Metadata-Version: 2.5
Name: bing-webmaster-ai-cli-mcp
Version: 0.1.3
Summary: bing-webmaster-ai-cli-mcp gives an AI agent read access to what Bing knows about your sites — traffic, indexing, crawl issues, inbound links, keywords — and a write path you choose: reviewed plan-and-apply by default, or direct writes when you opt in.
Project-URL: Changelog, https://github.com/stufently/bing-webmaster-ai-cli-mcp/blob/main/CHANGELOG.md
Project-URL: Documentation, https://github.com/stufently/bing-webmaster-ai-cli-mcp#readme
Project-URL: Issues, https://github.com/stufently/bing-webmaster-ai-cli-mcp/issues
Project-URL: Repository, https://github.com/stufently/bing-webmaster-ai-cli-mcp
License-Expression: MIT
License-File: LICENSE
Keywords: bing,bing-webmaster-tools,cli,indexnow,mcp,model-context-protocol,seo
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Programming Language :: Python :: 3.14
Requires-Python: >=3.12
Requires-Dist: click<9,>=8.2
Requires-Dist: httpx<1,>=0.28
Requires-Dist: mcp<3,>=2.1
Requires-Dist: pydantic-settings<3,>=2.6
Requires-Dist: pydantic<3,>=2.11
Requires-Dist: uvicorn<1,>=0.30
Provides-Extra: test
Requires-Dist: pytest-asyncio>=0.24; extra == 'test'
Requires-Dist: pytest>=8; extra == 'test'
Requires-Dist: ruff>=0.9; extra == 'test'
Description-Content-Type: text/markdown

# bing-webmaster-ai-cli-mcp

<!-- mcp-name: io.github.stufently/bing-webmaster-ai-cli-mcp -->

bing-webmaster-ai-cli-mcp gives an AI agent read access to what Bing knows about your sites — traffic, indexing, crawl issues, inbound links, keywords — and a write path you choose: reviewed plan-and-apply by default, or direct writes when you opt in.

It ships a Python 3.12+ CLI and MCP server for the JSON Bing Webmaster Tools API,
plus protocol-correct IndexNow submission. SOAP and POX are deliberately absent.

## Choosing a write path

Read tools always execute immediately. For everything that changes Bing state, one
setting picks between two paths:

```console
export BING_WM_ALLOW_WRITES=false   # default: writes are planned and applied by a human
export BING_WM_ALLOW_WRITES=true    # opt in: writes go straight through
```

**Read-only toward Bing is the default.** Without `BING_WM_ALLOW_WRITES=true` the server
lists no tool that changes Bing, and the CLI refuses one-step writes. Here is the reason.
An agent using this server also reads pages, anchor text, crawl issues and search queries
written by strangers. A confirmation an agent can send is a confirmation prompt injection
can send, so as long as the agent can write, a poisoned string in someone else's anchor
text can reach your Bing account. With the default, nothing an agent does changes Bing
until you have read the plan and run `bing-wm plan apply` yourself. Opt in only for an
agent you trust with your Bing account.

### Direct writes (`BING_WM_ALLOW_WRITES=true`)

The MCP server advertises one-step `bing_<operation>` tools, and the CLI accepts:

```console
$ bing-wm submit-url example.com https://example.com/new-page --json
{"applied": true, "plan_id": "…", "operation": "submit_url", "result": null}
```

A direct write is not idempotent: retrying one records a new plan and sends the change
again. If a response is lost, read `audit.jsonl` or `bing-wm plan list` before repeating
the call.

### Reviewed writes (default, `BING_WM_ALLOW_WRITES=false`)

The server advertises `bing_plan_<operation>` instead. Those tools send nothing:

```console
$ bing-wm plan submit-url example.com https://example.com/new-page --json
{
  "plan_id": "…",
  "summary": "submit url https://example.com/new-page for https://example.com",
  "apply_with": "bing-wm plan apply …"
}
$ bing-wm plan apply PLAN_ID
Apply this plan? [y/N]: y
```

No MCP tool takes a plan ID, in either mode: a plan recorded for review is applied or
rejected only at the CLI, by you. A direct write applies the plan it creates in the same
call and never touches one somebody else recorded. The plan path stays available while
direct writes are on, so an agent can still propose a change for review.

