20 Best Code Execution MCP Servers, Compared (November 2026)
The short answer
Most developers should install LLM Sandbox: it is published on PyPI as llm-sandbox with 106,859 PyPI downloads in a typical week, has 1,143 stars, its last commit was 10 days ago and it had 86 commits in the last 12 weeks. It exposes 3 tools with descriptions graded B and runs code across container backends and several languages. If you already work in JupyterLab and need an agent to edit and execute notebook cells in place, Jupyter MCP Server is the better fit, with 18 tools graded A and a commit 1 day ago. If large outputs are filling your agent's context, Context Mode is the one to use instead.
Whichever you choose, give it the narrowest access that still works (a read-only credential, a replica, a scratch account), and widen it only once you have watched what your agent actually asks for.
Glama operates the MCP registry these numbers are measured from, and sells MCP hosting and a gateway. No position on this page is paid for. How the registry is built.
Quick picks
- 1LLM Sandbox : Running LLM-generated Python, JavaScript, Java, C++, Go, or R code: execute_code isolates it in sandbox backends and captures visualizations automatically.
- 2Jupyter MCP Server : Existing JupyterLab users who need an agent to edit and execute notebook cells in place: it exposes cell-level edit and execution tools.
- 3Context Mode : Running code over large files and command output without flooding the agent's context: it sandboxes execution and returns only what the script prints.
- 4sandbox-mcp : Testing installers or untrusted Windows files in a disposable Windows Sandbox: it exposes one-shot install-test and detonation tools plus UI Automation and screenshot evidence.
- 5Cloudflare MCP Server : Managing Cloudflare services without the full OpenAPI spec in context: it documents code mode tools for searching the spec and executing API calls.
Which one, for your situation
| Your situation | What to use |
|---|---|
| You need Python, JavaScript, Java, C++, Go and R execution | LLM Sandbox: execute_code isolates code in sandbox backends and its llm-sandbox package had 106,859 PyPI downloads in a typical week. |
| You already use JupyterLab and want in-place notebook cell edits | Jupyter MCP Server: it exposes 18 tools graded A for cell-level notebook editing and execution, with a commit 1 day ago. |
| Large tool outputs are flooding your agent's context | Context Mode: it returns only what the script prints, has 11 tools graded A, and had 2,459 commits in the last 12 weeks. |
| You need to test an untrusted Windows installer in a disposable VM | sandbox-mcp: it exposes 27 tools graded A for install-test and detonation runs inside disposable Windows Sandbox machines, with a commit 0 days ago. |
| You want a persistent Linux container that survives between agent runs | Containarium-Cloud: it provisions SSH-reachable unprivileged LXC containers with an MCP server inside for shell commands, and had 1,320 commits in the last 12 weeks. |
| You need to drive a desktop GUI with mouse and keyboard | EdgeBox: it adds a VNC Ubuntu desktop with mouse, keyboard and screenshot control, though its last commit was 207 days ago and it had 0 commits in the last 12 weeks. |
Top MCP servers for Code Execution
| Profile | |||||||
|---|---|---|---|---|---|---|---|
| 1 | Community favourite | 1,143 | +12 | 106.9K | 10 days ago | 87.5 | |
| 2 | Community favourite | 1,325 | +20 | – | yesterday | 80.7 | |
| 3 | Community favourite | 34,083 | +2,967 | 26.1K | today | 77.3 | |
| 4 | Community favourite | 426 | – | – | today | 67.3 | |
| 5 | Community favourite | 1,092 | +137 | 7 | 3 days ago | 65.0 | |
| 6 | Community favourite | 318 | +11 | – | today | 57.6 | |
| 7 | Abandoned but popular | 344 | +5 | – | 345 days ago | 57.5 | |
| 8 | Steady | 234 | +6 | 0 | today | 56.1 | |
| 9 | Community favourite | 9,060 | +387 | – | 3 days ago | 55.9 | |
| 10 | Community favourite | 512 | +19 | 91 | 37 days ago | 55.8 | |
| 11 | Community favourite | 1,368 | +15 | – | 14 days ago | 49.6 | |
| 12 | Emerging | 60 | +11 | – | 2 days ago | 48.2 | |
| 13 | Emerging | 44 | +44 | – | 4 days ago | 46.1 | |
| 14 | Emerging | 53 | +33 | 81 | 2 days ago | 45.7 | |
| 15 | Community favourite | 66 | 0 | 623 | 2 days ago | 43.0 | |
