Skip to main content
Glama

Stateful Python REPL MCP Server

A Model Context Protocol (MCP) server giving AI agents a persistent Python REPL with honest execution semantics. Code runs in a subprocess kernel (Jupyter-style): variables survive across calls, runaway code is interruptible without losing state, and crashes never take the server down. Your other MCP servers — project, global and plugin alike — are callable in-code via a pre-injected mcp bridge.

Features

  • Persistent State: variables, imports, and functions survive across calls (~0.1s warm calls vs ~3s per fresh python3 spawn)

  • Real Timeouts: runaway code (sync or async) is interrupted at timeout seconds — KeyboardInterrupt, namespace state preserved. Cells that swallow the interrupt are killed and the kernel respawns with an explicit "variables cleared" notice

  • Crash Isolation: a segfault/OOM in REPL code kills only the kernel child; the server respawns it instantly

  • Top-level await: await client.get(url) directly — no asyncio.run() wrapper

  • Shell Composition: pre-injected sh() helper — json.loads(sh("gh pr view 1 --json title")) replaces cmd | python3 -c pipelines

  • Full Filesystem Access: open(), absolute paths, and ~ all work; cwd is your project

  • MCP Bridge: mcp.call("server", "tool", **args) reaches every MCP server Claude Code knows — project (./.mcp.json), user/global (~/.claude.json) and plugin-provided — each connected on demand, the first time you name it. Failures stay visible in mcp.failed / mcp.help()

  • Claude Code Plugin: one install bundles the server, a usage skill, and a Bash-nudge hook

Related MCP server: mcp-python-repl

Installation

# In Claude Code:
/plugin marketplace add iota-uz/repl-mcp
/plugin install python-repl@repl-mcp

Restart the session and all three components are active. Portable across machines — nothing is hand-edited in ~/.claude.json.

Migrating from a claude mcp add install? Remove the old entry first: claude mcp remove python-repl -s user. Keeping both registers two REPL server processes with duplicate tools and can skew versions between them.

What the plugin bundles:

Component

What it does

MCP server

execute_python tool, launched via uvx pinned to the release tag (cached after first run; the REPL's working directory is your project, not the plugin cache)

Skill (python-repl)

Teaches Claude when to reach for the REPL (instead of python3 -c / heredocs via Bash) and its gotchas — truncation limits, the on-demand mcp bridge, package installs

Nudge hook (PostToolUse)

When Claude runs inline Python through Bash (python3 -c, python3 - <<EOF, cmd | python3), injects a non-blocking reminder to use execute_python. Silent on python3 script.py, python3 -m ..., pytest

To update later: /plugin marketplace update repl-mcp then /plugin update python-repl@repl-mcp.

Claude Code (MCP server only)

claude mcp add python-repl -- uvx --from git+https://github.com/iota-uz/repl-mcp@v2.1.2 repl-mcp

Pin to a tag (as above) so uvx caches the build instead of fetching GitHub on every session start.

Codex CLI

codex mcp add python-repl -- uvx --from git+https://github.com/iota-uz/repl-mcp@v2.1.2 repl-mcp

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "python-repl": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/iota-uz/repl-mcp@v2.1.2", "repl-mcp"]
    }
  }
}

Manual (development)

git clone https://github.com/iota-uz/repl-mcp && cd repl-mcp
uv sync --extra dev
uv run repl-mcp                  # stdio transport (the only transport)

Usage

One tool: execute_python(code, reset=False, timeout=120).

# State persists across calls
execute_python(code="import httpx; data = (await httpx.AsyncClient().get(url)).json()")
execute_python(code="len(data['items'])")          # → 42

# Shell composition
execute_python(code="prs = json.loads(sh('gh pr list --json number,title'))")

# MCP bridge — see what's reachable (free, connects nothing)
execute_python(code="print(mcp.help())")
# Any scope: project, user/global, plugin. The named server starts on first call.
execute_python(code="mcp.call('github', 'create_issue', owner='me', repo='proj', title='Bug')")
execute_python(code="for f in files: mcp.call('telegram-mcp', 'download_media', **f)")

# Runaway code? Interrupted at timeout, state survives:
execute_python(code="while True: pass", timeout=5)
# → KeyboardInterrupt: execution interrupted. Namespace state ... preserved.