### What both paths share

A direct write is the same code with the human step removed: it records the same durable
plan and goes through the same apply boundary. A readable plan record, an append-only
audit trail, one-shot application, expiry, Bing's own submission quota, a site denylist
(`BING_WM_DENIED_SITES`) and restart-persistent local limits
(`BING_WM_MAX_WRITES_PER_DAY`) apply to both.

If the applying process is killed and leaves a lock behind, the command
`bing-wm plan unlock PLAN_ID` verifies that the recorded PID is gone before recovering
it. An unfinished plan becomes `unknown_outcome` and cannot be applied again; the
recovery is audited.

## Install

```console
BING_WM_ALLOW_WRITES=false uvx bing-webmaster-ai-cli-mcp
```

That one command starts the stdio MCP server. Every client block below runs the same
`uvx` command. Create an API key in Bing Webmaster Tools under Settings → API Access
and pass it as `BING_WM_API_KEY`. Writes are planned for review unless you also set
`BING_WM_ALLOW_WRITES=true` — see [Choosing a write path](#choosing-a-write-path).

The package is on [PyPI](https://pypi.org/project/bing-webmaster-ai-cli-mcp/) as
`bing-webmaster-ai-cli-mcp`. `uvx` caches what it installed and may keep running that
version; `uvx bing-webmaster-ai-cli-mcp@latest` checks PyPI again, and
`uvx bing-webmaster-ai-cli-mcp==0.1.3` pins one. This README follows `main`, so a change
listed under *Unreleased* in the [changelog](CHANGELOG.md) is not in the PyPI release
yet. Since 0.1.3, `BING_WM_ALLOW_WRITES=false` is the default and the command above and
the client blocks below would behave the same without it. They still spell it out:
0.1.2 and earlier default to `true`, so a `uvx` cache still holding one of them would go
back to direct writes if the setting were left out.
To run `main` itself, use
`uvx --from "git+https://github.com/stufently/bing-webmaster-ai-cli-mcp.git" bing-webmaster-ai-cli-mcp`.

pip (Python 3.12+) installs the same package and puts `bing-wm`,
`bing-webmaster-ai-cli-mcp` and `bing-webmaster-ai-cli-mcp-http` on `PATH` — in a
virtualenv, or with `pipx install bing-webmaster-ai-cli-mcp`:

```console
python -m pip install bing-webmaster-ai-cli-mcp
export BING_WM_API_KEY='…'
bing-wm sites list
```

The supported matrix is Python 3.12, 3.13, and 3.14. Development and images use the
exact versions in `constraints.txt`; published dependencies remain compatible floors.
Never put a real key in the repository. See [configuration](docs/configuration.md) for
all settings.

## Example prompts

- Why did my site lose clicks last week?
- Which pages is Bing failing to crawl, and why?
- Is this URL indexed, and when did Bing last fetch it?
- Submit this new page to Bing, but let me review the change first.

## Verified clients and platforms

Only what has actually been run is listed as verified. Everything else in
[Add the MCP server to your AI client](#add-the-mcp-server-to-your-ai-client) is a
documented config that nobody has checked end to end.

