cd /news/developer-tools/show-hn-coding-tools-mcp-give-any-ll… · home topics developer-tools article
[ARTICLE · art-33513] src=github.com ↗ pub= topic=developer-tools verified=true sentiment=↑ positive

Show HN: Coding Tools MCP – Give any LLM agent the ability to code

Developer xyTom released Coding Tools MCP, an open-source runtime server that exposes local coding primitives like file search, patch application, and test execution to any MCP-compatible LLM agent. The tool is model-neutral and designed for secure local development without external cloud dependencies.

read6 min views1 publishedJun 19, 2026

Coding Tools MCP is a model-neutral coding-agent runtime MCP server. It exposes local coding primitives to any MCP client:

inspect repo -> search/read files -> apply structured patches -> run tests/commands
-> interact with stdin sessions -> inspect git status/diff

It is not a prompt wrapper. It does not expose external agent accounts, memory, cloud tasks, web search, image generation, model routing, plugin marketplace, or subagent orchestration as MCP tools.

QuickstartMCP client configurationRemote MCPTools and schemasPermission modesExec command recipesDocker sandboxSecurity policySecurity boundaryCI and test commandsDogfoodSWE-bench evaluationKnown limitationsTroubleshootingExec troubleshootingCompetitive analysis- Normative MCP runtime profile: docs/profile-v0.1.md

Install the published command from PyPI:

curl -fsSL https://raw.githubusercontent.com/xyTom/coding-tools-mcp/main/scripts/install.sh | bash

Install and start local Streamable HTTP against a workspace:

curl -fsSL https://raw.githubusercontent.com/xyTom/coding-tools-mcp/main/scripts/install.sh \
  | bash -s -- --start --workspace /path/to/repo

Install and expose a read-only bearer-token tunnel:

curl -fsSL https://raw.githubusercontent.com/xyTom/coding-tools-mcp/main/scripts/install.sh \
  | bash -s -- --tunnel cloudflared --auto-install-tunnel --workspace /path/to/repo

Or, from this checkout:

scripts/install.sh

Run the published package without a persistent install:

uvx coding-tools-mcp --workspace .

Use stdio for MCP clients:

uvx coding-tools-mcp --stdio --workspace /path/to/repo

If you are working from this checkout instead of a published package:

make start

Pass a different workspace, host, port, or extra server flags with Make variables:

make start MCP_WORKSPACE=/path/to/repo MCP_PORT=8000 MCP_ARGS="--permission-mode trusted"

If dependencies are missing, install the runtime in editable mode:

python -m pip install -e ".[dev]"

HTTP endpoint:

http://127.0.0.1:8765/mcp

Install the optional image extra when you want view_image

auto-resize support:

python -m pip install -e ".[image]"

Stdio:

coding-tools-mcp --stdio --workspace /path/to/repo

Set CODING_TOOLS_MCP_TRACE=1

to emit redacted JSON tool-call trace events to stderr for local debugging. Logs stay off stdout so stdio JSON-RPC remains clean.

By default, exec_command

passes a core shell environment only. For local toolchains that depend on inherited environment variables, such as MSVC developer prompts, start with:

CODING_TOOLS_MCP_SHELL_ENV_INHERIT=all coding-tools-mcp --workspace /path/to/repo

inherit=all

still filters secret-looking and /startup variables unless dangerous mode is also enabled. For local development with dependency downloads, shell expansion, and inline interpreter snippets, use:

coding-tools-mcp --permission-mode trusted --workspace /path/to/repo

--allow-network

remains available as a compatibility flag when you only want to open network-looking commands. If your MCP client does not support permission elicitation and you explicitly want to disable exec_command

permission gates inside an isolated container or VM, start with:

coding-tools-mcp --permission-mode dangerous --workspace /path/to/repo

This disables exec_command

permission gates such as network-looking commands, destructive command checks, shell expansion, inline scripts, and sensitive env checks. Workspace path boundaries for direct file tools still apply. --dangerously-skip-all-permissions

remains as a compatibility alias.

Generic stdio client:

[mcp_servers.coding_tools]
command = "uvx"
args = ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]

Claude Code:

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

Cursor:

{
  "mcpServers": {
    "coding-tools": {
      "command": "uvx",
      "args": ["coding-tools-mcp", "--stdio", "--workspace", "/path/to/repo"]
    }
  }
}

Generic Streamable HTTP clients should use MCP protocol version 2025-06-18

and point at http://127.0.0.1:8765/mcp

.

