MCP servers

The job

Model Context Protocol servers give a coding agent extra abilities through a shared interface, such as searching code, recalling earlier work, fetching documents or operating another program. A developer adds one when the agent alone cannot see what it needs.

How they differ

One kind helps an agent understand a codebase. Codebase-memory-mcp, codegraph, ripwire and Trail-Code build maps of code so the agent reads fewer files, and xerj and zvec-grep search code and documents.

A second kind remembers. Agentmemory records coding-agent sessions so several agents can search one history, and gbrain keeps sourced facts that can be corrected or withdrawn. OpenLore serves shared Markdown notes over SSH and MCP. All three suit people who move between agents.

A third kind brings in outside content. Context7 fetches library documentation, while firecrawl and apify-mcp-server return web pages and scraped data through a service.

A fourth kind operates another program. Chrome-devtools-mcp drives Chrome for debugging and performance checks, agent-desktop reads desktop accessibility data, and artemis drives Android phones. Rea links an agent to reverse-engineering tools, and SkillSpector reports on skills before they are installed.

A fifth kind sits between agents, models and accounts. OmniRoute offers one endpoint in front of many model providers, headroom trims what reaches a model, and ruflo coordinates several agents. Codex-with-chatgpt and chat-on-steroids connect ChatGPT to local work, and PI-Desktop is a graphical workspace for coding agents.

Licences

On 10 November 2026, GitHub detected MIT or Apache-2.0 on most of the tools here. Firecrawl carried AGPL-3.0; PI-Desktop carried LGPL-3.0. A licence can change, and each tool's own page gives the latest reading with its date.

What to check first

Controlling a browser or desktop has its own cautions. Chrome-devtools-mcp warns that the client can inspect and change data held in the browser. Agent-desktop warns that what it captures from the screen or clipboard may be private. Ruflo says its plugin mods run with the account's own permissions and no sandbox.

In agentmemory's default shared scope, any connected agent can read what others stored. Setup often changes agent configuration. Agentmemory, artemis, codebase-memory-mcp and zvec-grep write settings for agents, and Trail-Code reads past Claude Code sessions and adds hooks. Codegraph installs through a piped shell script and has a telemetry section.

Data may leave the machine. Firecrawl calls a hosted API, context7 uses a hosted endpoint, and chrome-devtools-mcp sends usage statistics to Google by default. Codex-with-chatgpt exposes a local workspace through a Cloudflare tunnel, and headroom sends traffic to OpenAI unless changed. Chat-on-steroids warns of possible account suspension.

PI-Desktop, an early preview, has an agent that runs commands and takes marketplace plugins. Xerj's examples start its server with an insecure flag, OpenLore's quick start exposes SSH and web ports, and gbrain warns of a look-alike npm package.