pline.ai

pline.ai MCP server

Run this MCP server on your own machine to scrape pages, crawl sites, discover URLs, and search through the hosted pline.ai API. Each user supplies their own API key.

Your MCP client → local MCP process (stdio) → hosted pline.ai API (HTTPS)

The scraping backend stays on pline.ai's infrastructure. You do not need to run the orchestrator, browser fetcher, Redis, or Temporal locally.

Run with npx

The server is published to npm as pline-api: a small launcher that selects the prebuilt binary for your platform (macOS arm64/x64, Linux x64/arm64, Windows x64) from an optional dependency, falling back to a checksum-verified download from the GitHub Release. Node.js 18+ is the only requirement:

{

"mcpServers": { "pline.ai": { "command": "npx", "args": ["-y", "pline-api"], "env": { "PLINE_BASE_URL": "https://apix.pline.ai/v1", "PLINE_API_KEY": "your-own-api-key" } } } }

Arguments pass straight through to the binary (npx -y pline-api --http 127.0.0.1:8080), and PLINE_MCP_BINARY=/path/to/pline-mcp makes the launcher run a build of your own. The launcher source is in ../npm; per-platform packages are named pline-api- - .

Install a prebuilt binary

Use the platform downloads and checksum instructions in the repository setup guide. Extract pline-mcp (or pline-mcp.exe on Windows) and configure your MCP client with its absolute path below. You do not need Rust, Cargo, or a repository checkout for a prebuilt installation.

Run the Docker image

Each release is also published as a multi-platform image at ghcr.io/grepsr/pline-mcp (linux/amd64 and linux/arm64), which is the package the MCP Registry entry io.github.grepsr/pline-api points at. With Docker installed, an MCP client can run the server over stdio without downloading a binary:

{

"mcpServers": { "pline.ai": { "command": "docker", "args": ["run", "-i", "--rm", "-e", "PLINE_BASE_URL", "-e", "PLINE_API_KEY", "ghcr.io/grepsr/pline-mcp:0.3.12", "--stdio"], "env": { "PLINE_BASE_URL": "https://apix.pline.ai/v1", "PLINE_API_KEY": "your-own-api-key" } } } }

-e NAME without a value forwards the variable from the client-provided environment into the container, so the key never appears on the command line. The image runs as an unprivileged user and has no access to your filesystem, so save_dir and download_dir are unavailable unless you mount a volume. Pin the tag to a release version; latest moves. The same image serves Streamable HTTP when started without --stdio:

docker run --rm -p 8080:8080 -e PLINE_BASE_URL=https://apix.pline.ai/v1 ghcr.io/grepsr/pline-mcp:0.3.12

Build from Git

You need Git, a current stable Rust toolchain with Cargo, and an API key. The server calls the public API at https://apix.pline.ai/v1 unless PLINE_BASE_URL overrides it.

From the repository root:

git clone https://github.com/grepsr/pline-api.git

cd pline-api cargo build --manifest-path mcp-server/Cargo.toml --release --locked cp .env.example .env

Edit .env with your API URL and your own key, then load it before starting your MCP client:

set -a

. ./.env set +a

These environment-loading commands are for POSIX shells. On other platforms, set the same two environment variables in your shell or MCP client configuration. The server does not load .env automatically; the file is ignored by Git.

PLINE_BASE_URL is optional and defaults to the public API, https://apix.pline.ai/v1. Internal deployment addresses are not reachable from a user's machine.

Connect an MCP client

For a client with a mcpServers JSON configuration, add this entry. Replace the binary path, API URL, and key with your own values. On Windows the binary has an .exe suffix.

{

"mcpServers": { "pline.ai": { "command": "/absolute/path/to/pline-mcp", "args": ["--stdio"], "env": { "PLINE_BASE_URL": "https://apix.pline.ai/v1", "PLINE_API_KEY": "your-own-api-key" } } } }

Use an absolute binary path so the client can launch it from any working directory. For source builds, the executable is mcp-server/target/release/pline-mcp inside the checkout. Store real credentials in your local client settings, not in a committed configuration file.

