Download Latest Version atomic-agent-win32-x64.zip (43.7 MB)
Email in envelope

Get an email when there's a new version of Atomic Agent

Home / v0.1.69
Name Modified Size InfoDownloads / Week
Parent folder
atomic-agent-win32-x64.zip.sha256 2026-08-04 93 Bytes
0
atomic-agent-linux-x64.tar.gz.sha256 2026-08-04 96 Bytes
0
atomic-agent-win32-x64.zip 2026-08-04 43.7 MB
0
atomic-agent-linux-x64.tar.gz 2026-08-04 51.0 MB
0
atomic-agent-linux-arm64.tar.gz.sha256 2026-08-04 98 Bytes
0
atomic-agent-linux-arm64.tar.gz 2026-08-04 50.8 MB
0
atomic-agent-darwin-arm64.tar.gz.sha256 2026-08-04 99 Bytes
0
atomic-agent-darwin-arm64.tar.gz 2026-08-04 49.8 MB
0
README.md 2026-08-04 3.4 kB
0
v0.1.69 source code.tar.gz 2026-08-04 5.7 MB
0
v0.1.69 source code.zip 2026-08-04 6.6 MB
0
Totals: 11 Items   207.6 MB 0

Every change in this release came through the tracker in the past week: five community PRs and one from the team, covering the broken first-run experience on the default model, an SSRF gap, and long-requested quality-of-life for local-model setups.

Fixed

  • The default model could stall forever in managed mode. The bundled chat template for Qwen 3.5 4B was written for an older revision of the model and disabled its reasoning channel, which deadlocked the tool-call grammar: the GPU stayed busy, nothing ever reached the screen, and retries hid the failure for up to 15 minutes. The daemon now uses the model's own template, stalls abort quickly with a clear error, the auto-context floor is raised so bigger models get room to answer, and the slot count is read from the server instead of guessed (#48 by @sachin-detrax, fixes [#47]).

  • os.http.request skipped the SSRF guard that os.web.fetch enforces. With the default-open allowlist, a model-controlled request could reach private and internal addresses, including cloud metadata. Both tools now share the same host checks, and every redirect hop is re-validated (#46 by @Bartok9, fixes [#45]).

  • Reasoning models over OpenAI-compatible APIs could kill a session with model returned empty content. Qwen3.8 with preserve_thinking and DeepSeek-R1 can end a turn with the entire answer in the reasoning channel and an empty content; the runtime treated that as fatal. The answer is now salvaged from the reasoning body, and vision sub-calls got a realistic token budget so reasoning models stop returning truncated image descriptions (#51 by @sosidudku1).

Added

  • An External llama.cpp pane in the LLM tab. Point the chat route at your own llama-server and switch back, with a health probe first so an unreachable URL never steals a working setup. Switching away also stops the managed daemon and frees the VRAM it was holding (#50 by @sachin-detrax, fixes [#49]).

  • Model picker for OpenAI-compatible providers. The setup wizard now pulls the model list from the server instead of asking you to type ids from memory, degrades to typed entry when the endpoint refuses, and normalizes base URLs in the provider so a pasted trailing /v1 no longer breaks requests (#41 by @sachin-detrax, fixes [#31]).

  • Live progress in Telegram. A status bubble updates in place while the agent works (step counter, current tool) and is deleted when the reply lands, so history stays clean. Edits are throttled and deduped (#28 by @SergiioB).

This release also carries the error-reporting and installer hardening tagged as v0.1.68, which was never published separately.

Updating

macOS / Linux

curl -fsSL https://atomicagent.io/install | sh

Windows (PowerShell)

irm https://atomicagent.io/install.ps1 | iex

Existing installs update in place; the TUI prompts for a restart afterwards.

New Contributors

Thanks @sachin-detrax, @Bartok9 and @SergiioB for the fixes and features, and @Ooooze for the reviews and the release work.

Full Changelog: https://github.com/AtomicBot-ai/atomic-agent/compare/v0.1.67...v0.1.69

Source: README.md, updated 2026-08-04