bisibility

@bisibility/mcp

Part of bisibility - an open-source SEO platform for keyword research, backlink analysis, and Google rank tracking. This repository contains the MCP server that exposes bisibility tools to AI agents.

Docs · API reference · Roadmap

Status: Developer preview.

Distribution

[npm version] [Official MCP Registry] [Glama] [smithery badge]

Distribution status and release channels

Model Context Protocol server for the bisibility REST API. It exposes stdio tools backed by the published @bisibility/sdk package.

Requirements

• Node.js 18 or newer
• A bisibility API key
• A bisibility API v1 base URL

Install

Run the published stdio server without a global install:

npx -y @bisibility/mcp

For local development, install and build from this package directory:

From this package directory:

npm installnpm run build

The SDK is consumed from the npm registry:

"@bisibility/sdk": "^0.22.0"

MCPB bundle

Directories and clients that support MCPB can distribute bisibility as a one-click local bundle. The bundle includes the stdio server and all production dependencies, but never a bisibility credential. The client collects configuration during installation.

Build the upload artifact with:

npm cinpm run buildnpm run build:mcpb

The versioned file is written to artifacts/.

Environment

export BISIBILITY_API_KEY="bsb_pat_live_..."export BISIBILITY_BASE_URL="https://eu.bisibility.com/api/v1"export BISIBILITY_PROJECT_ID="prj_..."export BISIBILITY_MCP_READ_ONLY="1"export BISIBILITY_MCP_TOOLSETS="projects,keywords,checks,rank-history"

BISIBILITY_BASE_URL is optional and defaults to https://eu.bisibility.com/api/v1. For self-hosted installs, set it to your API v1 root, for example https://rank.example.com/api/v1. BISIBILITY_API_KEY accepts a project key (bsb_key_live_...) or personal access token (bsb_pat_live_...). Set optional BISIBILITY_PROJECT_ID as the default X-Bisibility-Project selector for project-implicit PAT tools; a tool's optional project_id argument overrides it for that call. Every resource ID accepted by the MCP server, including BISIBILITY_PROJECT_ID, must use the current typed public ID format: a canonical lowercase prefix plus _ and a 24-character lowercase suffix, for example prj_a1b2c3d4e5f6g7h8j9k0m2n3. Malformed IDs, mixed-case IDs, and wrong resource prefixes are rejected. Location selection uses the returned location_key, never a location ID. A key can end in @language to select a non-default language for the location, for example ES/Andalusia/Malaga@en; an unqualified key uses the country's default language.

The server consumes the public ID contract from the published @bisibility/sdk package. Do not replace it with a local SDK link.

BISIBILITY_MCP_READ_ONLY accepts 1, true, yes, or on, ignoring case. When enabled, write tools are not registered and do not appear in tools/list.

BISIBILITY_MCP_TOOLSETS is an optional comma-separated allowlist. Valid toolsets are account, agent-reports, alerts, analytics, backlinks, brand-visibility, checks, competitors, domain-overview, keywords, notifications, page-indexing, project-context, projects, prompt-explorer, prompt-research, prompt-tracker, providers, rank-history, saved-keywords, saved-views, signals, site-audits, sitemaps, system, team, tokens, and webhooks. An unknown value prevents the server from starting. When the variable is unset, all toolsets are registered. The toolset filter and read-only mode compose. The allowlist is a scope control, not a way to improve tool selection.

Run

npm run buildBISIBILITY_API_KEY="bsb_key_live_..." node dist/stdio.js

When installed as a package, the bin is:

bisibility-mcp

Connect

After building the repository, connect Codex to the local stdio server:

codex mcp add bisibility \  --env BISIBILITY_API_KEY="bsb_key_live_..." \  --env BISIBILITY_BASE_URL="https://bisibility.com/api/v1" \  -- node /absolute/path/to/bisibility-mcp/dist/stdio.js

The codex mcp add syntax above was verified against the Codex CLI. Replace the absolute path and credential before running it.

Other MCP clients commonly use a JSON configuration like this:

{  "mcpServers": {    "bisibility": {      "command": "npx",      "args": ["-y", "@bisibility/mcp"],      "env": {        "BISIBILITY_API_KEY": "bsb_key_live_...",        "BISIBILITY_BASE_URL": "https://eu.bisibility.com/api/v1"      }    }  }}

If bisibility-mcp is installed on the client's PATH, use the package bin:

{  "mcpServers": {    "bisibility": {      "command": "bisibility-mcp",      "env": {        "BISIBILITY_API_KEY": "bsb_key_live_...",        "BISIBILITY_BASE_URL": "https://eu.bisibility.com/api/v1"      }    }  }}

Tools

Tools use unprefixed snake_case names and can be filtered with BISIBILITY_MCP_TOOLSETS. Tools shared with the built-in /api/mcp endpoint use the same names. The two transports have different tool inventories; check tools/list before switching.

Domain Overview tools always require a nonnegative whole-cent max_cost_cents cap. Use estimate_only: true together with max_cost_cents: 0 for the safe estimate-first step; provider estimates and returned charges may still contain fractional cents.

