AskSakina Islamic Knowledge MCP Server

Specialist-AI-reviewed Islamic knowledge for AI agents. Quranic verses, a curated du'a corpus, and the 99 Names of Allah — reviewed across the four mainstream Sunni schools (Hanafi, Maliki, Shafi'i, Hanbali) through AskSakina's structured specialist-AI review chain, which is not a substitute for a qualified scholar. Every response is wrapped in a presentation contract so agents cannot silently misrepresent the content.

Architectural posture: "AskSakina ships records, agents ship answers." The server is a reference library, not an advisor. Each response includes a _sakina_meta envelope with disclaimer, LLM directives, presentation contract, and educational context. What the calling agent does with the record is its responsibility — but every response arms the agent with enough structural context to make mishandling difficult and trackable.

What it provides

Four lookup tools, two learning-journey prompts and one canonical resource over the Model Context Protocol. Every record is sourced from AskSakina's main app (asksakina.com), so each MCP release matches what the public app surfaced when it was built.

Tools

get_quran_verse

Verbatim Quranic verse lookup by surah:ayah. Returns the canonical AskSakina envelope with Arabic text, translation, surah name (Arabic + English), and the quran presentation contract — no paraphrasing, surah:ayah citation required, "Translation of the Meaning" labelling.

Parameters

Example response (truncated)

{
  "_sakina_meta": {
    "version": "1.0",
    "content_type": "quran_verse",
    "presentation_contract": {
      "quran": {
        "no_paraphrase": true,
        "require_citation": "surah:ayah",
        "translation_label": "Translation of the Meaning"
      }
    }
    /* … */
  },
  "content": {
    "surah": { "number": 2, "name_arabic": "البقرة", "name_english": "Al-Baqarah" },
    "ayah": 186,
    "arabic": "وَإِذَا سَأَلَكَ عِبَادِي عَنِّي فَإِنِّي قَرِيبٌ ۖ …",
    "translation": "And when My servants ask you concerning Me — indeed I am near. …"
  }
}

get_dua

Returns du'as matching a life context (anxiety, grief, morning, travel, etc.). Context is resolved against canonical category slugs, an alias map, and tag dimensions. A natural phrase ("grief after losing my mother") also resolves: the server normalises it and matches known category, alias, tag and synonym words and a few multi-word phrases inside it, deterministically and without AI. content.matched_by says how it resolved. Grief (WO#386): a query that names the querier's own grief ("I lost my mother", "grief after losing my mother", "my father passed away", "bereaved") resolves to context bereaved; a funeral rite or the person who has died ("funeral", "dua for the deceased", "visiting my mother's grave") stays deceased. Parents (WO#430): a query that names a parent (mother, mum, mom, father, dad, parent, parents) together with a death word (late, dead, passed, grave, burial, janaza, janazah, cemetery, funeral, died, death, deceased, lost) takes this grief route and never returns mothers or fathers; "miss" with a parent and no death word returns only the du'as that suit a living or a deceased parent; a plain parent query ("my mother", "parents") is unchanged. A death word on its own, such as "grave", resolves to nothing. bereaved returns seven du'as ruled by the fiqh reviewer (Gem 4), in its order, with a teaching_note on the masculine forms in the Arabic, and always carries at least the soft support note. Crisis detection runs first, on the full raw input. Self-harm and abuse get separate blocks (crisis_type): the abuse block carries domestic-abuse lines and an abuse-specific directive. On any hard crisis (self-harm or abuse) no du'as are returned: content_type is crisis_resource_only and the crisis resource is the whole response. Each block lists 24/7 lines first and states the hours of any limited-hours line. If the context contains a crisis keyword (matched against the main app's detectCrisis keyword list), the response includes a mandatory crisis_resource block with the appropriate hotline and prosocial directive.

Parameters

Result shaping (2.2.0). A context lookup returns limit records of the route's existing order (bundle order; Gem 4's fixed order on bereaved), starting at offset. When more follow the page, content.total_available gives the whole count, content.truncated is true, and content.more carries the note "N more du'as for this context; call get_dua with offset and limit to see them." The same note is in CRITICAL_RULES. content.offset is present when it is not 0. Every record of a category is reachable by paging.

