Serveur MCP OwnSERP

OwnSERP

An MCP server for the Serper.dev search API, with per-call credit accounting.

Status

Stable at 2.0. Every release ships from a tagged commit after the full local gate passes. The semantic version promise covers the command line, the profile schema, the ledger format, and the tool result shape. The Rust API of the ownserp-core crate is not covered by it, so a minor release may rename a public item there. From 1.0, the command surface, the configuration schema, the ledger format, and the tool result shape follow semantic versioning: a breaking change to any of them takes a major version, a new tool, flag, or setting takes a minor version, and a fix takes a patch. CHANGELOG.md records every change.

Contents

• Overview
• Requirements
• Installation
• Upgrading
• Quick start
• Connect a client
• API key
• Tools
• Response shape and field presets
• Example
• Cost accounting and the ledger
• Configuration
• HTTP deployment
• Security
• Troubleshooting
• Development
• Distribution
• Support
• Contributing
• License

Overview

OwnSERP exposes the Serper.dev search endpoints as MCP (Model Context Protocol) tools. A model can search the web, images, news, places, maps, videos, shopping, scholar, patents, reviews, and lens, extract a page, resolve a location name, and read its own spend, all through one local server.

It is built for agents that spend metered credits. Every tool description states its price, every result carries a cost object, and a local ledger records each call. Before a call spends anything, the server can check a per-call ceiling and a session budget and refuse the call when either would be passed. Both are off until you set them, through per_call_credit_ceiling and session_budget_credits in the profile, or OWN_SERPER_PER_CALL_CREDIT_CEILING and OWN_SERPER_SESSION_BUDGET_CREDITS in the environment. Results are shaped for a model rather than for a browser: invisible characters are stripped, long text is capped and marked, absent fields are omitted, and key order is fixed.

The server is one binary and needs no language runtime. It speaks MCP revision 2026-07-28 and answers clients on 2025-11-25. It runs over stdio by default, or over Streamable HTTP with a bearer token gate.

Requirements

• To run a published binary: no runtime, no toolchain. Prebuilt binaries cover macOS, Linux, and Windows.
• To build from source: Rust 1.89 or newer, pinned to 1.99.1 in rust-toolchain.toml.
• A Serper.dev account and API key from https://serper.dev/api-keys.
• A client that speaks MCP over stdio, or any HTTP client for the Streamable HTTP transport.

Installation

Every install path uses a prebuilt binary: no Rust toolchain and no compile step.

Release archives cover macOS ARM64 and x86_64, Linux x86_64 with glibc, Linux ARM64 with glibc, Linux x86_64 and ARM64 with static musl, and Windows x86_64.

npm

npx @devops-infinity/ownserp serve

The package downloads the archive for the current platform on install, checks it against the release checksum, and runs the binary. Node 22 or newer.

Homebrew

brew install devops-infinity/tap/ownserp

macOS and Linux both install the release binary. The tap is updated by hand after each release, so it can lag the release by a few minutes.

Shell installer

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/devops-infinity/ownserp/releases/latest/download/ownserp-installer.sh | sh

On Windows, in PowerShell:

powershell -ExecutionPolicy Bypass -c "irm https://github.com/devops-infinity/ownserp/releases/latest/download/ownserp-installer.ps1 | iex"

Release archive

VERSION=2.4.0
BASE="https://github.com/devops-infinity/ownserp/releases/download/v$VERSION"
TARGET=x86_64-unknown-linux-gnu
curl -LO "$BASE/ownserp-$TARGET.tar.gz"
curl -LO "$BASE/ownserp-$TARGET.tar.gz.sha256"
sha256sum -c "ownserp-$TARGET.tar.gz.sha256"
tar -xzf "ownserp-$TARGET.tar.gz"
install -m 0755 "ownserp-$TARGET/ownserp" ~/.local/bin/ownserp

Set TARGET to aarch64-apple-darwin or x86_64-apple-darwin on macOS, aarch64-unknown-linux-gnu or x86_64-unknown-linux-musl or aarch64-unknown-linux-musl on those platforms, and use the .zip file with x86_64-pc-windows-msvc on Windows.

