M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
A

Atomic Agent

Community
Not a standard server
A local-first AI agent that runs on your machine.

This is not a typical connect-and-use MCP server; read the source repository before using it.

How to install ↓ Compare GitHub source repository ↗
Category
Other #187 of 230
Stars
★ 3.4k Very popular
Runtime
Node.js 25.7+ · Prebuilt binary
Credentials
No credential needed
License
MIT
Last commit
44FMRS · D

The supplied material describes Atomic Agent as a local-first AI agent and MCP client, not as an installable MCP server implementation. It offers broad automation, local-model support, approvals, and observability, but requires substantial system access and remains a developer preview.

Strongest · Documentation 13/20 Weakest · Reliability 3/20

Reliability
3/20
Security and permissions
10/20
Maintenance
10/20
Documentation
13/20
Setup experience
8/20
Why each score
Reliability 3/20
The supplied material contains no cached server manifest, MCP initialization flow, tool list, or committed test evidence. The README primarily describes a local agent that connects to external MCP servers, not a declared MCP server itself. Handshake behavior, tool contracts, error handling, and key execution paths therefore cannot be verified, so the static calibration requires a substantial deduction.
Security and permissions 10/20
The README states that shell, filesystem writes, patches, HTTP requests, process termination, and untrusted MCP tools are approval-gated, and it discloses local state and network-egress categories. Points are deducted because implementation evidence and precise least-privilege boundaries are absent; skills and shell inherit the agent environment including .env secrets, and the README explicitly says redaction is not complete isolation. Browser, arbitrary HTTP, and external MCP data flows remain configuration- and approval-dependent.
Maintenance 10/20
The supplied facts indicate that the project is not archived, has an MIT license, GitHub issue and release workflow references, and a version entry, while identifying the software as a developer preview. Points are deducted because there is no commit history, maintainer-response evidence, dependency-update policy, security-response channel, or stability record; the README says APIs, commands, and configuration are still changing.
Documentation 13/20
The README covers installation, execution, platform requirements, configuration, secret storage, capabilities, privacy and egress, limitations, troubleshooting, and several usage modes, including an MCP client configuration example. Points are deducted because there is no server manifest, concrete tool parameter/return contract, mainstream MCP-client connection example, or verifiable test evidence; it also documents developer-preview status and platform constraints.
Setup experience 8/20
The material provides installers for macOS/Linux and Windows, a run command, state-directory configuration, and an MCP client configuration snippet. Points are deducted because operation depends on Node, llama-server, Chromium, git, and additional model/browser setup, and the material does not demonstrate that this repository can run as an MCP server for mainstream clients. No manifest or standard MCP server launch configuration is supplied.

Static review · not runListed 2026-08-14

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or codeUses the networkControls a browser

Best for

  • Users who want the control loop and state kept on their device
  • Users with llama.cpp, Ollama, LM Studio, or another OpenAI-compatible model backend
  • Developers needing browser, filesystem, document, and task automation
  • Technical users willing to review approvals, traces, and network egress

Not for

  • Users looking for a standalone MCP server that can be installed directly into an MCP client
  • Production integrations requiring stable APIs, because the project is still a developer preview
  • Environments that cannot allow agent access to browsers, files, shells, or networks
  • Systems without Node.js, Chromium, git, or an accessible model backend

Required permissions

  • Read and write access to the local state directory and SQLite databases
  • Permission to run approved shell commands and perform some process operations
  • Access to the configured Chromium browser, browser profile, and websites
  • Network access to configured llama-server, cloud models, search providers, or HTTP endpoints
  • Permissions required by external MCP servers whose tools are invoked
  • Optional macOS Accessibility, Screen Recording, Automation, or Reminders permissions

Risks and side effects

  • Browser navigation, HTTP tools, cloud models, search providers, MCP, Telegram, update checks, and analytics can create network egress
  • Shell commands and skills inherit the agent process environment and may access `.env` secrets
  • Traces and `.env` under the state directory should be treated as sensitive; the project says redaction and environment filtering are not complete isolation
  • Filesystem writes, patches, archive extraction, process termination, HTTP requests, and untrusted MCP tools are policy-gated, but approval errors can still cause harm
  • The project is a developer preview, so commands, configuration, and behavior may change

Setup

Before you start

Runtime：Node.js 25.7+ · Prebuilt binary

ATOMIC_AGENT_LLAMA_API_KEY optionalsecret Optional bearer token for the external llama-server; supply your own.
TELEGRAM_BOT_TOKEN optionalsecret Bot token for the Telegram remote-control channel; create a bot via @BotFather and put it in <stateDir>/.env.
Other optional settings (7)
ATOMIC_AGENT_STATE_DIR optional Sets the state directory (config, memory, tasks, traces, etc.); defaults to ~/.atomic-agent.
ATOMIC_AGENT_LLAMA_URL optional URL of an external llama-server, e.g. http://127.0.0.1:8080.
ATOMIC_AGENT_LLAMA_MAX_TOKENS optional Caps the number of completion tokens per call.
ATOMIC_AGENT_BROWSER_CHANNEL optional Selects the browser channel: chrome, msedge, or chromium.
ATOMIC_AGENT_BROWSER_EXECUTABLE_PATH optional Explicit path to a Chromium-family browser executable.
ATOMIC_AGENT_BROWSER_CDP_URL optional Attach to an already-running browser via CDP.
ATOMIC_AGENT_UPDATE_CHECK_ON_STARTUP optional Set to false to skip the startup GitHub Releases update check.

On macOS or Linux, run curl -fsSL https://atomicagent.io/install | sh; on Windows PowerShell, run irm https://atomicagent.io/install.ps1 | iex. Then run atomic-agent. The project is a developer preview, so integrations should pin a release.

Check that it works

After installing, run `atomic-agent`; if the TUI starts and shows the interactive console (try /help to list commands), the install worked. You can also run `atomic-agent models list` to confirm managed models are visible.

Troubleshooting

  1. Verify Node.js, git, a Chromium-family browser, and an accessible llama-server or other model backend
  2. Check the browser executable, browser channel, and CDP settings
  3. On Linux, verify dependencies for ripgrep, clipboard, notifications, and window control
  4. If Chrome fails in a Linux environment, check sandbox configuration and the documented `--no-sandbox` option
  5. Check the model service health, model name, port, and OpenAI-compatible path
  6. Check MCP command, arguments, enabled state, and trust level; inspect the surfaced stderr tail when a server fails to connect
  7. Review approval policies, repeated-call protection, logs, and local NDJSON traces
  8. For problems, collect error logs and system specifications, then use the GitHub issue tracker or Discord

Things to try

Once connected, you can ask your AI assistant things like:

  • Open the browser, search for today's tech news, and summarize the key points
  • Extract the text from the PDF on my desktop
  • Fix the bug in this source file, but ask me for approval before making changes
  • Create a cron job that sends me a morning digest via Telegram at 9am

Use cases

Running long-context, multi-step AI tasks locally
Automating browser actions and web access
Reading, editing, and searching local files
Running approval-gated shell commands
Connecting external tools, resources, and prompts through MCP
Maintaining local memory and scheduled workflows

Overview

Atomic Agent is a local-first AI agent that runs on the user's device with local or cloud models. It can drive browsers, read and edit files, run approved shell commands, extract document text, retain memory across sessions, schedule tasks, and connect to external MCP servers. The project is described as an MCP client, not as an MCP server with a declared transport.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 659 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision 7899d173e0a0 Data synced 2026-10-19 Read the FMRS scoring method