| 16 | Steady | 33 | +1 | – | 34 days ago | 41.4 | |
| 17 | Steady | 104 | 0 | 2K | 15 days ago | 41.0 | |
| 18 | Abandoned but popular | 212 | 0 | – | 207 days ago | 40.9 | |
| 19 | Abandoned but popular | 453 | 0 | – | 296 days ago | 40.8 | |
| 20 | Steady | 3 | 0 | 38 | 12 days ago | 40.3 |
The ranking, with the evidence
- 1vndee/llm-sandbox ↗
LLM Sandbox
Community favouriteBest for: Running LLM-generated Python, JavaScript, Java, C++, Go, or R code: execute_code isolates it in sandbox backends and captures visualizations automatically.
It exposes three tools, execute_code, get_supported_languages, and get_language_details, for running LLM-generated code in an isolated sandbox across container and local Linux and Windows backends and multiple programming languages. Before selecting it, plan backend setup: the Python package installs with pip, Docker support needs the docker extra, and Bubblewrap or NsJail need system executables.
Score 87.5: show every number behind it
- Adoption100 / 100 · weight 40%
- GitHub stars76
- PyPI downloads100
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum54 / 100 · weight 14%
- Stars gained, relative to size40
- Stars gained, absolute45
- PyPI download trend89
- Tool quality63 / 100 · weight 13%
- Tool description quality55
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 87.5
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 87.5
- 2datalayer/jupyter-mcp-server ↗
Jupyter MCP Server
Community favouriteBest for: Existing JupyterLab users who need an agent to edit and execute notebook cells in place: it exposes cell-level edit and execution tools.
Jupyter MCP Server exposes 18 tools for connecting to a Jupyter server, activating notebooks, editing and executing cells, and running code directly in a kernel. Sandbox execution requires pinning code-sandboxes to a version that matches jupyter-mcp-server.
Score 80.7: show every number behind it
- Adoption78 / 100 · weight 40%
- GitHub stars78
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance100 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade100
- Commit cadence100
- Momentum50 / 100 · weight 14%
- Stars gained, relative to size48
- Stars gained, absolute53
- npm download trendnot measuredno download history for the selected registry
- Tool quality73 / 100 · weight 13%
- Tool description quality65
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 80.7
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 80.7
- 3mksglu/context-mode ↗
Context Mode
Community favouriteBest for: Running code over large files and command output without flooding the agent's context: it sandboxes execution and returns only what the script prints.
Context Mode exposes 11 tools that run code in a sandboxed subprocess in JavaScript, shell, TypeScript, Python and Perl, index file and URL content into a local SQLite knowledge base using BM25 over FTS5, and search that base with a multi-strategy ranking pipeline. Before installing, note that the plugin runs locally under a source-available ELv2 license, while its Insight dashboard opens a hosted analytics layer and Gateway is a separate hosted service with its own account and subscription.
Score 77.3: show every number behind it
- Adoption100 / 100 · weight 40%
- GitHub stars100
- npm downloads94
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance100 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade100
- Commit cadence100
- Momentum100 / 100 · weight 14%
- Stars gained, relative to size100
- Stars gained, absolute100
- npm download trend100
- Tool quality93 / 100 · weight 13%
- Tool description quality85
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 99.1
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 77.3
- 4shanselman/sandbox-mcp ↗
sandbox-mcp
Community favouriteBest for: Testing installers or untrusted Windows files in a disposable Windows Sandbox: it exposes one-shot install-test and detonation tools plus UI Automation and screenshot evidence.
sandbox-mcp provides a standalone CLI and an optional MCP adapter with 27 tools for booting a fresh Windows Sandbox, copying and running applications, driving their UI through UI Automation, capturing screenshots and recordings, inspecting crashes, and collecting reports. It requires Windows Sandbox on the host, and setup may ask before administrator changes or a restart.