# Missing package? Install into the running env:
execute_python(code="sh('uv pip install openpyxl')")

Notes:

  • The mcp bridge sees claude.ai host connectors (Notion/Gmail/Drive/chrome) not at all — those are server-managed with nothing on disk, so call their tools directly. Everything configured locally is reachable; see Scopes below.

  • mcp.call arguments must be JSON-serializable (they cross the kernel process boundary).

  • Output truncates at 50KB (stdout) / 20KB (return values) — aggregate in-REPL.

  • reset=True clears variables but keeps sh/mcp.

MCP bridge scopes

Discovery mirrors Claude Code's own config layout. On a name collision the highest-precedence scope wins; the loser stays reachable as project:name / user:name / plugin:id:name.

Precedence

Scope

Source

1

local

~/.claude.jsonprojects["<cwd>"].mcpServers

2

project

<cwd>/.mcp.jsonmcpServers (or --config)

3

user (global)

~/.claude.jsonmcpServers

4

plugin

each enabled plugin's .claude-plugin/plugin.jsonmcpServers

Servers listed in disabledMcpjsonServers are skipped. This REPL server itself is always excluded, so mcp.call can never fork a nested bridge.

Discovery runs at startup and spawns nothing — a server process starts only when you name it in mcp.call() (~1-3s the first time, warm after). print(mcp.help()) shows every available server with its scope and status without connecting anything.

Security: in-REPL code can now start any of your configured MCP servers with your credentials. Narrow it with --mcp-scope project,local (or --mcp-scope none to disable the bridge entirely).

Architecture (v2: subprocess kernel)

MCP client ── stdio ──► PARENT (FastMCP, pure async)        CHILD (owns namespace)
                          execute_python ── EXECUTE ──────►  exec / await cell
                                       ◄──── RESULT ──────   captured output
                          timeout: SIGINT ────────────────►  KeyboardInterrupt
                          crash: respawn + clear notice      (state survives)
                     MCP sessions (on demand)  ◄─ MCP_CALL ─ in-code mcp.* proxy

The server's event loop never blocks on REPL code; in-cell mcp.* calls are serviced on an independent channel while the cell runs. See CLAUDE.md for the full development guide.

v2.1.2 changes

  • Restores startup with MCP Python SDK v2 / FastMCP v4.

  • Migrates Streamable HTTP clients to streamable_http_client and httpx2.

  • Constrains dependency major versions so future resolver changes cannot silently select an incompatible SDK generation.

v2.1.1 changes

Discoverability fixes — v2.1.0 made global servers reachable, but an agent still had to know that:

  • The mcp bridge is named in the first paragraph of the execute_python description. Clients that defer tools show agents a truncated description; everything from Helpers: down was being cut, so the bridge was invisible exactly when it mattered

  • repr(mcp) now names the reachable servers instead of just listing its own methods

  • mcp.servers renders as <available: [...] | live: [...]> — a bare list read as "these are running"

v2.1.0 changes

  • Global MCP servers are reachable: the bridge merges user-scope (~/.claude.json), local, project and plugin configs instead of only ./.mcp.json

  • On-demand connect: naming a server starts that one server; a session that never touches mcp.* still spawns zero child processes. Failed connects are remembered briefly so a loop over a dead server doesn't pay the timeout each iteration

  • mcp.servers now lists what is available (any scope), not just what happens to be connected

  • ${VAR} expansion applies to command/args/url too; unset vars fail the connect with a clear reason instead of exec'ing an empty command

  • New --mcp-scope flag (all by default)

v2.0.0 breaking changes

  • Removed (zero observed usage across real agent transcripts): workspace/git/ast_utils/code pre-injected utilities (use open()/pathlib/sh('git …')), %magic commands and object? queries, the inject parameter, mcp.tools.<server>.<tool> dot-style access and discover_tools() (use mcp.call/mcp.list_tools), SSE transport (stdio only)

  • Changed: execution moved to a subprocess kernel — timeout is now actually enforced; kernel restarts are reported explicitly

  • Added: top-level await, mcp.failed, lazy MCP connect

  • Install footprint dropped ~350MB (tree-sitter removed)

Development

uv run pytest tests/ -v          # full suite

See CLAUDE.md for architecture details, test map, gotchas, and the release process.

