield answers 400 naming the new one, and old backup
archives import with the mapping applied. - Agents may see less than before. Whole-file reads, entity reads and
chunk lists are now gated for agents under an untrusted provider (they
were open), and public scopes exclude Sensitive-marked documents even
with a trusted provider. If your agent suddenly misses content, that is
the gate working: approve the provider (per-destination opt-in switches
are coming in a later release) or read the same content in the app, which
is never gated. - Tag counts grow with your vocabulary now (see above); links written
under the old merging scheme stay until the item is re-processed.
v1.5.0
AnchorCore 1.5.0
Search no longer needs Ollama. AnchorCore can now embed with a built-in model
that runs inside the app, re-syncing an edited document skips the work that
didn't change, and a "Report issue" button turns a wrong answer into a
shareable, redacted bug report.
Highlights
- Built-in embedding model — no Ollama needed for search. Press
Download built-in model in the setup wizard or in Settings →
Embeddings (about 550 MB to download, about 275 MB on disk). It is the same
nomic-embed-textv1.5 model Ollama serves, running inside the app (Metal
on macOS, CPU on Windows and Linux), so its vectors are identical: switching
between the two never re-indexes anything. When Ollama is running and has
nomic-embed-text, AnchorCore keeps using it, because it indexes large
folders faster; the built-in model takes over when Ollama is not there.
The download only happens when you ask for it, is checked against pinned
checksums, and resumes if interrupted. Offline? Copy the three model files
into the folder the card shows and press the button. Headless:
anchorcore embed-model fetch. New settingembed_provider:auto
(default),builtin, orollama. - The Team image embeds out of the box. The Docker image now ships the
built-in model, so hybrid search works without any Ollama. If the host runs
Ollama,.env.examplenow marks it as trusted so sensitive sources still
embed through it. - Small edits re-sync fast. When a document changes, every passage whose
text didn't change keeps its stored vector. Fixing one character in a large
file now re-embeds only the passages around it (measured: the embedding step
went from 136 s to under a second on a 4,500-passage file). - Report a wrong answer. Every answer in Ask has a Report issue link
that downloads a diagnostics file for that query: which search legs found
which results, their ranks and scores, index counts and versions. It
contains no document text, titles, paths, settings or secrets (documents
and sources appear as pseudonyms), only your question, so you can send it
to support after a glance. Agents get the same file from
GET /system/report. - Strong keyword matches are no longer buried. A passage that matches
several of your query's words strongly (often in a PDF) can now rank above
weaker matches from higher-authority sources instead of being held below
them. - Meeting notes with only an "Action items" section count as records.
Their action items now show up under Actions.
Fixed
- The folder watcher no longer re-scans every folder every second; it reacts
to file events and double-checks once a minute (ANCHOR_WATCH_FALLBACK_SECS).
Idle CPU is near zero, and removed or disabled sources stop being watched. - Starting a second copy of the app no longer opens a browser tab onto the
first one; it exits with "is AnchorCore already running? try --port". - MCP Connect now works when the app runs with a custom
--data-dir: the
client entry gets the same--data-dir, soanchorcore mcpfinds the app. - macOS: the menu-bar icon is now a monochrome glyph that follows light and
dark menu bars, and remembers where you ⌘-dragged it. - Large documents no longer hold the database write lock while being cleaned
and split into passages.
Upgrade notes
- One-time entity cleanup, with a backup first. On first start, 1.5.0
copies your database toexports/safety-backup-pre-r45.10-<time>.db(never
deleted automatically), then removes decisions and actions that older
versions extracted with a language model or from documents that aren't
records, and marks the remaining old ones as unverified (verified means a
person confirmed it). The counts are in the System tab's event log
(componentmigration). Run Reclassify on each source afterwards to
rebuild decisions and actions from document structure. - Embedding batches can now overlap for remote embedding APIs
(ANCHOR_EMBED_CONCURRENCY, default 2). Local Ollama speed is unchanged. - The app binary is a few MB larger (the built-in model's runtime). The model
itself is downloaded separately, only when you ask.
v1.4.1
AnchorCore 1.4.1
Syncing and reclassifying are much faster: tagging used to slow down as your
collection grew, and on large collections it could cost several seconds per
file. It now takes milliseconds, whatever the size.
Highlights
- Reclassify and sync run up to ~40% faster. Tagging a large file dropped
from about 10 seconds to half a second on the measured corpus, and unlike
before, it no longer gets slower the further a long run gets. A run that
crawled at a dozen seconds per file should now move at the speed of its
embeddings instead.
Fixed
- Tagging re-read a setting once per existing tag. Every new tag opened a
throwaway database connection for every tag already stored, so the cost grew
with the collection. The setting is now read once per file. Stored tags and
results are unchanged, only faster.
Upgrade notes
- No migration, no behavior change. Update, restart, and re-run; a sync or
reclassify that is already running keeps the old speed until you restart the
app. - Long runs and sleep: a laptop that falls asleep pauses syncing entirely
(nothing progresses while asleep). For a multi-hour first sync or
reclassify, keep the machine awake until the job finishes.