For remote MCP clients and local development over an HTTPS tunnel, keep the server bound to loopback and expose the tunnel URL with the safest profile your client can use. Anonymous tunnel testing should use read-only

mode:

CODING_TOOLS_MCP_AUTH_MODE=noauth \
CODING_TOOLS_MCP_TOOL_PROFILE=read-only \
./scripts/tunnel.sh cloudflared /path/to/repo

Configure the remote MCP client with the HTTPS tunnel URL:

URL: https://<tunnel-host>/mcp

The tunnel scripts support cloudflared

, ngrok

, and Microsoft Dev Tunnel. If the selected tunnel CLI is missing, the script asks before installing it:

scripts/tunnel.sh cloudflared /path/to/repo
scripts/tunnel.sh ngrok /path/to/repo
scripts/tunnel.sh devtunnel /path/to/repo

For clients that support custom headers, use bearer-token auth with Authorization: Bearer <token>

. For MCP clients that speak OAuth 2.1 Authorization Code + PKCE, use CODING_TOOLS_MCP_AUTH_MODE=oauth

with scripts/tunnel.sh

(or scripts/install.sh --auth-mode oauth

). The server can infer its OAuth issuer from the tunnel request URL, so one-shot tunnels like cloudflared work without setting CODING_TOOLS_MCP_SERVER_URL

before startup; set it only when you want to pin a stable issuer. The script prints a generated OAuth password, accepts any non-empty client_id by default, and lets you opt into CODING_TOOLS_MCP_OAUTH_CLIENT_ID

/CODING_TOOLS_MCP_OAUTH_CLIENT_SECRET

only when you need to lock down a confidential client. Clients that cannot send custom bearer headers and do not speak OAuth should use anonymous read-only

mode only for local/testing tunnels, or be placed behind an external auth proxy for production use.

See docs/remote-mcp.md for the exact modes and security notes.

full

: exposes all tools with truthful annotations. This is the default for backward compatibility.read-only

: recommended for remote or safe-mode clients; exposes only inspection tools, git read tools, image viewing, and default-cwd helpers.compat-readonly-all

: exposes all tools but advertises every tool as read-only for clients that gate availability onreadOnlyHint

. This is not a safety mode; mutation-capable tools such asapply_patch

,exec_command

,write_stdin

, andkill_session

can still mutate local state.

P0 tools exposed by default:

server_info

get_default_cwd

set_default_cwd

read_file

list_dir

list_files

search_text

apply_patch

exec_command

write_stdin

kill_session

git_status

git_diff

git_log

git_show

git_blame

request_permissions

Additional image tool exposed by default:

view_image

For input/output schemas and result envelopes, see docs/tools-and-schemas.md and docs/profile-v0.1.md.

The runtime binds one workspace root per server process. Paths are workspace-relative by default. Absolute paths, ..

traversal, and symlink escapes are rejected. Recursive listing/search excludes .git

, .reference

, node_modules

, target

, dist

, build outputs, virtualenvs, and common caches by default.

exec_command

runs under policy controls with workspace-bound cwd, configurable shell environment inheritance, timeout, output caps, sensitive-value and /startup environment rejection, destructive command checks, network-looking command checks, shell-expansion permission gates, indirect absolute-path checks, cancellation/kill cleanup, session deadline watchdogs, and bounded session buffers. On Linux hosts with Landlock support it also applies filesystem confinement; on Windows, macOS, or Linux hosts without Landlock, command results include a warning and external sandboxing is required before running untrusted commands. This is still not a complete OS/container sandbox; see SECURITY.md.

--permission-mode safe

is the default. --permission-mode trusted

opens local-development gates while keeping secret filtering and destructive-command checks. --permission-mode dangerous

disables exec_command

permission gates for operators who accept that risk inside an isolated runner. Do not use dangerous mode for untrusted workspaces or untrusted MCP clients.

make compliance

Compliance and CI commands are documented in docs/ci-and-tests.md. The checked-in report files are generated artifacts; inspect their suite

field before treating them as full compliance evidence.

Dogfood and SWE-bench notes live in docs/dogfood.md, docs/swe-bench.md, and BENCHMARK.md. This repository does not claim a model-generated SWE-bench leaderboard result.

make lint
make typecheck
make test
make compliance
make ci

See docs/ci-and-tests.md for the full test matrix.

This project is source-available, not open source. See LICENSE. Internal evaluation, development, testing, and security review are permitted; redistribution, hosted third-party service use, and production commercial use require prior written permission.

── more in #developer-tools 4 stories · sorted by recency
── more on @xytom 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-coding-tools…] indexed:0 read:6min 2026-06-19 ·