Score 67.3: show every number behind it
- Adoption66 / 100 · weight 40%
- GitHub stars66
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum≈0 / 100 · weight 14%
- Stars gained, relative to sizenot measuredno snapshot history yet
- Stars gained, absolutenot measuredno snapshot history yet
- npm download trendnot measuredno download history for the selected registry
- Tool quality71 / 100 · weight 13%
- Tool description quality63
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 67.3
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 67.3
- 5cloudflare/mcp ↗
Cloudflare MCP Server
Community favouriteBest for: Managing Cloudflare services without the full OpenAPI spec in context: it documents code mode tools for searching the spec and executing API calls.
It uses a code mode interface documented with docs, search, and execute tools to query Cloudflare developer documentation and the API spec and to call Cloudflare APIs. Access requires OAuth authorization or a Cloudflare API token, and API tokens with Client IP Address Filtering enabled are not supported.
Score 65.0: show every number behind it
- Adoption77 / 100 · weight 40%
- GitHub stars76
- npm downloads10downloads show none of the weekday rhythm human traffic has; halved
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum82 / 100 · weight 14%
- Stars gained, relative to size100
- Stars gained, absolute86
- npm download trend44
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integrates100
- Weighted mean of the five
- 83.3
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 65.0
- 6FootprintAI/Containarium ↗
Containarium-Cloud
Community favouriteBest for: Agents that need a persistent Linux sandbox between runs: it provisions SSH-reachable unprivileged containers, plus an MCP server inside the box for shell commands.
Containarium exposes an admin surface over MCP and ships a second MCP server that lives inside each container, so an agent can run shell commands and edit files in a box that persists between runs. Choosing it means running the daemon on an Ubuntu 24.04 VM with sudo access (or using the hosted cloud offering), and tenant-to-tenant isolation is not on by default.
Score 57.6: show every number behind it
- Adoption63 / 100 · weight 40%
- GitHub stars63
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum56 / 100 · weight 14%
- Stars gained, relative to size64
- Stars gained, absolute44
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 73.8
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 57.6
Best for: Running untrusted Python data-analysis workflows that need MCP tool access: it exposes one run_python tool backed by a stateful rootless sandbox.
It exposes one tool, run_python, which executes Python in a stateful, persistent rootless sandbox and proxies stdio MCP server tools. Its isolation depends on Podman or Docker containers, so a container runtime is part of the setup.
Score 57.5: show every number behind it
- Adoption63 / 100 · weight 40%
- GitHub stars63
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance32 / 100 · weight 24%
- Last commit touching this server42
- Repository maintenance grade30
- Commit cadence5
- Momentum37 / 100 · weight 14%
- Stars gained, relative to size40
- Stars gained, absolute31
- npm download trendnot measuredno download history for the selected registry
- Tool quality81 / 100 · weight 13%
- Tool description quality73
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 57.5
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 57.5
Best for: Teams running multiple MCP clients across shared company tools: a permission-aware, event-sourced Lobu endpoint gives Claude Code, Codex, ChatGPT, and custom agents the same memory.