Example response (truncated)

{
  "_sakina_meta": {
    "content_type": "dua_collection",
    "presentation_contract": {
      "hadith": { "require_grading": "per_record" /* true when every record is graded */ }
    }
    /* … */
  },
  "content": {
    "context": "stress-anxiety",
    "duas": [
      {
        "dua_block": "Arabic: …\nTransliteration: …\nTranslation: …\nOrigin: Quran 21:87\nSource: Quran 21:87\nGrading: Quranic",
        "title": "Dua to Deal with Distress & Depression",
        "arabic": "وَذَا ٱلنُّونِ إِذ ذَّهَبَ …",
        "source": "Quran 21:87",
        "grading": "Quranic",
        "grading_status": "quranic",
        "origin": "quran",
        "quran_citation": "Quran 21:87"
      },
      {
        "title": "Dua to Remove Sorrow & Grief:",
        "source": "Source not recorded in AskSakina corpus",
        "grading": "Not graded in AskSakina corpus",
        "grading_status": "not_recorded"
        /* … */
      }
      /* … */
    ],
    "total_results": 3,
    "total_available": 12,
    "truncated": true,
    "more": "9 more du'as for this context; call get_dua with offset and limit to see them.",
    "grading_summary": { "quranic": 2, "graded": 0, "not_a_narration": 0, "not_recorded": 1 }
  }
  // crisis_resource block added if context triggers the crisis keyword list
}

Grading status. Every du'a record carries grading_status: quranic (the text is a Quranic verse or clause), graded (a hadith grading is recorded), not_a_narration (a general supplication, not a narration from the Prophet; it carries a grading_directive saying so), or not_recorded (the AskSakina corpus holds no grading for it). Most of the corpus is not_recorded today; the Source and Grading lines say so rather than implying a grading exists. Du'as whose Quranic reference has been verified carry origin: "quran" and a quran_citation, and their Arabic is the Tanzil Uthmani text resolved from AskSakina's scripture module at bundle time.

get_name_of_allah

Look up by number (1–99) or string (transliteration, Arabic, or English meaning). Returns Arabic, transliteration, locale-aware meaning, reflection, and Quranic references.

Parameters — one of:

Plus optional locale (tr returns the English meaning and reflection).

Example response (truncated)

{
  "_sakina_meta": { "content_type": "name_of_allah" /* … */ },
  "content": {
    "number": 29,
    "arabic": "الحَكَم",
    "transliteration": "Al-Hakam",
    "meaning": "The Judge",
    "reflection": "Allah is the absolute Judge…",
    "quran_references": ["6:114", "13:41"]
  }
}

find_verses

Topic lookup beside get_quran_verse's reference lookup (WO#386). The query is normalised and each word is matched to a theme of AskSakina's thematic verse index (anxiety, gratitude, patience, trust, grief, hope, forgiveness, morning, friday, ramadan, encouragement), by theme name or a fixed synonym list. Deterministic: no embedding, no AI call.

Each verse carries reference (surah:ayah), arabic (the full ayah from the scripture module, Tanzil Uthmani, byte for byte), translation (full-ayah Pickthall, fetched as get_quran_verse fetches it), the surah names, and a relevance_note saying which theme lists it. The note is not commentary. If the upstream is unreachable, the Arabic is still served and translation is null with a note.

No match returns content_type: "not_found" with verses: [], no_results: true and the list of available themes, not an error. Crisis detection runs first, as on get_dua: any hard crisis (self-harm or abuse) returns crisis_resource_only, with no verses and no theme list.

Parameters

Resource: sakina://about

Read this resource before using the tools. It contains the seven core directives plus three Gem-cleared educational dawah texts. Most MCP clients fetch it automatically the first time they connect.

Server instructions (2.2.0)

The initialize result carries instructions (1,529 characters, under the 1,801 limit): the quoting and teaching rules, the da'wah adab block ruled on 20 Oct 2026 (verbatim), and the open-door line with its conditions (only on the user's interest in Islam itself, at most once per conversation, never in distress, never as a condition of an answer, never after a decline). A client that surfaces instructions gives its model these rules once per session; the per-response rules in _sakina_meta.llm_directives apply either way.