| Client / runtime | OS | How it runs | Status | Evidence |
|---|---|---|---|---|
| Claude Code 2.1 | Ubuntu 26.04 LTS, x86_64 | Docker image built locally from this repo, stdio, direct writes opted in | Verified | Daily use on the maintainer's host, reads and writes in the audit trail |
| `bing-wm` CLI | Ubuntu 26.04 LTS, x86_64 | Same Docker image | Verified | Maintainer's host |
| Codex CLI 0.162 | Ubuntu 26.04 LTS, x86_64 | Same Docker image, stdio | Configured, not checked separately | Maintainer's host config |
| Python 3.12, 3.13, 3.14 | Ubuntu (`ubuntu-latest`) | Lint, tests, stdio MCP smoke test | Verified in CI | `ci.yml` |
| Docker image | Ubuntu (`ubuntu-latest`) | Image builds and does not run as root | Verified in CI | `ci.yml`, job `container` |
| `uvx bing-webmaster-ai-cli-mcp` (PyPI 0.1.2) | Linux x86_64 | Install path from [Install](#install), uv 0.13 in Docker | MCP `initialize` answered, no client attached | Maintainer's host, 2026-10-12 |
| Claude Desktop (`.mcpb`) | macOS, Windows | Desktop Extension | Not verified | Bundle is built in CI, never installed in a real Desktop |
| Cursor, Windsurf, Zed | any | Configs below | Not verified | Documented only |
| macOS, Windows | — | Any client | Not verified | No CI runner on these systems |

## Add the MCP server to your AI client

Every JSON block below runs the same `uvx` command as [Install](#install).
`BING_WM_ALLOW_WRITES` is `false` here (the default, spelled out), so writes are planned
and you apply them. Set it to `true` for direct writes — see
[Choosing a write path](#choosing-a-write-path). Keep the key in a user-level config,
not in a project file you commit. The server exposes 34 Bing read tools, one local
read-only tool (`bing_indexnow_key_plan`), plan inspection, and one write tool per
supported operation — `bing_plan_<operation>` tools by default, or direct
`bing_<operation>` tools when `BING_WM_ALLOW_WRITES=true`. It exposes no plan
application or rejection tool. Restart the server after changing the setting so the
client refreshes its tool list.

### Claude Code

Project file: `.mcp.json` in the project root.

```json
{
  "mcpServers": {
    "bing-webmaster": {
      "command": "uvx",
      "args": [
        "bing-webmaster-ai-cli-mcp"
      ],
      "env": {
        "BING_WM_API_KEY": "your-api-key",
        "BING_WM_ALLOW_WRITES": "false"
      }
    }
  }
}
```

Or from the terminal:

```console
claude mcp add bing-webmaster --scope user \
  --env BING_WM_API_KEY=your-api-key --env BING_WM_ALLOW_WRITES=false \
  -- uvx bing-webmaster-ai-cli-mcp
```

### Claude Desktop

File: `claude_desktop_config.json` — macOS
`~/Library/Application Support/Claude/claude_desktop_config.json`, Windows
`%APPDATA%\Claude\claude_desktop_config.json`, Linux
`~/.config/Claude/claude_desktop_config.json`.

```json
{
  "mcpServers": {
    "bing-webmaster": {
      "command": "uvx",
      "args": [
        "bing-webmaster-ai-cli-mcp"
      ],
      "env": {
        "BING_WM_API_KEY": "your-api-key",
        "BING_WM_ALLOW_WRITES": "false"
      }
    }
  }
}
```

One click, without editing JSON: download the `.mcpb` from
https://github.com/stufently/bing-webmaster-ai-cli-mcp/releases/latest
and open it. Claude Desktop installs the extension and asks for the API key.
Direct writes stay off unless you turn Allow direct writes on. While they are off
the extension only records plans. Applying one needs the `bing-wm` CLI, which the
extension does not put on your `PATH`: install it with pip as in [Install](#install),
export the same key as `BING_WM_API_KEY` in that shell, then run `bing-wm plan apply`.

The release includes `checksums.txt` with the SHA256 of every `.mcpb`.
Download that file into the same directory as the bundle and check it before
opening the extension:

```bash
sha256sum -c --ignore-missing checksums.txt   # run next to the downloaded .mcpb
```

On macOS, filter the bundle's line and check it with `shasum -a 256 -c -`.