Lobu documents an MCP endpoint that clients including Claude Code, Codex, OpenCode, ChatGPT, Claude Desktop, and Cursor connect to through
npx @lobu/cli@latest connect, providing shared event-sourced memory, permission-aware context from connected sources, and delegation to persistent specialists. Setup requires completing OAuth and connecting the sources to share; the server's tool list is not documented.Score 56.1: show every number behind it
- Adoption59 / 100 · weight 40%
- GitHub stars59
- npm downloads0downloads show none of the weekday rhythm human traffic has; halved
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance100 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade100
- Commit cadence100
- Momentum44 / 100 · weight 14%
- Stars gained, relative to size50
- Stars gained, absolute33
- npm download trend44
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integrates100
- Weighted mean of the five
- 71.9
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 56.1
Best for: Running untrusted AI-agent or user code locally: it documents a CLI and SDKs for launching microVMs from OCI images with hardware isolation.
The README documents an
msbCLI and TypeScript, Rust, Python, and Go SDKs for running OCI-compatible images in local microVMs, plus a link to a separate MCP server. Before choosing it, note that the project labels itself beta software and requires macOS on Apple Silicon, Linux with KVM enabled, or Windows with WHP enabled.Score 55.9: show every number behind it
- Adoption99 / 100 · weight 40%
- GitHub stars99
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum91 / 100 · weight 14%
- Stars gained, relative to size85
- Stars gained, absolute100
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 93.2
- × relevance: the keyword is tagged here
- 0.60
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 55.9
Best for: Testing and debugging a running Godot game: it exposes runtime GDScript evaluation, node property inspection, input simulation, and screenshot tools.
The server connects AI assistants to Godot projects through 100 tools for launching and running projects, editing scenes and project settings, and inspecting or manipulating a running game via GDScript evaluation, node properties, signals, input, and screenshots. Before choosing it, note that the README states Godot 4.4 or later is required.
Score 55.8: show every number behind it
- Adoption71 / 100 · weight 40%
- GitHub stars68
- npm downloads21downloads show none of the weekday rhythm human traffic has; halved
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance77 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade55
- Commit cadence40
- Momentum63 / 100 · weight 14%
- Stars gained, relative to size69
- Stars gained, absolute52
- npm download trend66
- Tool quality53 / 100 · weight 13%
- Tool description quality45
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 71.5
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 55.8
- 11
Best for: Security researcher training on MCP vulnerabilities: it provides deliberately vulnerable Streamable HTTP challenge servers for prompt injection, tool poisoning, and code execution exercises.
DVMCP is an educational project that runs deliberately vulnerable MCP challenge servers as Streamable HTTP endpoints, with Docker setup and challenges covering prompt injection, tool poisoning, excessive permissions, and malicious code execution. It assumes a Docker or Linux lab environment for educational use only, and the README notes instability on Windows.
Score 49.6: show every number behind it
- Adoption78 / 100 · weight 40%
- GitHub stars78
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance71 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade30
- Commit cadence40
- Momentum44 / 100 · weight 14%
- Stars gained, relative to size41
- Stars gained, absolute49
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust0 / 100 · weight 9%
- License0
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 63.6
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 49.6
- 12
Best for: Delegating execution-heavy coding work from Claude Code or Codex CLI to DeepSeek: delegate_to_deepseek gives DeepSeek its own file and Bash tool loop for the task.
It exposes two MCP tools, ping and delegate_to_deepseek, the latter running a focused task through DeepSeek's own file and Bash tool loop. Before installing it, Python 3.10 to 3.12 must already be present, and DeepSeek's hosted API needs a key in ~/.deepseek-mcp/config.json on POSIX or DEEPSEEK_API_KEY on Windows.
Score 48.2: show every number behind it
- Adoption45 / 100 · weight 40%
- GitHub stars45
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum76 / 100 · weight 14%
- Stars gained, relative to size97
- Stars gained, absolute44
- npm download trendnot measuredno download history for the selected registry
- Tool quality96 / 100 · weight 13%
- Tool description quality88
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 72.7
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: modest but real audience
- 0.85
- Composite score
- 48.2
Best for: RF/microwave design automation: it synthesizes device geometry from closed-form physics, drives solver adapters, runs results through quality gates, and optimizes dimensions, all callable over MCP.
rfauto exposes RF/microwave design and simulation workflows over MCP: closed-form synthesis of device geometry from a YAML recipe, adapters for solver engines such as HFSS, openEMS, Elmer and ngspice, optimizers and tuning campaigns, and quality gates that report a failing result with the gate name and margin. Before choosing it, note that no npm package is published and no PyPI package is attributed to it, so installation is from the repository on Python 3.10, 3.11 or 3.12, with the commercial engines treated as optional extras that a built-in fake channel substitutes for.
