Skip to content

Atlas v1.0.6

Choose a tag to compare

View all tags
@github-actions github-actions released this 30 Sep 20:11
· 5 commits to main since this release
v1.0.6
f1a5d20
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Atlas 1.0.6

Web search without an API key, in one click. Profile → Advanced → Web search → Set up free search. Atlas installs free-search-mcp in the background (and uv if needed), runs a test search, and switches over only once that works; it takes about 10 seconds. Queries still go to public search engines, with SafeSearch set to strict. You can switch back to a Tavily key at any time.

See what the model is thinking. Replies from thinking models show a Thinking panel: open while the model thinks, then collapsed to "Thought for Ns". Click to expand. This works with LM Studio, Ollama, OpenRouter, Claude and Gemini. Chat-template tokens that some local models leaked into answers (such as <|begin_of_box|>) are removed.

"Continue" continues. After a failed reply, "continue", "try again" or "go on" re-runs that request instead of starting over. In Web search, a bare "continue" repeats your previous search question and never searches the word itself.

Much better research and search with local models:

  • Research builds a real multi-query plan with local thinking models; before, it silently fell back to searching your raw question.
  • Questions that name someone and then say "his" or "their" are no longer refused as "needs a subject".
  • Citations written as (S1), 【S1】, [Source 1] or a bare S1 are recognized, and Web search gets one repair attempt, so fewer answers end up as "Partial: missing citations".
  • The model no longer calls your sources "untrusted evidence".

Also: npm run build:server uses the project's server/.venv automatically, and the install steps are updated for macOS 15 and later.

Atlas 1.0.5

  • Research keeps the conversation. Research answers now see the chat's earlier turns, so follow-ups like "and what about the second one?" stay on topic. Search queries still use at most the last few turns, and a bare "do" asks what to search for instead of searching the word.
  • Optional keyless web search. Search and Research can use an installed free-search-mcp instead of Tavily (ATLAS_WEB_SEARCH_PROVIDER=free-search-mcp). Tavily stays the default. Keyless search still sends queries to external search engines, and it is not yet verified inside the packaged app.
  • New safe tools: a calculator and a side-by-side source comparison for tool-capable models.
  • Research digs deeper when evidence is thin, with one extra bounded search round.
  • Chat polish: a new greeting on the empty chat and a styled copy-code button.
  • Memory works in the downloaded app. 1.0.4 is the first build made with a Python that can load the memory store's database extension. Earlier downloads (1.0.1) were built without it.

Atlas 1.0.4

Models and routing

  • The model picker has Auto (default). It uses local models first and only uses your connected cloud providers after you turn on Allow cloud models in Auto. Picking a specific model always uses exactly that model.
  • A reasoning slider (Off, Low, Med, High, Max) in the model picker controls how much the model thinks. It applies to OpenAI reasoning models, Claude (extended thinking), Gemini 2.5/3, OpenRouter, and local models that support it. In Auto, the slider is a maximum and Auto picks less for simple messages.
  • Auto now searches the web by itself when a message needs current information. It uses tool calls when the model supports them and the search pipeline otherwise; without a search key it answers without searching and says why.
  • Fixed new chats failing with "No permitted model" when a cloud model was shown as selected.

Conversation memory

  • Every message now carries the chat's own recent history, so context survives app and server restarts, model switches, Auto routing changes, and web searches or plan runs in between. Recreated sessions also keep the chat's memory summary.

Chat

  • Replies show the steps of a run (planning, searches, tools) as chips, which are saved with the chat.
  • Retry regenerates a reply in place and brings the original back if the retry fails. Cited sources appear as chips.
  • Select text in a reply to Explain, Shorten, or Ask about it.
  • Plan mode: tick the steps to run, then Approve & run, Revise, or Dismiss.
  • "Memory used" details are shown as cards; clicking a past conversation opens it.
  • Refreshed buttons and layout in Profile and Library.

Also includes 1.0.3 (not published separately):

  • Auto text chat can use a configured cloud model when you explicitly select it; otherwise Auto stays local-only.
  • One intent selector replaces the separate mode and tool controls in the composer.
  • Agent runs that fail now say why (unsupported model, missing search key, timeout, or tool limit) instead of a generic error.
  • Grouped citations such as [S1, S3] are validated against the sources actually issued and link correctly in chat and export; exports render all heading levels. Search and Research ask who "his/her/their" refers to instead of searching a bare pronoun.
  • Refreshed onboarding: the "Atlas" wordmark is no longer cropped, and account setup uses a WebGL flowing-gradient background (a still frame when reduced motion is on, a static gradient if WebGL is unavailable).

Cloud billing and Atlas-managed inference are not available. Reasoning levels were checked against provider documentation and automated tests, not against every model.

Release pipeline

A version-matching v* tag triggers offline backend tests, desktop typechecking
and tests, an Apple Silicon bundle build, embedded-server smoke test, code-signature
verification, and DMG integrity/SHA-256 checks. The checksum detects accidental changes but is not a cryptographic signature of the publisher. GitHub publishes the release only
after those gates pass.

Install on Apple Silicon

This free release is ad-hoc signed and is not notarized, so macOS will show a
security warning.

  1. Download the .dmg and matching .sha256 file from Assets below.
  2. Verify the download with shasum -a 256 -c <downloaded-dmg>.sha256.
  3. Open the DMG and drag Atlas to Applications.
  4. Open Atlas from Applications. When macOS says it "could not verify" Atlas, click Done (not Move to Bin).
  5. Open System Settings → Privacy & Security, scroll to Security, click Open Anyway next to the Atlas message, authenticate, then click Open. This is needed once per install. (On macOS 14 and earlier, Control-click Atlas and choose Open instead.)

Only bypass Gatekeeper for a build downloaded from the official Atlas repository. This release is for Apple Silicon Macs; it does not include automatic updates.

Assets 4
Loading