contextfs
ContextFS gives LLM agents a persistent, structured filesystem via MCP, enabling sandboxed file operations, workspace management, and memory/skills storage without direct machine access.
README
contextfs
Context filesystem server/client for LLM agents — MCP-compatible, multi-tenant, zero-dependency dashboard.
Author: Javier Leandro Arancibia — intrane.fr · arancibiajav@gmail.com
Architecture
graph TD
Agent[LLM Agent] -->|MCP SSE/Stdio| Server[ContextFS Server]
Server -->|WebSocket| WS1[WS Client A]
Server -->|WebSocket| WS2[WS Client B]
WS1 -->|Sandboxed| FS1[Workspace A1]
WS1 -->|Sandboxed| FS2[Workspace A2]
ContextFS follows a Hub-and-Spoke architecture:
- Server (Hub): Manages authentication, virtual client registration, and sticky scheduling.
- WS Client (Spoke): Physically executes tool calls in a sandboxed environment.
- Local Mode: For single-machine use, the server acts as its own spoke.
What is contextfs?
contextfs gives LLM agents a persistent, structured filesystem they can read from, write to, and search — without giving them direct access to your machine.
It exposes a standard MCP (Model Context Protocol) surface over SSE or stdio, so any MCP-compatible LLM client (Claude Desktop, Cursor, custom agents) can use it out of the box.
Key concepts
| Concept | Description |
|---|---|
| WS Client | A machine/process that physically executes tool calls. Connects to the server via WebSocket. |
| Virtual Client | A logical tenant (agent, user, project). Has its own API key, workspace, skills, and memory. Multiple VCs can be served by the same WS client. |
| Workspace | A directory scoped to a virtual client. All file operations are sandboxed within it. |
| Local mode | Server executes tools in-process — no WS client needed. Ideal for single-machine use. |
Quick start
Option A — Local mode (simplest, single machine)
npx contextfs server --local --mcp sse
# Create a virtual client
curl -s -X POST http://localhost:3010/api/virtual-clients \
-H 'Content-Type: application/json' \
-d '{"name":"my-agent"}' | jq .
# → copy "id" and "apiKey"
# Open dashboard
open http://localhost:3010
# Chat
CONTEXTFS_VC_ID=<id> CONTEXTFS_VC_KEY=<key> OPENROUTER_API_KEY=sk-or-... \
npx contextfs chat --mcp-server http://localhost:3010
Option B — Remote WS client mode
# Terminal 1 — server
npx contextfs server --mcp sse
# Terminal 2 — create entities
curl -s -X POST http://localhost:3010/api/ws-clients \
-H 'Content-Type: application/json' -d '{"name":"worker-1"}' | jq .
# → copy wsc id + apiKey
curl -s -X POST http://localhost:3010/api/virtual-clients \
-H 'Content-Type: application/json' -d '{"name":"agent-1"}' | jq .