Score 46.1: show every number behind it
- Adoption41 / 100 · weight 40%
- GitHub stars41
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum87 / 100 · weight 14%
- Stars gained, relative to size100
- Stars gained, absolute67
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 69.5
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: modest but real audience
- 0.85
- Composite score
- 46.1
- 14
Best for: For AI agents working with a running Roblox game client: it executes Lua, decompiles scripts, searches instances, and intercepts remotes.
Roblox Executor MCP is an MCP server whose README documents a local dashboard and capabilities for running Lua code, inspecting and decompiling scripts, searching instances, intercepting remotes, clicking GUI controls, and capturing screenshots. It assumes a Roblox executor with loadstring and HTTP request support, and its screenshot capture is limited to Windows.
Score 45.7: show every number behind it
- Adoption46 / 100 · weight 40%
- GitHub stars43
- npm downloads20downloads show none of the weekday rhythm human traffic has; halved
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance92 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence85
- Momentum73 / 100 · weight 14%
- Stars gained, relative to size100
- Stars gained, absolute62
- npm download trend39
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 68.9
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: modest but real audience
- 0.85
- Composite score
- 45.7
Best for: Sending transactional email from an MCP client with Mailtrap: it exposes send-email plus sandbox send and message inspection tools.
Mailtrap Email Sending exposes eight tools for sending email, managing templates, and testing messages in a sandbox inbox. It requires a Mailtrap API token, and template management and sandbox message listing or showing also require a Mailtrap account ID.
Score 43.0: show every number behind it
- Adoption66 / 100 · weight 40%
- GitHub stars46
- npm downloads59
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance95 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence100
- Momentum25 / 100 · weight 14%
- Stars gained, relative to size0
- Stars gained, absolute0
- npm download trend100
- Tool quality76 / 100 · weight 13%
- Tool description quality68
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 71.7
- × relevance: the keyword is tagged here
- 0.60
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 43.0
Best for: Using a running Jupyter Server to expose selected Python functions as MCP tools: it dynamically registers configured functions and provides a stdio proxy.
Jupyter Server MCP is a Jupyter Server extension that dynamically loads and registers Python functions as MCP tools, serving them over Streamable HTTP and through a stdio proxy. It must be installed into the same environment as Jupyter Server or JupyterLab, and the available tools depend on functions configured through mcp_tools or exposed via package entrypoints.
Score 41.4: show every number behind it
- Adoption38 / 100 · weight 40%
- GitHub stars38
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance92 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence85
- Momentum26 / 100 · weight 14%
- Stars gained, relative to size34
- Stars gained, absolute14
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 59.1
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: small audience
- 0.70
- Composite score
- 41.4
Best for: Local Python scripting with LLM assistance across system or conda environments: it exposes environment listing, code and file execution, package installation, and file read/write tools.
The server exposes tools to list Python environments and installed packages, execute inline or file-based Python, install packages, read and write files, and list Python files in a directory. Executed Python runs with the server user's permissions and is not sandboxed, and working-directory path checks cover file tools and run_python_code(save_as=...), so isolate the server or run only trusted code.
Score 41.0: show every number behind it
- Adoption56 / 100 · weight 40%
- GitHub stars51
- PyPI downloads35PyPI downloads show no weekday rhythm; halved
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance83 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence40
- Momentum10 / 100 · weight 14%
- Stars gained, relative to size0
- Stars gained, absolute0
- PyPI download trend40
- Tool quality71 / 100 · weight 13%
- Tool description quality63
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 61.8
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: modest but real audience
- 0.85
- Composite score
- 41.0
- 18BIGPPWONG/EdgeBox ↗
EdgeBox
Abandoned but popularBest for: Agents that must drive a desktop GUI, not just a shell: it adds a VNC Ubuntu desktop with mouse, keyboard and screenshot control.