The binary prints its shell completions and manual page on demand with ownserp completions <shell> and ownserp man. Install them where your shell and manual reader look, under ~/.local/share/bash-completion/completions/, a directory on fpath, ~/.config/fish/completions/, and ~/.local/share/man/man1/.

When the MINISIGN_KEY repository secret is configured, each release signs sha256.sum with minisign. To check a download, verify the signature against the public key in this repository, then check the archive against the file:

minisign -Vm sha256.sum -P RWSPyQWAsXwP5vjgAFI/PAAtbFAd5tA4BlhyFrT+MCPNmE2J+jgT26kC
grep "ownserp-x86_64-unknown-linux-gnu.tar.gz" sha256.sum | sha256sum --check

Container

docker run --rm -p 127.0.0.1:8777:8777 \
  -e SERPER_API_KEY -e OWN_SERPER_BEARER_TOKENS \
  -e OWN_SERPER_ALLOWED_HOSTS=127.0.0.1,localhost \
  ghcr.io/devops-infinity/ownserp:latest

The image serves Streamable HTTP and refuses a public bind without a bearer token, so both variables are required. OWN_SERPER_ALLOWED_HOSTS must name every Host a client sends, because the image binds 0.0.0.0 and the allowlist derived from that address matches no real Host; add the public hostname, or the name the proxy sends, to the same list. The package is public, so docker pull needs no login.

Windows

Use the PowerShell installer above, or download ownserp-x86_64-pc-windows-msvc.zip from the release page, extract ownserp.exe, and put the folder on your PATH. A client configuration that cannot resolve the name uses the absolute path to ownserp.exe.

MCPB bundle

The bundle carries the Linux x86_64 binary only. On another architecture, install the archive for that platform instead.

Download ownserp-<version>.mcpb from the release page and open it in a client that supports MCP Bundles. The bundle asks for the API key and stores it in the OS keychain.

Upgrading

ownserp update asks GitHub for the latest release and prints the command that fits this install: brew upgrade ownserp, cargo install --locked ownserp, the npm global install, or the container pull. It checks and reports only, so the upgrade runs through the channel you installed with and that channel's own signing and review stay in force. ownserp update --json gives current, latest, updateAvailable, and command for scripts. The check calls api.github.com and nothing else, and the server itself still sends no telemetry.

Quick start

export SERPER_API_KEY="your key"
npx @devops-infinity/ownserp doctor
npx @devops-infinity/ownserp serve

The first command sets the key for the shell. The second prints the resolved settings, the key source with the key masked, and the answer from both Serper hosts. The third starts the server on stdio, which is the command an MCP client launches, so it keeps running until the client stops it.

To connect a client, add the server as shown in the next section.

Connect a client

Every client runs ownserp serve as a stdio process and passes SERPER_API_KEY in the environment. When a desktop client cannot resolve ownserp on its PATH, use the absolute path printed by command -v ownserp.

Claude Code

claude mcp add --scope project ownserp -- ownserp serve

Or commit a .mcp.json at the project root:

{
  "mcpServers": {
    "ownserp": {
      "command": "ownserp",
      "args": ["serve"]
    }
  }
}

Claude Desktop

Add this to claude_desktop_config.json, at ~/Library/Application Support/Claude/ on macOS or %APPDATA%\Claude\ on Windows. Use the absolute path to the installed binary, because a desktop application does not inherit the shell PATH:

{
  "mcpServers": {
    "ownserp": {
      "command": "/absolute/path/to/ownserp",
      "args": ["serve"],
      "env": { "SERPER_API_KEY": "your key" }
    }
  }
}

The npm package installs the prebuilt binary for the current platform. Point command at ownserp when the client inherits your PATH, or at the absolute path command -v ownserp prints when it does not.

OpenAI Codex CLI

In ~/.codex/config.toml, or a project .codex/config.toml:

[mcp_servers.ownserp]
command = "ownserp"
args = ["serve"]
env_vars = ["SERPER_API_KEY"]

env_vars passes the variable from the environment, so the key never enters a command line.