For rank results, read execution status and observation_completeness together with position. A null position with unknown or legacy coverage does not establish absence within Top N. Only complete coverage supports that conclusion; a positive position remains valid with truncated coverage. latest_successful_check means the most recent completed observation, which can itself have unknown coverage.

Backlinks history is optional. history_unavailable marks a history failure with confirmed cost while summary and links remain available. If accounting is unknown, the request stays failed (isError) and may carry the completed summary as current-response evidence: cost_cents is null and rows are not requested. That evidence is not a cached snapshot or a settled total, cannot be extended with load_more_backlink_rows, and does not survive reload.

All protected tools use the configured BISIBILITY_API_KEY. Tools that accept an idempotency_key forward it as the API Idempotency-Key request option. The key is required for create_prompt_tracker_run, retry_prompt_tracker_run, and prompt_research_suggestions_generate; generation requires a UUID. Other tools expose optional keys where supported by their schemas.

The list above is asserted by a test (test/tools.test.ts), so it stays in sync with the registered tool surface.

Security

The credential's scope is the primary authorization control, and the server accepts two kinds. A project key (bsb_key_live_...) belongs to exactly one project, which bounds the damage from a leak and makes it a good fit for a single-project or machine setup. A personal access token (bsb_pat_live_...) covers the projects you are a member of, so one token serves them all; its effective access in each project is the lower of the token's scope and your role there, meaning a token never grants more than the person behind it. Use BISIBILITY_PROJECT_ID to set the default project for a token that spans several, as described under Environment above.

Whichever kind you use, create it with a read scope for assistant use whenever possible, and do not grant admin unless the assistant needs an administrative API operation. Server-side filtering narrows the tools presented to the model, but it does not expand or replace the permissions of the configured credential.

An agent that receives a write-scoped or admin credential can create and change project data. Provider-backed Domain Overview tools can also spend the project's own provider budget on a cache miss. Every Domain Overview call requires an explicit nonnegative integer max_cost_cents. For the safe estimate-first step, call analyze_domain_overview with estimate_only: true and max_cost_cents: 0; use the same zero cap for a cache-only attempt. A positive value caps the permitted provider charge.

Brand Visibility reads a limited provider-observed dataset. Prompt Explorer runs synthetic prompts; it does not measure all user conversations or observed Brand Visibility. Observed visibility requires an explicit whole-cent max_cost_cents cap. Prompt comparison uses that cap in default hard-cap mode; explicit provider actual-cost mode uses acknowledged non-guaranteed estimates and an advisory limit. Start with estimate_only: true and a zero cap. Site audits crawl only the project domain within a bounded page count and deadline; they do not run Lighthouse or paid provider work. Saved reports and project context retain the project's existing authorization rules.

The destructive surface includes delete_project, delete_keyword, bulk_update_keywords when its operation is delete, delete_webhook, delete_alert_rule, delete_saved_keyword, delete_saved_view, remove_team_member, remove_competitor, disconnect_provider, revoke_api_key, revoke_personal_token, revoke_team_invite, and revoke_migration_token. Revoking the credential used by the server can immediately lock the server out.

Use read-only mode and a narrow toolset allowlist as defense in depth:

export BISIBILITY_MCP_READ_ONLY="1"export BISIBILITY_MCP_TOOLSETS="projects,keywords,checks,rank-history,alerts"

Content returned by the API is untrusted input that reaches the model. Project names, keyword phrases, alert text, and similar tracked data can contain hostile instructions intended to steer an agent that also has write tools. Keep credentials and registered toolsets as narrow as the workflow permits, and enable client-side confirmation for tool calls when the MCP client supports it.

HTTP Transport

This package ships stdio transport only. The code is split so an HTTP transport can reuse the same server factory:

import { createBisibilityClientFromEnv, createBisibilityMcpServer } from "@bisibility/mcp";
const client = createBisibilityClientFromEnv();const server = createBisibilityMcpServer({ client });

An HTTP entry point can wrap that server with the official SDK HTTP transport, such as StreamableHTTPServerTransport, in a Node, Worker, or hosted adapter.

Development

npm run lintnpm run typechecknpm testnpm run build

License

Licensed under the Apache License, Version 2.0. See LICENSE and NOTICE.

Prompt Tracker

The prompt-tracker toolset exposes 22 project-scoped tools for topics, immutable prompt revisions, disabled schedules, spend-free preview, consented runs, samples, history, trends and exports. Read-only sessions retain preview and evidence reads. Launch and retry require reviewed credential/budget/consent revisions, explicit paid consent, and a stable idempotency_key, forwarded through the SDK as Idempotency-Key. No operation starts an automatic paid baseline.

get_ai_research_catalog reads free model and locale capabilities. compare_ai_prompts accepts catalog model IDs, optional web search and locale hints, and an explicit actual-cost policy. Actual-cost consent acknowledges that its estimate is not a guaranteed maximum.

Prompt Research

The prompt-research toolset exposes prompt suggestions. suggest_prompt_research_prompts reads the reviewed project context and returns heuristic templates with no measured popularity. accept_prompt_research_suggestions saves explicitly accepted drafts and starts no provider work.