EdgeBox is a local desktop application built on the E2B Code Interpreter project that exposes, over an MCP HTTP interface, a bash shell, Python and Node.js code execution, an isolated per-session filesystem, and GUI automation tools for mouse, keyboard and screenshot capture inside a VNC desktop with Chrome and VS Code preinstalled. Its sandbox runs as a Docker container, so Docker is a prerequisite to confirm before installing, and no npm or PyPI package is published, which means setup comes from the application release rather than a package manager.
Score 40.9: show every number behind it
- Adoption58 / 100 · weight 40%
- GitHub stars58
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance46 / 100 · weight 24%
- Last commit touching this server68
- Repository maintenance grade30
- Commit cadence5
- Momentum0 / 100 · weight 14%
- Stars gained, relative to size0
- Stars gained, absolute0
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 52.4
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 40.9
- 19instavm/open-skills ↗
open-skills
Abandoned but popularBest for: Processing sensitive documents locally on an Apple Silicon Mac: it runs Claude Skills inside Apple's native containers, so files never leave the machine.
It runs an MCP server at http://open-skills.local:8222/mcp, with configuration examples for Claude Desktop, OpenCode, Gemini CLI, OpenAI's Python agents library, Qwen CLI and Kiro, and it executes Anthropic-style skills stored under ~/.open-skills/assets/skills. The constraint worth knowing first is scope: it installs only on a Mac with macOS on Apple Silicon and Python 3.10 or newer, through a shell script in the cloned repository, since no npm or PyPI package is published.
Score 40.8: show every number behind it
- Adoption66 / 100 · weight 40%
- GitHub stars66
- npm downloadsnot measuredno npm package
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance32 / 100 · weight 24%
- Last commit touching this server42
- Repository maintenance grade30
- Commit cadence5
- Momentum0 / 100 · weight 14%
- Stars gained, relative to size0
- Stars gained, absolute0
- npm download trendnot measuredno download history for the selected registry
- Tool quality≈70 / 100 · weight 13%
- Tool description quality≈70tool descriptions not yet scored
- Built and inspected by Glamanot measurednever built and inspected by Glama
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 52.3
- × relevance: the keyword is declared here
- 0.78
- × continuity: actively changing
- 1.00
- × evidence: widely adopted
- 1.00
- Composite score
- 40.8
Best for: Teams running AI agents that need isolated shell access with stable named Linux computers: it exposes terminal_execute and supports persistent computers by ID.
It exposes one tool, terminal_execute, which runs bash commands in an isolated Debian Linux sandbox and returns stdout, stderr, exit code, and duration, with optional computer_id and temporary inputs for named or persistent use. Each call is independent, so shell variables and working directory do not persist between calls, and Docker is the default backend.
Score 40.3: show every number behind it
- Adoption36 / 100 · weight 40%
- GitHub stars15
- PyPI downloads34
- Used through Glamanot measurednot used through Glama in the last 30 days
- Maintenance88 / 100 · weight 24%
- Last commit touching this server100
- Repository maintenance grade80
- Commit cadence65
- Momentum0 / 100 · weight 14%
- Stars gained, relative to size0
- Stars gained, absolute0
- PyPI download trendnot measuredno download history for the selected registry
- Tool quality100 / 100 · weight 13%
- Tool description quality93
- Built and inspected by Glama100
- Trust100 / 100 · weight 9%
- License100
- Published by the vendor it integratesnot measurednot published by the vendor it integrates
- Weighted mean of the five
- 57.6
- × relevance: the keyword is dedicated here
- 1.00
- × continuity: actively changing
- 1.00
- × evidence: small audience
- 0.70
- Composite score
- 40.3
Questions people ask
Should I pick LLM Sandbox or Context Mode?