Command Code

commandcode mcp add-json --scope project ownserp '{"type":"stdio","command":"ownserp","args":["serve"]}'

Cursor

In ~/.cursor/mcp.json, or .cursor/mcp.json for one project:

{
  "mcpServers": {
    "ownserp": {
      "command": "ownserp",
      "args": ["serve"],
      "env": { "SERPER_API_KEY": "your key" }
    }
  }
}

Windsurf (Cascade)

In ~/.config/devin/mcp_config.json on Linux and macOS, or %APPDATA%\devin\mcp_config.json on Windows:

{
  "mcpServers": {
    "ownserp": {
      "command": "ownserp",
      "args": ["serve"],
      "env": { "SERPER_API_KEY": "your key" }
    }
  }
}

VS Code

Run MCP: Add Server from the Command Palette, or add .vscode/mcp.json to the workspace:

{
  "servers": {
    "ownserp": {
      "type": "stdio",
      "command": "ownserp",
      "args": ["serve"],
      "env": { "SERPER_API_KEY": "your key" }
    }
  }
}

Zed

In settings.json:

{
  "context_servers": {
    "ownserp": {
      "command": "ownserp",
      "args": ["serve"],
      "env": { "SERPER_API_KEY": "your key" }
    }
  }
}

Other MCP clients

Use stdio with command ownserp and argument serve, and set SERPER_API_KEY in the client's environment block or in the shell the client inherits.

API key

Three sources work, and the process environment wins over the keychain.

• Environment variable, in the shell or in the client's env block:

export SERPER_API_KEY="your key"

• OS keychain, read from stdin so the key does not enter shell history:

printf '%s' "$SERPER_API_KEY" | ownserp config set-api-key

• Client JSON config, where the client injects the env block it holds.

The key is never read from a command line argument, never written to the profile file, and never printed unmasked. On Linux the keychain store is Secret Service, so a desktop session bus must be running; on a headless server, set SERPER_API_KEY instead. Use one named key per project or agent at https://serper.dev/api-keys, because a key cannot be scoped to a smaller spend.

Tools

Every tool description in the tool list states the price, so a model can budget before spending.

• serper_search, 1 credit. Web results with knowledge graph, answer box, and related searches.
• serper_images, 1 credit, 2 when num is above 10. Image results with dimensions and source.
• serper_news, 1 credit. Dated news results.
• serper_places, 1 credit. Local businesses filtered by a Serper location name.
• serper_maps, 3 credits. Places anchored to a map viewport in @lat,lng,zoomz form.
• serper_videos, 1 credit. Video results with channel and duration.
• serper_shopping, 2 credits. Product listings with price and delivery.
• serper_scholar, 1 credit. Papers with publication info and citation counts.
• serper_patents, 1 credit. Patents with assignee and dates.
• serper_autocomplete, 1 credit. Query suggestions, taking only q, gl, and hl.
• serper_reviews, 1 credit. Reviews for one place, with cursor pagination through nextPageToken.
• serper_lens, 3 credits. Reverse image search for a public image URL.
• serper_scrape, 2, 6, or 10 credits by page difficulty. Page content as Markdown or text, with an optional title, images, links, and videos, and the actual credits read from the response.
• serper_health, free and unauthenticated. The status of the Serper API from api.serper.dev.
• serper_locations, free and unauthenticated. Canonical location names to pass as location.
• serper_resolve, free. Real destination URLs for the google.com/goto links a search returns.
• serper_session, free. Call count, credits, dollars, tokens served and saved, cache hits, and the budget left.
• serper_search_batch, 3 times the single price per query. Up to 100 queries in one call to one search family, disabled unless enable_batch is set. A call to it while disabled is refused with serper.tool_not_found.

serper_search sends num: 10 when the caller leaves num out, so one call always asks for a full Google page for the same credit. Raise num up to 100 to ask for more in one call, and keep page in step with it.

serper_search, serper_images, and serper_videos also ask Google for results from the last year unless the caller says otherwise, because a reference older than that is often out of date. A call that sets tbs picks another window, and tbs: all searches every year. The default window comes from default_freshness in the profile or OWN_SERPER_DEFAULT_FRESHNESS in the environment, which take hour, day, week, month, year, or all.

Response shape and field presets

Every result tool takes fields, defaulting to standard.