Teaching (2.2.0, "teach, don't hand off")

Where a teaching entry exists for a record (src/teach/entries.ts; six records today), the record carries a teach block and a related list, with or without audience. The directives give the teaching order: quote the record first, exactly (Arabic, translation, source, grading); then explain only from the teach block, as context from AskSakina's notes, not tafsir or a ruling; where the notes end, say so and point to a recognised tafsir or a qualified scholar; offer the reflection question verbatim without answering it; offer at most one related next step. Nothing attaches under a support block, on a grief route or on a crisis response, and the directives tell the agent to quote records only once the user has disclosed distress, grief or crisis. Where no teach block is supplied, the agent may explain only from the record's own supplied fields, attributed to AskSakina's notes, then hand over the same way.

Prompts: learning journeys (2.4.0)

Two prompts, for clients that show MCP prompts (for example as slash commands). Each returns one user message that tells the agent to call one tool, then to present the response by the rules that response already carries (_sakina_meta.llm_directives.CRITICAL_RULES and presentation_contract, which hold the teaching order whenever a teach block is served) and by the server's da'wah adab, quoted verbatim. A prompt adds no religious content: everything the user sees comes from the tool's record and its Gem-approved teach entry, if one exists.

• Arguments are strings of digits, checked before the prompt runs; anything else is refused with -32602 and the value is not echoed. Whether an ayah exists in its surah is answered by the tool (not_found), as for a direct call.
• If a response carries a crisis_resource block (crisis or soft support), the message tells the agent to follow that block's directive and add nothing: no reflection question and no next step.
• Prompt arguments are never logged or counted. /stats counts only how many times each prompt was fetched, by name.
• Two further journeys, what_muslims_believe and first_steps_new_muslim, are parked: they need authored content that has not been reviewed.

Response envelope

Every tool returns the same shape:

{
  "_sakina_meta": {
    "version": "1.0",
    "source": "AskSakina Islamic Knowledge Server (asksakina.com)",
    "content_type": "quran_verse" | "dua_collection" | "name_of_allah" | "not_found" | "crisis_resource_only",
    "disclaimer": "AskSakina provides curated Islamic reference content for educational purposes. …",
    "llm_directives": { "CRITICAL_RULES": ["…"] },
    "presentation_contract": {
      "quran":         { "no_paraphrase": true, "require_citation": "surah:ayah" /* … */ },
      "hadith":        { "require_grading": true, "require_source": true /* … */ },
      "name_of_allah": { "no_paraphrase_arabic": true /* … */ }
    },
    "educational_context": "…"
  },
  "content": { /* tool-specific record; a record with a teaching entry also carries "teach" and "related" (2.2.0) */ },
  // get_dua and find_verses, when crisis keywords detected:
  "crisis_resource": { "level": "hard" | "soft", "crisis_type": "self-harm" | "abuse", "locale": "…", "directive": "…", "text": "…" }
}

The envelope is the contract: _sakina_meta.presentation_contract tells the agent how to render the record, disclaimer tells the agent what claims it cannot make on its own authority, and crisis_resource overrides everything when a user's distress crosses a safety threshold.

Quick start

Run via npx (no install)

npx -y @asksakina/islamic-knowledge-mcp --stdio

With --stdio the package speaks MCP over stdin/stdout; wire it into any MCP client. Without the flag it starts the Streamable HTTP server on $PORT (default 3030, POST /mcp, GET /health), as the hosted endpoint runs. Before 2.1.0 the sakina-mcp command exited without starting either (WO#444).

Connect via remote streamable HTTP

https://sakina-mcp.fly.dev/mcp

For agents that prefer a hosted endpoint over a local subprocess. Same tools, same envelope, same rate limits.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "sakina": {
      "command": "npx",
      "args": ["-y", "@asksakina/islamic-knowledge-mcp", "--stdio"]
    }
  }
}

Restart Claude Desktop. The four tools and the sakina://about resource appear in the tool picker.

Cursor

Add to ~/.cursor/mcp.json (or workspace .cursor/mcp.json):

{
  "mcpServers": {
    "sakina": {
      "command": "npx",
      "args": ["-y", "@asksakina/islamic-knowledge-mcp", "--stdio"]
    }
  }
}

VS Code (with an MCP-aware extension)

Add to .vscode/mcp.json in your workspace:

{
  "servers": {
    "sakina-islamic-knowledge": {
      "command": "npx",
      "args": ["-y", "@asksakina/islamic-knowledge-mcp", "--stdio"]
    }
  }
}

Generic MCP client over remote HTTP

{
  "mcpServers": {
    "sakina": {
      "type": "streamable-http",
      "url": "https://sakina-mcp.fly.dev/mcp"
    }
  }
}

Use cases

The MCP server is designed for any agent or application that needs curated Islamic source material, returned with explicit instructions to quote it exactly rather than paraphrase it.

