Atomic Agent Files

27 PRs since v0.1.72, most of them driven by requests and reports from the community in Discord. This is the largest release the project has shipped so far, and it is mostly about model setup and control: a full live model catalog with a filterable picker, a five-level permission ladder, and automatic fallback between providers.

Highlights

Five-level permission ladder

Approvals used to be a single switch: either the agent asked about everything, or it asked about nothing. That is a hard choice to make well, so we replaced it with a ladder of five levels. L1 (paranoid) asks before every action and is the default. L2 lets the agent edit files inside the project you are working in. L3 adds your home directory and HTTP requests. L4 adds shell commands and scripts. L5 is full trust and runs everything. You can move between levels on the Privacy tab or with /privacy level, and it applies live with no restart. Right inside an approval prompt you also get two shortcuts: s allows a whole category for the rest of the session, and a allows one specific shell command shape. Those session grants live in memory only, which is safer than flipping a permanent global toggle. Hardline actions like rm -rf / are blocked at every level, and config.json and .env always ask no matter where you are on the ladder. (#97, [#100], [#101])

Automatic provider fallback

Before this release, if your active provider went down (a 429, a 5xx, or a network error), the turn failed and you had to switch providers by hand. Now you can set up an ordered chain of providers, with an optional local model at the end, and the agent fails over to the next one on the list when the current one stops responding. It sticks with whichever provider is working, and quietly retries your primary about once every five minutes so you drift back to it once it recovers. You edit the chain in the TUI, on the Fallback tab of the LLM tab. Cooldowns back off as failures repeat: 30 seconds, then 1 minute, then 5 minutes. (#102, [#90], closes [#70])

Full live model catalog with a filterable picker

The picker used to show only a slice of what your provider actually offers: about 12 of 305 models on OpenRouter, 13 of 337 on aimlapi. It now pulls the full live catalog and renders it as an inline list right in the panel, so you are not guessing at what is available. Start typing to filter the list by substring, case-insensitive, and it narrows as you go. There are also 10 built-in provider presets, so you can add a provider without hunting down its base URL. (#67, [#68], [#83], [#91], [#95], [#80])

Everything else

Model selection and providers

• Ten built-in provider presets (Nous, Groq, DeepSeek, Together AI, Fireworks AI, Cerebras, Mistral, xAI, Ollama Cloud, LM Studio); each key lives in its own env var, so a second service never overwrites the first one's key. (#80)
• One canonical /model command; /models and /local become quiet aliases. (#92)
• /model works more than once per session (picker wiring plus an editor cursor fix). (#93)
• Provider presets are sorted alphabetically in the add-provider wizard. (#94)
• Shorter provider wizard: presets skip the base-URL screen, and keyless or local presets skip the key screen. (#96)
• Brand the OpenRouter app as "Atomic Agent" and add marketplace categories. (#98)

Permissions and safety

• The five-level ladder ships as agent.approvalLevel (config v37) with a transparent migration from the old toggle. (#97)
• Review follow-ups for the ladder: injected trust paths, deduped filesystem approvals, category forwarding, honest capability reporting. (#100)
• Session-scoped grants at the approval prompt (s for a category, a for a shell command shape), held in memory only. (#101)
• Privacy tab "approve everything" toggle, applied live and persisted. (#81)

Reliability

• Type cloud provider errors (OpenAiHttpError) and give the cloud path a retry budget, so a 429 no longer reads as a tool bug and a timeout no longer reads as a cancel. (#90)
• Validate the /health body and identify OpenAI-compatible runners, which fixes a permanent freeze against runners that answer 200 on every path (KoboldCpp). Closes [#65], [#66]. (#73)
• Retry transient .env read errors on Windows (EPERM/EACCES/EBUSY/EAGAIN) and surface a persistent failure loudly, so stored keys no longer vanish silently. (#82)

Agent capabilities

• /tools lists the built-in capabilities so they are discoverable. Closes [#71]. (#74)
• os.fs.locate_project resolves a project name to a path from bounded local sources, with no disk-wide scan. Closes [#77]. (#86)
• Opt-in Telegram reports for scheduled task outcomes. Closes [#78]. (#87)
• Advertise chat scrolling in the footer (fn+up/down on macOS, pgup/pgdn elsewhere). (#72)

Analytics

• Attach token counts and estimated spend to the message_sent event. (#85)

Docs

• Refresh the README, remove an inaccurate data claim, document the default-on anonymous analytics and how to disable it, and add a way to book a call with the team. (#84)
• Document the model picker filter. (#88)

Fixes

• Stop LLM panel rows from garbling on a narrow terminal (width-aware rendering). (#99)

Install / Update

macOS / Linux:

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

Windows (PowerShell):

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

Atomic Agent updates itself in place. After an update the TUI prompts you to restart.

Full Changelog: https://github.com/AtomicBot-ai/atomic-agent/compare/v0.1.72...v0.2.0

Built by @sosidudku1 and @Ooooze.