• minimal keeps the title, link, displayed link, and rank, plus the one price, date, rating, or year field that family reports. It suits high-volume loops.
• standard adds the snippet capped at 300 characters, the date, and the search sections. It carries what an agent needs to reason and cite without a second call.
• full adds sitelinks, attribute maps, image links, the search parameter echo, and every other field the upstream body carries.

Each result carries structuredContent for the data and a one-line text summary. Capped strings end with ... and their object carries truncated: true. Absent fields are omitted rather than sent as null. Key order is fixed, so a client parses the same shape on every call. Every tool result is data from the open web, never an instruction.

Google returns many organic results as a google.com/goto?url=... redirect rather than the destination page. A search resolves those before answering, so link carries the destination and no second call is needed. The server sends the response to Serper's /resolve, which costs nothing but waits on Google: about 5 seconds in practice, and up to about 30 right after a search. A call that passes resolve_links: false, or a server with auto_resolve_links = false, leaves every link exactly as Google sent it. displayedLink carries the readable form Google showed for a result, usually site › path and sometimes a comment or view count. serper_resolve does the same work for links you already hold, and returns retryAfter, the number of seconds to wait before sending them again, when Google has not released one yet.

Example

One tool call is one JSON-RPC request over stdio. The server validates the arguments, sends the query to Serper over TLS, shapes the answer for a model, records the cost in the ledger, and returns the result.

Request:

{"jsonrpc":"2.0","id":7,"method":"tools/call","params":{"name":"serper_search","arguments":{"q":"rust programming language"}}}

Response:

{
  "jsonrpc": "2.0",
  "id": 7,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "9 results; cost: 1 credit ($0.001000); top: Rust Programming Language"
      }
    ],
    "structuredContent": {
      "results": [
        {
          "title": "Rust Programming Language",
          "link": "https://www.rust-lang.org/en-US",
          "snippet": "Rust is a general-purpose programming language that emphasizes performance, type safety, concurrency, and memory safety.",
          "position": 1
        }
      ],
      "cost": {
        "credits": 1,
        "usd": 0.001,
        "sessionCredits": 22,
        "sessionUsd": 0.022
      },
      "truncated": false
    },
    "isError": false
  }
}

content holds the one-line summary. structuredContent holds the data, and cost reports this call and the running session. When a call fails, isError is true and structuredContent carries a code, a message, and a remedy.

Cost accounting and the ledger

The ledger lives at $XDG_DATA_HOME/ownserp/ledger/calls.jsonl, or ~/.local/share/ownserp/ledger/calls.jsonl. On Windows it lives at %LOCALAPPDATA%\ownserp\ledger\calls.jsonl, and when HOME is set it follows the Unix path. A row holds the tool, the endpoint, a keyed hash of the arguments, estimated credits, actual credits when the response reports them, duration, status, bytes, and token counts. Query text is never written.

A call that Serper rejects with a 4xx or 5xx status, answers with a rate limit, or answers with a redirect is recorded at zero credits, because nothing was billed. A timeout or a cancel is recorded at the estimate, because the request may have reached Serper and been billed.

ownserp audit verify

On Unix the ledger directory is owner-only (0700) and the file is owner-only (0600). On Windows it sits under %LOCALAPPDATA%, so it inherits the user profile's access rules. One line is capped at 1 MiB. Any number of ownserp serve processes on one machine share this one file, one hash chain, and one running total. A process takes the file lock only while it reads the rows written since its last call and appends its own row, so no process holds the lock for as long as it runs. A final row that only lost its newline is kept, because its hash proves it was written; anything less complete is dropped under that lock, and the chain check covers every complete row. Each row carries the format version, and a file written by a newer binary is refused by name instead of being read as damage. On Windows a running server does not notice a rotate, so restart it to pick up the new chain. The arguments hash is an HMAC, a keyed hash, keyed by args.key, created beside the ledger directory with owner-only permissions, so the ledger alone cannot confirm a guessed query.

session_budget_credits counts every credit in the chain, whichever process wrote it. A credit that another process has reserved but not yet written is invisible to that check, so the budget can be passed by at most the number of running processes times the per-call credit ceiling. That bound holds only when a ceiling is set. With no ceiling it is the process count times the largest single reservation, which is 900 credits for a 100-query maps or lens batch. The response cache stays per process: each process keeps its own cache. OWN_SERPER_LEDGER_DIR still gives a client its own separate ledger when that is what you want.