• Build an Islamic chatbot with cited references where recorded. Pipe get_quran_verse and get_dua into your conversation flow; the presentation contract instructs the agent to quote exactly and keep gradings attached, and each response states its content type and, for du'as, each record's grading status.
• Add prayer-time-aware spiritual content to your agent. Combine the Quranic and du'a tools with a prayer-time API to surface Allah's words at the right moment of the day.
• Recommend du'as based on user context. The crisis-keyword filter in get_dua handles the dangerous edge cases (self-harm language, abuse) so your assistant is directed not to reply with a generic du'a to a safety-critical message.
• Answer questions about the 99 Names. get_name_of_allah returns the canonical Arabic, transliteration, meaning, and reflection — plus Quranic references — for both number-based and name-based lookups.
• Stay theologically inclusive. The server is designed not to present a contested ruling as the universal answer.

Content integrity

New and changed content is reviewed by AskSakina's "Gem" specialist-AI review chain before shipping. Four of its reviewers gate every Islamic-knowledge release:

1. Quran & Translation Verification. Ayah accuracy, surah/ayah citation correctness, hadith grading, source review.
2. Fiqh Diversity (Multi-Madhab). Cross-school accuracy, no single-madhab framing on contested topics, inclusive language.
3. Islamic Psychology & Pastoral Care. Comfort window safety, crisis filter coverage, no guilt-based motivation.
4. Explorer / New-to-Islam Accessibility. Glossing of Arabic terms for non-Muslim audiences, plain-English fiqh, zero-assumed-belief reflections.

Architectural behaviours implemented in code:

• The Unicode prophet salutation (U+FDFA, ﷺ) is replaced with (peace be upon him) at every data-loader boundary. The response builder rejects any output containing the symbol.
• Each du'a record carries a grading_status (quranic, graded, not_a_narration, not_recorded). The hadith presentation contract sets require_grading: true only when every returned record is graded or Quranic, and "per_record" otherwise, so the envelope never claims a grading the data does not hold.
• Crisis keyword detection runs on every get_dua request — the same keyword list used by the main app's safety gate.

Sacred Use License. This server is distributed under a Sacred Use License (see LICENSE). Permitted uses centre on dawah, education, personal worship, and respectful integration into Muslim-serving applications. The license forbids monetisation that frames Islamic knowledge as scarce or paywalled. A public summary of the license terms lives at https://www.asksakina.com/en/mcp.

Privacy. See Privacy & Analytics.

Running locally for development

cd mcp-server
npm install
npm run bundle-data     # snapshot data from main repo into data/
npm run typecheck       # tsc --noEmit
npm run start           # HTTP server on :3030 via tsx
npm run test:tools      # exercise each tool, prints sample JSON

# Or build + run from compiled output:
npm run build           # tsc -> dist/, postbuild adds shebang + dist/package.json
npm run start:dist      # node dist/server.js

POST /mcp accepts standard MCP JSON-RPC 2.0. GET /health returns { "status": "ok", "name", "version" }.

Rate limiting

Default: 60 requests per minute per IP.

• UPSTASH_REDIS_REST_URL + UPSTASH_REDIS_REST_TOKEN set → distributed Upstash Redis fixed window.
• Otherwise → in-memory token bucket, per process.

Responses include X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset. A 429 with code -32029 is returned on exceed.

Authentication

v1 ships unauthenticated. An optional X-Sakina-App-Id header is accepted but not enforced or logged in v1. This becomes required in v2.

Privacy & Analytics

This service operates with an aggregate-only data model (WO#360; daily totals WO#456; teaching counters WO#477). For each tool, the server maintains a running count of calls by outcome (ok, not_found, error), the summed handler time to report operational averages, and the number of responses that carried a teaching block. Across all tools it also counts how many calls carried each value of the optional audience argument (curious, revert, practising, unknown, or none), and how many times each learning prompt was fetched. These are totals only, reported via the /stats endpoint.