Releases from the next tag on also carry a GitHub build provenance attestation for each
`.mcpb`, signed in the release workflow. Check that a bundle was built by this
repository's workflow and not swapped afterwards (needs the GitHub CLI):

```bash
gh attestation verify <bundle>.mcpb --repo stufently/bing-webmaster-ai-cli-mcp
```

### Cursor

File: `~/.cursor/mcp.json` (every project) or `.cursor/mcp.json` (this project).

```json
{
  "mcpServers": {
    "bing-webmaster": {
      "command": "uvx",
      "args": [
        "bing-webmaster-ai-cli-mcp"
      ],
      "env": {
        "BING_WM_API_KEY": "your-api-key",
        "BING_WM_ALLOW_WRITES": "false"
      }
    }
  }
}
```

### Windsurf

File: `mcp_config.json` — macOS and Linux `~/.config/devin/mcp_config.json`
(or `$XDG_CONFIG_HOME/devin/mcp_config.json`), Windows
`%APPDATA%\devin\mcp_config.json`. Older builds read
`~/.codeium/windsurf/mcp_config.json`.

```json
{
  "mcpServers": {
    "bing-webmaster": {
      "command": "uvx",
      "args": [
        "bing-webmaster-ai-cli-mcp"
      ],
      "env": {
        "BING_WM_API_KEY": "your-api-key",
        "BING_WM_ALLOW_WRITES": "false"
      }
    }
  }
}
```

### Zed

File: `settings.json` — Linux and macOS `~/.config/zed/settings.json`.

```json
{
  "context_servers": {
    "bing-webmaster": {
      "command": "uvx",
      "args": [
        "bing-webmaster-ai-cli-mcp"
      ],
      "env": {
        "BING_WM_API_KEY": "your-api-key",
        "BING_WM_ALLOW_WRITES": "false"
      }
    }
  }
}
```

An optional Streamable HTTP entry point is also available once the package is on
`PATH` (the pip install above; `uvx` does not install a lasting command):

```console
export BING_WM_HTTP_BEARER_TOKEN='a-random-token-of-at-least-32-characters'
bing-webmaster-ai-cli-mcp-http
```

It refuses non-loopback bind addresses and unauthenticated requests.

## Common questions

### Which crawl issues does my site have?

```console
bing-wm crawl issues example.com --json
```

The result carries Bing's rows unchanged plus a category breakdown built from
Microsoft's `UrlWithCrawlIssues.CrawlIssues` flags — redirects, 4xx, 5xx, robots.txt
blocks, malware, DNS and timeout errors — with a count per category, a count per raw
`HttpCode`, and an `other` bucket so nothing Bing sends is dropped. A 4xx row is split
further into `http_404` or `http_403` from its own `HttpCode`, alongside the broad
`http_4xx` rather than instead of it. Bing has no `noindex` crawl-issue flag, so there
is no such category and this project does not invent one.

### How do I check if Bing has indexed my page?

```console
bing-wm index url example.com https://example.com/page
```

### How do I submit URLs to Bing from the command line?

With writes enabled, from one URL or a newline-delimited file:

```console
bing-wm submit-url example.com https://example.com/page
bing-wm submit-urls example.com --file PATH
```

With `BING_WM_ALLOW_WRITES=false`, create a plan, review it, then apply it:

```console
bing-wm plan submit-url example.com https://example.com/page
bing-wm plan submit-urls example.com --file PATH
bing-wm plan show PLAN_ID
bing-wm plan apply PLAN_ID
```

The planner calls `GetUrlSubmissionQuota`; no quota number is hardcoded. Bing's method
documentation also limits one `SubmitUrlBatch` call to 500 URLs.

### Does IndexNow work with Google?

Google does not participate in IndexNow. The live IndexNow registry currently lists
Bing, Yandex, Seznam, Naver, Yep, Internet Archive, and Amazonbot. Submission through
`api.indexnow.org` fans out to participating engines.

Generate a key, host the displayed UTF-8 key file, then submit:

```console
bing-wm indexnow key example.com
bing-wm indexnow key example.com --key KEY   # is the key file live yet?
bing-wm indexnow submit example.com --file PATH --key KEY
```

`indexnow key` — and the matching `bing_indexnow_key_plan` MCP tool — only computes and
checks. It talks to neither Bing nor `api.indexnow.org`, so it is not a write and needs
no plan. The generated key is printed once and stored nowhere: save it, serve it at the
printed URL, then submit.