LLM Sandbox is published on PyPI as llm-sandbox and had 106,859 PyPI downloads in a typical week, exposes 3 tools graded B, and its last commit was 10 days ago. Context Mode had 34,083 stars, 26,082 npm downloads in a typical week, 11 tools graded A and 2,459 commits in the last 12 weeks, with a commit 0 days ago. Pick LLM Sandbox when you want a broad language sandbox; pick Context Mode when the problem is large outputs filling the agent's context, since it returns only what the script prints.
What does Jupyter MCP Server need before it will run?
It is built for JupyterLab environments and exposes cell-level edit and execution tools, with 18 tools graded A. Neither a PyPI package nor an npm package is attributed to it, so do not assume a pip or npm install will get you there. It had a commit 1 day ago and 113 commits in the last 12 weeks.
Which servers are Abandoned but popular, and how active are they?
MCP Server Code Execution Mode, EdgeBox and open-skills are each Abandoned but popular. Their last commits were 345 days ago, 207 days ago and 296 days ago respectively, and each had 0 commits in the last 12 weeks. None of the three is archived.
Jupyter MCP Server or Jupyter Server MCP for notebook work?
Jupyter MCP Server is for JupyterLab users who need an agent to edit and execute notebook cells in place; it exposes 18 tools graded A, with a commit 1 day ago and 113 commits in the last 12 weeks. Jupyter Server MCP registers configured Python functions as tools from a running Jupyter Server and provides a stdio proxy; it had a commit 34 days ago and 12 commits in the last 12 weeks. Choose Jupyter MCP Server for notebook cells and Jupyter Server MCP for selected Python functions.
What should I be careful about with sandbox-mcp?
sandbox-mcp boots disposable Windows Sandbox machines and drives their UI, with 27 tools graded A and a commit 0 days ago. Its one-shot tools run install tests and detonate untrusted files, so an agent pointed at it can execute those files by design. It is specific to Windows Sandbox, not a general Linux container runner.
Which server gives a persistent sandbox across runs?
Containarium-Cloud provisions persistent, SSH-reachable unprivileged LXC containers with an MCP server inside for shell commands; it had 1,320 commits in the last 12 weeks and a commit 0 days ago. mcp-sandbox-computer-vm-for-ai also offers named Linux computers with stable IDs and lifecycle control across Docker, Fly Machines, Modal and E2B backends, with 6 commits in the last 12 weeks and 38 PyPI downloads in a typical week. Containarium-Cloud had more commit activity in that period.
Other comparisons
- Best Figma MCP servers
- Best Obsidian MCP servers
- Best YouTube MCP servers
- Best Jira MCP servers
- Best PostgreSQL MCP servers
- Best MySQL MCP servers
- Best GitHub MCP servers
- Best Slack MCP servers
- Best Notion MCP servers
- Best Linear MCP servers
- Best Discord MCP servers
- Best MongoDB MCP servers
- Best CRM MCP servers
- Best Payments MCP servers
- Best Ecommerce MCP servers
- Best Social Media MCP servers
- Best Browser Automation MCP servers
- Best CMS MCP servers
- Best Email MCP servers
- Best PDF MCP servers
- Best Kubernetes MCP servers
- Best Docker MCP servers
- Best AWS MCP servers
- Best Spreadsheets MCP servers
- Best SEO MCP servers
- Best Web Search MCP servers
- Best Web Scraping MCP servers
- Best Vector Database MCP servers
- Best Calendar MCP servers
- Best Observability MCP servers
- Best Image Generation MCP servers
- Best Speech & Transcription MCP servers
- Best Maps MCP servers
- Best Travel MCP servers
- Best Stock Market MCP servers
- Best Web Analytics MCP servers
- Best Memory MCP servers
- Best Documentation MCP servers
- Best Security MCP servers