# → copy vc id + apiKey
# Terminal 3 — WS client (can be on a different machine)
CONTEXTFS_API_KEY=<wsc-key> \
npx contextfs client \
--url ws://localhost:3010 \
--ws-client-id <wsc-id> \
--api-key <wsc-key>
# Terminal 4 — chat
CONTEXTFS_VC_ID=<vc-id> CONTEXTFS_VC_KEY=<vc-key> OPENROUTER_API_KEY=sk-or-... \
npx contextfs chat --mcp-server http://localhost:3010
CLI reference
contextfs server
contextfs server [options]
Options:
--port <port> HTTP + WS port (default: 3010, env: PORT)
--local Local mode: tools run in-process, no WS clients (env: CONTEXTFS_LOCAL=1)
--mcp [sse] Enable MCP server
(default: stdio — requires --vc-id + --vc-key)
(sse — VC creds per-connection: ?vcId=&vcKey= or headers)
--vc-id <id> Virtual client ID for stdio MCP (env: CONTEXTFS_VC_ID)
--vc-key <key> Virtual client API key for stdio MCP (env: CONTEXTFS_VC_KEY)
--insecure Enable contextfs.bash_script_once tool
--verbose Verbose logging
contextfs client
contextfs client [options]
Options:
--url <wsUrl> Server WebSocket URL (required, env: CONTEXTFS_SERVER_URL)
--ws-client-id <id> WS client ID (required, env: CONTEXTFS_WS_CLIENT_ID)
--api-key <key> WS client API key (required, env: CONTEXTFS_API_KEY)
--cwd <path> Workspace root (default: ~/.contextfs/workspaces/<wsc-id>)
--insecure Enable contextfs.bash_script_once
--verbose Verbose logging
contextfs chat
contextfs chat [options]
Options:
--mcp-server <url> MCP server base URL (default: http://localhost:3010)
--vc-id <id> Virtual client ID (env: CONTEXTFS_VC_ID)
--vc-key <key> Virtual client API key (env: CONTEXTFS_VC_KEY)
--model <model> LLM model via OpenRouter (env: CONTEXTFS_MODEL)
--message <text> Non-interactive: send single message, exit 0 (alias: -m)
--stdin Read message from stdin
--output json Output { message, toolCalls, durationMs } JSON
--no-tools Disable tool calls (pure LLM)
--verbose Verbose logging
Interactive commands: /exit, /clear, /tools, /history
MCP integration
SSE (recommended for web/agent use)
GET http://localhost:3010/mcp/sse?vcId=<id>&vcKey=<key>
POST http://localhost:3010/mcp/message?sessionId=<sessionId>
Or via headers:
GET http://localhost:3010/mcp/sse
X-VC-ID: <id>
Authorization: Bearer <key>
stdio (for Claude Desktop, Cursor, etc.)
{
"mcpServers": {
"contextfs": {
"command": "npx",
"args": ["contextfs", "server", "--local", "--mcp", "--vc-id", "<id>", "--vc-key", "<key>"]
}
}
}
Tools reference
All tools are sandboxed within the virtual client's active workspace root.
| Tool | Description |
|---|---|
contextfs.list |
List files/directories. Supports recursive, depth, glob filter. (RTK Optimized) |
contextfs.read |
Read file content. Supports line ranges, byte limits, and automatic filtering for large files (>500 lines). |
contextfs.summarize |
(New) Get an intelligent summary of a code file (signatures, docstrings, complexity) at 90% lower token cost. |
contextfs.write |
Write or append to a file. |
contextfs.list_workspaces |
List available workspaces for the current virtual client. |
contextfs.use_workspace |
Switch the active workspace for the current session. |
contextfs.save_skill |
Save a reusable skill as Markdown under /skills/. |
contextfs.list_skills |
List skills, optionally filtered by tag. |
contextfs.save_memory |
Persist a memory entry under /memory/YYYY/MM/. |
contextfs.search_memory |
Full-text keyword search across all memory files. |
contextfs.memory_summary |
Metadata summary of all memory entries. |
contextfs.memory_by_date |
Retrieve memories from a specific year/month. |
contextfs.memory_by_tag |
Find all memories with a specific tag. |
contextfs.bash_script_once |
Execute a one-shot bash script (requires --insecure). (RTK Optimized for tests) |
RTK Integration (Token Optimization)
ContextFS integrates RTK (Rust Token Killer) to significantly reduce token consumption (60-90%) when agents interact with the filesystem.
How it works
When running in a Docker container or where the rtk binary is available:
- Core Commands:
ls,grep,git, anddockerare automatically proxied through RTK to strip redundant metadata and formatting. - Test Optimization:
npm test,cargo test, andpytestoutputs are filtered to show only the first 5 failures and a summary, preventing token blowup on large suites. - Intelligent Summarization: The
contextfs.summarizetool leverages RTK's structural analysis to provide code overviews without reading full file content. - Ultra-Compact Mode: Force maximum compression by setting
CONTEXTFS_RTK_ULTRA_COMPACT=trueor passing the-uflag to supported commands.
