Atomic Agent is a local-first AI agent for the desktop: the control loop and all state (sessions, memory, tasks, traces, skills, browser profile, and config) run on your machine as plain files and SQLite databases. It browses the web, reads and edits files, runs approved shell commands, extracts text from documents, remembers context across sessions, schedules follow-ups, and calls external tools over MCP. It runs on a local model, a cloud model, or both at once in Fusion mode, where one model plans and a pool of workers executes.
The agent loop is built to keep small quantized models useful over long, multi-step work. One inference returns one JSON array of tool calls, grammar-constrained (GBNF) on a local llama-server and native tool calling on cloud providers. The core runs them, with independent reads in parallel and risky actions behind approval, then summarizes results instead of pasting them back in full. The prompt keeps a byte-stable prefix so the server reuses its KV-cache, clips conversation and state into a fixed budget, and keeps sessions, memory, and traces outside the prompt. Long jobs continue past 25-step checkpoints while they make progress, bounded by a per-task ceiling of 1000 steps or 2 hours by default.
Local inference runs on the project's own TurboQuant build of llama.cpp: low-bit KV-cache quantization compresses the cache up to about 6.4x versus F16, Lloyd-Max weight quantization fits small models on consumer hardware, and speculative decoding heads for Gemma 4 and Qwen 3.6 add 30-50% throughput. A managed mode downloads and updates the backend and curated GGUF models. Cloud providers include OpenAI-compatible endpoints, OpenRouter, and Gemini, with presets for Anthropic, Groq, DeepSeek, Mistral, xAI, Together, Ollama, and LM Studio, and existing Claude Code and OpenAI Codex subscriptions can be driven through their signed-in CLIs.
On the GAIA validation Level 1 split (53 tasks), Atomic Agent and Hermes drove the same local qwen-3.6-35b-a3b model under the same step budget and timeout:
| Metric | Atomic Agent | Hermes |
|---|---|---|
| Accuracy | 37/53 = 69.8% | 31/53 = 58.5% |
| Avg wall time per task | ~217 s | ~351 s |
| Head-to-head wins | 15 Atomic Agent only | 9 Hermes only |
The same loop on the same split with smaller local models, Atomic Agent alone (different versions per row):
| Chat model | Accuracy | Avg wall time per task |
|---|---|---|
| qwen-3.6-35b-a3b (UD-Q4_K_XL) | 37/53 = 69.8% | ~217 s |
| qwen-3.5-9b (Q4_K_M) | 28/53 = 52.8% | ~152 s |
| gemma-4-12b (it-qat UD-Q4_K_XL) | 24/53 = 45.3% | ~423 s |
It ships as a TUI and CLI, a Tauri desktop sidecar, an HTTP server for embedding in other apps, and Telegram and Discord channels with approval buttons. First-run import brings skills, sessions, memory, MCP servers, and cron jobs over from Claude Code, Codex, Pi, Oh My Pi, Hermes, and OpenClaw, depending on the source. Every action lands in append-only NDJSON traces, dangerous actions pass approval gates, a no-progress guard vetoes repeated identical tool calls, and files a write replaces are kept for restore.
Features
- Browser: navigates, clicks, types, manages tabs, and reads compact ARIA snapshots through playwright-core in Chrome, Edge, Brave, or Chromium
- Web and HTTP: web search through Exa, DuckDuckGo, Brave, or SearXNG, plus SSRF-guarded page fetch and HTTP requests separate from the browser
- Filesystem and shell: read, write, edit, patch, glob, grep, diff, archive extraction, approved shell commands, and process inspection
- Documents: local text extraction from PDF, DOC, DOCX, XLSX, PPTX, ODT, RTF, and plain text
- Git and GitHub: read-only repository inspection, approval-gated local writes, optional remote sync, and pull request and issue actions once a token is saved
- Memory: profile facts, SQLite and FTS5 notes with optional embeddings for hybrid recall, links, lessons, procedures, voting, dedup and eviction, and Obsidian export
- Tasks: durable deferred turns, cron schedules, intervals, webhooks, and agent-created reminders
- Skills and MCP: Markdown skill playbooks installable from ClawHub or GitHub, plus external MCP servers whose tools join the same registry
- Fusion: one model plans and delegates wide reads, drafts, and tests to a pool of workers, then checks and merges their results
- Providers: local llama-server by default, OpenAI-compatible, OpenRouter, and Gemini providers, presets for Anthropic, Groq, DeepSeek, Mistral, xAI, Together, Ollama, and LM Studio, and Claude Code and Codex subscriptions through their CLIs
- Channels and integrations: Telegram and Discord remote control with approval buttons, Composio toolkits over MCP, and an agent e-mail inbox
- Safety and observability: approval gates, append-only NDJSON traces, prompt drift replay, a failure taxonomy, a no-progress guard, restore copies of replaced files, and Trash instead of delete
