PreviewLocal-first AI coding and operator agentSource check · September 13, 2026

Atomic Agent

An open-source, local-first agent runtime for coding and computer tasks, built around llama.cpp, a terminal UI and approval-gated tools.

TerminalCode EditingRepo UnderstandingGitBrowserMcpLocal ModelsSelf Hosting

Product briefing

What it is and how it fits a coding workflow

Atomic Agent is a local-first operator runtime that can handle repository work through filesystem, shell and Git tools; it is not an IDE autocomplete extension. Its CLI and terminal UI run a local model through llama.cpp by default, while configured OpenAI-compatible, OpenRouter and Gemini providers—and signed-in Claude Code or Codex CLIs—offer cloud alternatives. The runtime uses grammar-constrained tool calls, compacted state, local memory and inspectable traces to make tool-heavy sessions more manageable. Browser automation, MCP servers, skills, scheduled tasks and optional single-user Telegram control extend it beyond coding into desktop and operations workflows. Risky actions such as shell commands, file writes and untrusted MCP calls are approval-gated, but users still need to review permissions and integrations. “Local-first” does not mean every action stays offline: web searches, browser navigation and any configured hosted model send requests outside the machine. The project identifies itself as a developer preview, so check the current release and platform support before relying on it for a production team. Its published benchmark results are project-reported; compare them with your own tasks and hardware rather than treating them as an independent ranking.

Best fit

  • Trying local GGUF models as an agent that can act on a code repository
  • Developers who want inspectable memory, traces and approval gates
  • Combining coding, browser research, MCP tools and personal automations in one local runtime

Probably not your fit if…

A mature IDE with inline completions, guaranteed offline behavior without configuration, or a production-critical workflow that requires a stable non-preview release.

What stands out

  • ↗ Local llama.cpp path with optional hosted and subscription-CLI providers
  • ↗ Filesystem, shell, browser, MCP, skills and scheduled-task tools
  • ↗ Local state and traces with approval gates around risky actions

Trade-offs to check

  • ! Developer-preview status means interfaces and model support can change
  • ! Cloud providers, browser activity and search introduce network egress; inspect analytics and privacy settings
  • ! A local agent with shell and filesystem access still needs least-privilege permissions and human review

Primary references

Verify features at the source

This profile is an editorial comparison, not a performance benchmark. Product features, plans and availability can change; verify your required workflow before adopting it.

GitHub source metadata

AtomicBot-ai/atomic-agent

Local-first AI agent runtime using llama.cpp, local models and approval-gated tools.

Primary language
TypeScript
GitHub license metadata
MIT
ai-agentbrowser-automationclillama-cpplocal-firstlocal-llmmcptuitypescript

Related field notes

Further reading

Explore related tools

Other profiles for similar workflows

Search all coding agents →