Native Fallback
Integration is non-intrusive. If RTK is unavailable, fails, or a command is not supported, ContextFS automatically falls back to native execution to ensure reliability.
REST API reference
Base URL: http://localhost:3010/api
WS Clients
GET /api/ws-clients List all WS clients
POST /api/ws-clients Create WS client → returns apiKey (once)
DELETE /api/ws-clients/:id Delete WS client
POST /api/ws-clients/:id/regen-key Regenerate API key → returns new apiKey
Virtual Clients
GET /api/virtual-clients List all virtual clients
POST /api/virtual-clients Create virtual client → returns apiKey (once)
DELETE /api/virtual-clients/:id Delete virtual client + owned workspaces
POST /api/virtual-clients/:id/regen-key Regenerate API key
Workspaces
GET /api/virtual-clients/:vcId/workspaces List workspaces
POST /api/virtual-clients/:vcId/workspaces Create workspace
DELETE /api/virtual-clients/:vcId/workspaces/:wsId Delete workspace
Dispatch
POST /api/dispatch
Body: { virtualClientId, virtualClientApiKey, tool, params, timeoutMs? }
Dispatches a tool call to the assigned WS client and waits for the response.
Status + MCP
GET /api/status Summary counts
GET /mcp/sessions Active MCP SSE sessions
Security model
- WS client API keys — authenticate each WS connection. Validated on every WebSocket message.
- Virtual client API keys — authenticate MCP sessions (
?vcKey=) and REST dispatch calls. Each virtual client is fully isolated from others. - Path sandboxing — all file operations are resolved within the workspace root. Any path traversal attempt (
../) returns an error. bash_script_once— disabled by default. Requires explicit--insecureflag on both server and client.- API keys are shown only once — on creation and on regen. Store them immediately.
Documentation
- MIGRATION.md — Upgrading from prototypes to v1.
- RUNBOOK.md — Operational guide, scaling, and maintenance.
Dashboard
Open http://localhost:3010 after starting the server. The dashboard shows:
- WS clients with live status, CPU load, RAM usage, heartbeat time
- Virtual clients with assignment status
- Workspaces per virtual client
- Active MCP sessions
No login required — serve behind a reverse proxy with authentication for production use.
Data directory
All state is persisted to ~/.contextfs/:
~/.contextfs/
├── ws-clients.json WS client registry
├── virtual-clients.json Virtual client registry
├── workspaces.json Workspace registry
├── chat-config.json Chat TUI config (API key, model)
├── .machine-id Persistent client identity
└── workspaces/
└── <vcId>/
└── <wsId>/
├── skills/ Saved skills (.md files)
├── memory/ Memory entries (YYYY/MM/*.md)
└── ... Your files
Environment variables
| Variable | Description | Default |
|---|---|---|
PORT |
Server HTTP port | 3010 |
CONTEXTFS_LOCAL |
Enable local mode (1) |
— |
CONTEXTFS_INSECURE |
Enable bash_script_once (1) |
— |
CONTEXTFS_SERVER_URL |
WS server URL for client | — |
CONTEXTFS_WS_CLIENT_ID |
WS client ID | — |
CONTEXTFS_API_KEY |
WS client API key | — |
CONTEXTFS_VC_ID |
Virtual client ID for chat/MCP | — |
CONTEXTFS_VC_KEY |
Virtual client API key for chat/MCP | — |
CONTEXTFS_MCP_SERVER |
MCP server base URL for chat | http://localhost:3010 |
CONTEXTFS_MODEL |
LLM model for chat | google/gemini-2.5-flash-preview |
OPENROUTER_API_KEY |
OpenRouter API key for chat | — |
CONTEXTFS_RTK_ENABLED |
Enable RTK optimization (true/false) |
true (auto-detect) |
CONTEXTFS_RTK_ULTRA_COMPACT |
Enable ultra-compact mode (true) |
false |
VERBOSE |
Enable verbose logging (1) |
— |
Requirements
- Node.js ≥ 18
- No build step required
- No native modules
License
MIT © Javier Leandro Arancibia
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.