To close a ledger without losing it, run ownserp audit rotate. It verifies the chain and archives the file beside the original with a timestamp. Every running server opens a fresh chain on its next write. When the chain is already damaged, ownserp audit rotate --archive-damaged archives the file up to the last row that verified, so a fresh chain can start and the spend history stays readable. ownserp audit remove --yes deletes the live ledger, its archives, and the arguments key, and prints each path it removed; without --yes it only says what it would remove.

Dollar figures come from the pack tier: starter (default), standard, scale, ultimate, or custom with SERPER_USD_PER_1K.

Configuration

Environment variables:

• SERPER_API_KEY
• SERPER_USD_PER_1K
• OWN_SERPER_PROFILE
• OWN_SERPER_BIND
• OWN_SERPER_CACHE_TTL_SECS
• OWN_SERPER_LEDGER_DIR
• OWN_SERPER_BEARER_TOKENS
• OWN_SERPER_ALLOWED_HOSTS
• OWN_SERPER_LOG_FORMAT, set to json for one JSON object per log line

Profile file at $XDG_CONFIG_HOME/ownserp/profile.toml, or ~/.config/ownserp/profile.toml. On Windows it is %APPDATA%\ownserp\profile.toml when HOME is unset:

version = 1
pack = "starter"
usd_per_1k = 1.0
cache_ttl_secs = 0
cache_max_bytes = 16777216
ledger_dir = "/absolute/path/to/ownserp-ledger"
per_call_credit_ceiling = 10
session_budget_credits = 500
enable_batch = false
http = false
bind = "127.0.0.1:8777"
request_timeout_secs = 30
default_freshness = "year"
auto_resolve_links = true

version is the profile schema version. It is 1 today, and a profile that names a newer version is refused by name.

ledger_dir takes an absolute path. The file is a TOML document, so a ~ or a variable is stored literally rather than expanded.

session_budget_credits counts every credit written to the current ledger chain, across restarts, and ownserp audit rotate starts a fresh count. allowed_hosts lists the Host names HTTP mode accepts; leave it unset to derive them from the bind address.

A custom pack with no usd_per_1k reports credits with no dollar figure, because the rate is unknown.

The cache is off by default. cache_ttl_secs above zero turns it on; a hit answers with cacheHit: true and charges nothing. A call that sets tbs itself skips the cache, and the default freshness window does not.

ownserp doctor prints the resolved settings, including the default freshness window, masks the key, names where the key came from, and probes both hosts. ownserp config show prints the key source, the masked key, the credit pack, the ledger directory, and the profile path.

HTTP deployment

OWN_SERPER_BEARER_TOKENS="a random value of at least 32 characters" \
  ownserp serve --http --bind 127.0.0.1:8777

/healthz is open. /readyz and /metrics require the bearer token whenever OWN_SERPER_BEARER_TOKENS is set, and every other path is the MCP endpoint. The server refuses a non-loopback bind without a token and refuses a token shorter than 32 characters. It checks the Host header against the allowlist on every protected route, caps a request body at 1 MiB and gives the body 30 seconds to arrive, accepts at most 256 connections, holds at most 128 requests at once, derives each request deadline from request_timeout_secs and the retry count, and cancels the call when the deadline passes, rejects a request that carries an Origin header, and marks every response x-content-type-options: nosniff with cache-control: no-store. Every bearer token on one server shares that server's ledger, session budget, and cache. Behind a proxy, list the names it sends in OWN_SERPER_ALLOWED_HOSTS and let the proxy enforce its own header and idle timeouts.

Security

The server holds one API key, sends queries only to Serper over TLS, and sends no telemetry. The TLS root set is bundled into the binary, so a container without ca-certificates still verifies Serper; platform roots are added when they load. Results are treated as untrusted text. SECURITY.md states the reporting process, and the threat model states the assets, trust boundaries, controls, and residual risks.