License

MIT

Available Tools

1 tool
execute_pythonExecute PythonA

Persistent Python REPL — use instead of python3 -c, heredocs or cmd | python3 via Bash. Also the way to BATCH MCP WORK: the injected mcp bridge reaches your project, global (user-scope) and plugin MCP servers, so one loop replaces N separate tool calls — for f in files: mcp.call('telegram-mcp', 'download_media', **f).

State (variables, imports, functions) survives across calls: a warm call takes ~0.1s vs ~3s for each fresh python3 Bash spawn. Full filesystem access — open(), absolute paths, and ~ all work. Top-level await is supported (e.g. await client.get(url) with httpx).

ParametersJSON Schema
NameRequiredDescriptionDefault
codeYesPython code to execute
resetNoClear namespace (keeps sh/mcp helpers)
timeoutNoMax execution seconds (default 120). Enforced for real: runaway code is interrupted (KeyboardInterrupt) with namespace state preserved.

TDQS

A4.6/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

With no annotations, the description carries the full burden, and it delivers: state persists across calls, performance characteristics are quantified, full filesystem access is disclosed, and top-level await is supported. The schema further clarifies that timeouts interrupt with KeyboardInterrupt and preserve namespace state.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Dense but purposeful: every sentence adds value, from purpose and alternatives to the MCP batch example and performance claims. Core concepts are front-loaded, and the inline example earns its place.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a complex tool, coverage is strong: persistence, MCP bridge, filesystem access, await, reset behavior, and timeout semantics are all covered. The only notable gap is that the description does not state what the tool returns (stdout, exceptions, etc.), which matters slightly more given no output schema is present.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, and the schema fully documents code, reset (clears namespace, keeps sh/mcp helpers), and timeout (real enforcement, KeyboardInterrupt, preserved state). The description enriches what code can do (await, mcp bridge) but adds no parameter-specific semantics, so the baseline 3 applies.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description opens with a specific verb and resource: 'Persistent Python REPL'. It immediately differentiates itself from alternatives like `python3 -c`, heredocs, and `cmd | python3` via Bash, and highlights distinctive features (state persistence, MCP bridge, top-level await).

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

It explicitly says 'use instead of' Bash Python invocations and explains why (persistence and ~0.1s warm vs ~3s spawn). It also declares itself the way to batch MCP work, giving a concrete example, so an agent has clear grounds to select it over alternatives.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 1 tool updatev2.1.1
    • First observedexecute_python

TDQS

A4.7/5.0

Scored across 1 tool

Disambiguation5/5

With only one tool, there is no possibility of confusion or overlap. The tool's purpose is unambiguous and clearly described.

Naming Consistency5/5

The single tool follows a clear verb_noun pattern (execute_python), and there are no other tools to create inconsistency.

Tool Count5/5

The server is purpose-built as a Python REPL, and a single tool fully satisfies that narrow scope. The tool is powerful and not trivial, so the count is appropriate.

Completeness5/5

The tool covers all expected REPL functionality: persistent state, filesystem access, top-level await, and MCP bridging. There are no obvious gaps for the stated domain.

Maintenance

ActivityMaintained
ResponsivenessUnresponsive

Related MCP Connectors

Related MCP Servers

  • A
    license
    A
    quality
    D
    maintenance
    Universal Python code execution MCP server that lets LLMs write and run Python for any task, with auto-install packages, streaming output, and automatic file display.
    9
    1
    MIT
  • A
    license
    A
    quality
    D
    maintenance
    A production-grade MCP server providing a persistent Python REPL with multi-session support, sandboxing, and timeout protection, enabling LLM agents to execute Python code across multiple turns with variables that persist between calls.
    12
    1
    MIT
  • A
    license
    Not graded
    quality
    D
    maintenance
    A secure, production-grade MCP server that provides filesystem operations, AST math evaluation, and system diagnostics for LLM agents.
    MIT
  • F
    license
    C
    quality
    D
    maintenance
    An MCP server that enables AI agents to execute terminal commands and Python code on the host system, leveraging Agent Zero's battle-tested implementation with session management and smart output handling.
    4
    -