Or, with `BING_WM_ALLOW_WRITES=false`:

```console
bing-wm plan indexnow example.com --file PATH --key KEY
bing-wm plan apply PLAN_ID
```

The submission checks the exact key-file URL without following redirects before it sends
the batch. IndexNow hosts must use multi-label DNS names rather than IP literals or
single-label names, and the key file must
contain only the key. Batches above 10,000 URLs, ambiguous dot-segment paths, and URLs
outside the authorized host or key subpath are rejected locally. There is intentionally
no unverified 2,048-character URL cap. IndexNow's protocol tells callers to resubmit a
valid request after a non-success response, so an HTTP 5xx leaves the plan pending; the
CLI never retries it automatically.

## Output safety

Anchor text, crawl-issue URLs, titles, messages, and query strings may be controlled by
strangers. The shared operation layer removes control and bidirectional-formatting
characters, truncates extreme values, and returns these fields as
`{"value": "…", "untrusted": true}`. Consumers must treat them as data, never as
instructions.

Verification secrets never leave the machine by accident. `GetUserSites` returns
`AuthenticationCode` and `DnsVerificationCode` beside every site and `GetSiteRoles`
returns `DelegatedCode`; whoever holds one can claim the site in another Bing account.
All three are replaced with `[redacted: verification secret]` in every response, and the
only way to see one is `--reveal-verification-codes` typed by an operator on
`bing-wm sites list|show|roles` or `bing-wm plan show|list|apply`. No MCP tool takes that
argument, so no model — and no text a model read — can ask for a code.

Redaction is an exit boundary, not a step on the read path. The same filter covers a
write's result, the arguments a plan records — an `add_site_roles` plan holds the
`authentication_code` it will send, and showing the plan does not show the code — and
Bing's own error text, in case Bing quotes back the code it rejected. The literals to
hide come from the request body, so the cover does not depend on anyone predicting which
field a secret will next arrive in. The plan record on disk keeps the real value; a plan
that lost its code could not be applied.

The API key is covered at the same boundary, under its own marker
`[redacted: API credential]`. Microsoft documents the key only as a query-string
parameter, so every request URL carries a live credential — and a proxy naming the
address it could not reach, or Bing quoting the request back in an error, would otherwise
put that URL into the error message, the terminal and the audit trail. An error carrying
no credential reads exactly as it did before.

An empty answer is never presented as a measurement. Some of the older endpoints return
an empty collection for accounts that demonstrably have data: `bing_link_counts`,
`bing_crawl_issues` and `bing_fetched_urls` all came back empty for a site whose
`bing_crawl_stats` reported 1700 inbound links in the same minute. A read that returned
no rows carries `empty_response: {rows_returned: 0, measured: false, note: …}` beside
`result` over MCP, and prints the note on stderr at the CLI, so "Bing returned nothing"
cannot be reported as "no problems found". Reads that answer with a single record —
`bing_url_info`, `bing_crawl_settings`, the quotas, `bing_keyword` — are never labelled:
an array inside one record, such as `CrawlRate`, is a field of that record
and not a row Bing withheld. `bing_url_info` labels the same trap in one field:
`HttpStatus: 0` means Bing reports no status, not `200`, and the row says so with
`http_status_reported`.

## Coverage and references

- [Agent skill: how to drive this server](SKILL.md)
- [CLI commands, MCP tools, and write operations](docs/operations.md)
- [Complete 62-method Microsoft interface transcription](docs/api-surface.md)
- [Product boundaries](docs/product-boundaries.md)
- [Configuration](docs/configuration.md)

The Microsoft interface contains 62 methods: 59 supported here and three obsolete
deep-link methods deliberately excluded. Keyword research is standalone because its
official signatures contain no site parameter. Content submission is exposed because
it remains present in Microsoft's current interface; account-side eligibility can only
be confirmed by Bing for a particular account.

## Licence

MIT.