Model prompt generation uses prompt_research_suggestions_preview followed by prompt_research_suggestions_generate. Preview reviews all five context fields, selected competitors, model and market without paid provider work. Generation requires the complete unexpired frozen preview, explicit paid consent and a stable UUID idempotency_key. Forecasts have no guaranteed maximum; generated drafts carry no measured popularity. Accepted edits retain the trusted generation_reference; provider datasets use a verified provider_dataset_reference.

Response evidence compatibility

Rank checks expose optional observation_completeness, and keywords expose optional latest_check and latest_successful_check. Older responses can omit these fields. Only complete coverage supports absence within the requested depth; a positive position remains valid with unknown or truncated coverage. A newer failure retains the prior success.

Site backlinks normally have 12 history months. A successful partial snapshot has history: [] and history_unavailable: true; preserve its summary, rows and confirmed cost. Page scope does not request history. Missing availability metadata uses the legacy default. A failed summary remains an HTTP error with current-response evidence in problem details: cost_cents: null is an unknown total, while known_summary_cost_cents is confirmed.

Versioned retained history export (Unreleased)

Opt in with version 2 to retain report quality, metric versions, frozen selection and signed continuation cursors. Keep stream, basis, from/to and keyword selection unchanged on every page. The default remains legacy v1 rows. Observations include bounded normalized items; opaque CID strings retain all original digits. The recorded-at attempts stream also contains actual run/item variants without checks, including cancelled, deferred and blocked outcomes. Full-project exports retain explicit project-only runs; selected exports report unresolved membership exclusions. CSV contains a complete bounded report manifest and accepts no cursor. Oversized sources fail with 413, and changed signed snapshots fail with 409.

See the rank-history contract.

The served MCP tool returns JSON. Complete CSV downloads use the REST API or SDK.

Retained Local reads

The five read-only list_local_monitor_runs, get_local_monitor_run_results, list_local_monitor_point_listings, list_local_monitor_point_history and export_local_monitor_run tools call the SDK's actual Local API methods. They preserve original lmn/lmr/lmt scope, null/unknown quality, exact string CID and retained source metadata. History as_of/from/to use canonical UTC milliseconds and occurrence_created_at; preserve the full selection with a signed cursor (1..8000 characters). The export tool returns existing Local point-report-v1 JSON only. Reads never start provider work or schedules. These source changes are unreleased; packaged support follows the SDK release.

Reviewed defaults

Use preview_project_defaults to read the current draft and impact; it performs no settings writes or provider work, but the server still requires settings permission. After the caller explicitly approves that draft, call update_project_defaults with the same settings and review containing the returned expected_source_hash, string expected_revision, preview_token, a stable idempotency_key, and confirmed: true. Missing confirmation returns 428; stale or expired reviews return 409. For a lost response, reuse the exact body and command key. The tools never confirm automatically.

The SDK requires the server capability features.reviewed_project_defaults: "v1" before sending either preview or reviewed-update PATCH. It refuses older servers rather than risking a write disguised as preview. This marker does not grant settings permission.

Search performance totals

MCP v0.20.0 adds the read-only get_search_performance_totals tool in the analytics toolset. It reads complete web search totals across all queries, pages, countries and devices for inclusive Pacific start_date and end_date values (at most 366 days). It requires project_id and an existing Search Console connection. Set the string compare_previous to "true" to include the preceding equal-length period.

The response preserves clicks, impressions, CTR, impression-weighted position, actual data dates, source and freshness for each period. Complete finalized aggregate coverage serves stored data; otherwise one bounded live date series supplies both periods. Null finalized_through means unknown finality. Empty covered periods return zero metrics and null data dates. Reads never shift the requested dates or start an import. Hosted API support is required.

Page indexing

See the Page indexing tool reference.

The page-indexing toolset reads Search Console sitemaps and explicitly inspects URLs. Results are Google's indexed state, not a live test, and may be partial or delayed. list_search_index_sitemaps reads the connected property's sitemaps; get_search_index_sitemap accepts the full sitemap URL in path. inspect_search_index_urls requires write scope and accepts 1-10 URLs, plus an optional idempotency_key (1-200 characters) for safe retries. Inspect each per-URL result for errors, even on success. Manual and background inspections share the project's daily allowance; inspect per-URL errors for rate-limit reset evidence. Read-only mode hides URL inspection.

get_page_indexing_report, list_page_indexing_urls and get_page_indexing_settings read stored results and settings without calling Google. The URL list supports cursor paging, group, reason, sitemap, source and substring filters, plus changed_since (an RFC 3339 date-time) and optional history via history="true". changed_since includes URLs with a stored tracked change at or after the timestamp, before pagination. Check completeness and freshness before interpreting totals. update_page_indexing_settings changes the background budget mode or fixed daily limit and accepts an optional idempotency_key (1-200 characters). Manual and background inspections share an estimated property counter with a Pacific midnight reset. Available modes are returned in settings; all modes, including available_quota, are available on Cloud and self-host.