For Claude Code launched from the repository root, the root .mcp.json runs the server through Cargo and expands the two environment variables from your shell. The first release build can take a while; building it first avoids waiting during the client's connection attempt.

The MCP client launches and manages the stdio process. You do not need to start an HTTP listener or expose a port. Logs go to stderr (RUST_LOG=debug for more detail).

Update

For prebuilt installs, download the new archive, verify its checksum, and replace the executable. For source builds:

git pull --ff-only

cargo build --manifest-path mcp-server/Cargo.toml --release --locked

Restart or reconnect the MCP client to load the rebuilt binary.

Tools

Tool results contain structured JSON. API errors return the HTTP status and the service's detail message. The client sends your key to the configured API as x-api-key.

scrape remembers sessions per caller and host for the process lifetime. A second scrape of the same host reuses the session unless you pass session_id or set new_session=true. Same-host scrapes are serialized locally. A 4xx response on a remembered session drops it for the next call.

Over stdio, save_dir writes scrape outputs to your local filesystem and returns their paths. Without it, outputs are clipped at max_chars (default 60000); truncated_outputs lists clipped fields. Over stdio, the status tools can download finished job artifacts to a local download_dir. HTTP requests with save_dir or download_dir are rejected before contacting the API.

Read pline://guide for usage guidance and pline://reference for the API fields and constraints. The documents are also available in docs/guide.md and docs/reference.md.

Hosted OAuth connection

The remote MCP endpoint can use Pline sign-in instead of asking every agent user to paste an API key. After sign-in, each user selects a workspace and an active API key. Follow docs/oauth-hosting.md to configure the Pline web app and host.

Configuration

The executable is pline-mcp; its public MCP identity is pline.ai. Set PLINE_BASE_URL and PLINE_API_KEY for local stdio use. For HTTP mode, PLINE_MCP_HTTP_ADDR sets the listener address and PLINE_MCP_ALLOWED_HOSTS accepts a comma-separated list of allowed hosts. OAuth mode also uses PLINE_PLATFORM_BASE_URL (Pline web app origin), PLINE_SUPABASE_URL (Supabase project origin), and PLINE_MCP_PUBLIC_URL (public MCP origin).

Where this server is listed

The server is published to the official MCP Registry as io.github.grepsr/pline-api with the npm package and the GHCR image above as its packages, and the repository root is installable as a Claude Code plugin (pline-ai) that runs the npm package. The root .cursor/mcp.json and .vscode/mcp.json configure npx -y pline-api for Cursor and VS Code workspaces. PUBLISHING.md describes how each listing is produced and kept current.

HTTP mode

The binary also supports Streamable HTTP, for example for local transport development:

PLINE_BASE_URL=https://apix.pline.ai/v1 \

./mcp-server/target/release/pline-mcp --http 127.0.0.1:8080

/health reports liveness. /openapi.json serves a public OpenAPI 3.1 description of the server, generated from the live tool list. /mcp accepts the caller's API key through Authorization: Bearer or x-api-key; the legacy / /mcp route also exists, but puts credentials in URLs and access logs. The server forwards each caller's key rather than using the stdio environment key. A Supabase OAuth access token is also accepted when the platform and OAuth environment variables are configured. The user selects an active workspace API key in Pline's OAuth consent screen, and the server resolves it for each request through the authenticated Pline platform. The MCP process does not keep a shared API key.

Configure Supabase OAuth 2.1 with dynamic client registration enabled and /oauth/consent as its authorization UI path. Apply API service migrations with uv run alembic upgrade head from the pline-hades-api-service repository before enabling the hosted endpoint.

HTTP transport is stateless: no Mcp-Session-Id is issued, and standalone GET streams and DELETE session requests return 405. Each POST uses its own API key; remembered scrape sessions remain scoped to that key. Session replay and MCP cancellation notifications are not supported in HTTP mode. The Git installation above uses stdio.

Develop

Run from mcp-server/:

cargo test --locked

cargo clippy --locked --all-targets --all-features -- -D warnings cargo fmt --check

The guide and reference are embedded with include_str!; rebuild after editing them.