• No persistent tracking: No per-request logs or client identifiers are written to disk or any database by the application.
• Ephemeral IP handling: Client IP addresses are used only to enforce the rate limit. Each is held as a rate-limit key in Upstash Redis for the rate-limit window (60 seconds), then expired. If Upstash is unavailable, the in-memory fallback holds it in process memory for the same window, then discards it. Raw IP addresses are never persisted by the application, and never written to logs.
• No payload logging: Request parameters (such as context values, verse references, Name numbers, prompt arguments or locale) and response payloads are never captured or logged. The one argument that is counted is audience, and only as a tally of its four fixed values plus "none"; the value itself is never stored with the call, the tool, or anything else.
• Since-start counters: The totals /stats reports under by_tool exist only in RAM and are reset when the server process restarts.
• Daily totals on the server's own disk: On the hosted server, the number of calls to each tool on each UTC day, by outcome (ok, not_found, error), their summed handler time, the number that carried a teaching block, the tally of audience values, and the number of fetches of each learning prompt are written to the server's own disk so that they survive restarts. They are kept for at most 90 days; older days are deleted. They contain no IP addresses, user agents, client identifiers, times of individual calls, parameters or payloads, and cannot be linked to any request or person. They are written only where MCP_USAGE_DIR is set: a local npx or --stdio run writes nothing.
• Nothing per person: No counter is kept per client, session, user or conversation. The teaching, audience and prompt counters are the same kind of number as the per-tool call counts: one running total across every caller.

Regulatory Note (UK GDPR): Because nothing identifiable is held beyond the rate-limit window, there are no records that could be accessed or erased. Note that while the application itself does not persist personal data, our infrastructure provider (Fly.io) transiently processes IP addresses at the edge layer for standard network routing and platform security.

/stats endpoint

Authenticated summary endpoint:

curl -H "Authorization: Bearer $MCP_STATS_TOKEN" https://sakina-mcp.fly.dev/stats

Requires the MCP_STATS_TOKEN env var to be set (otherwise the endpoint returns 503 and refuses to serve unauthenticated). Token comparison is constant-time.

Payload shape:

{
  "privacy": "aggregate-only counters; no per-request records, no IPs, no user-agents, no client identifiers, no geolocation, no params beyond a tally of audience values, no payloads; since-start totals reset on restart; daily totals (per tool: ok, not_found, error, summed duration, teach served; plus audience tally and prompt fetches by name) kept on the server's own disk for at most 90 days",
  "since": "2026-09-25T…",
  "total_requests": N,
  "by_tool": {
    "get_quran_verse":   { "total": N, "ok": N, "not_found": N, "error": N, "avg_response_ms": N, "teach_served": N },
    "get_dua":           { "total": N, "ok": N, "not_found": N, "error": N, "avg_response_ms": N, "teach_served": N },
    "get_name_of_allah": { "total": N, "ok": N, "not_found": N, "error": N, "avg_response_ms": N, "teach_served": N },
    "find_verses":       { "total": N, "ok": N, "not_found": N, "error": N, "avg_response_ms": N, "teach_served": N }
  },
  "audience": { "curious": N, "revert": N, "practising": N, "unknown": N, "absent": N },
  "prompts": { "understand_a_verse": N, "explore_names_of_allah": N },
  "rate_limiter": {
    "state": "upstash",
    "upstash_configured": true,
    "upstash_failures": 0,
    "last_failure": null,
    "last_failure_at": null,
    "last_success_at": "2026-09-25T…",
    "startup_check": "Upstash reachable (PING ok, 12 ms); state: upstash"
  },
  "basmala_prefix_mismatch": 0,
  "daily": {
    "retention_days": 90,
    "days": [
      { "date": "2026-10-12", "total": N,
        "by_tool": { "get_dua": { "total": N, "ok": N, "not_found": N, "error": N, "avg_response_ms": N, "teach_served": N }, … },
        "audience": { "curious": N, "revert": N, "practising": N, "unknown": N, "absent": N },
        "prompts": { "understand_a_verse": N, "explore_names_of_allah": N } }
    ]
  }
}

daily lists the stored days, newest first, with today's not-yet-written calls included; it is null when MCP_USAGE_DIR is unset, and the privacy line then names the in-memory counters only ("aggregate-only in-memory counters, including teach-served per tool, a tally of audience values and prompt fetches by name; … resets on restart").

teach_served counts responses that carried a teach block (WO#488). audience tallies the optional audience argument of every tool call by its four fixed values, with absent for a call that sent none; it is one total across all tools, never joined to a tool, a call or a time finer than the UTC day. prompts counts fetches of each learning prompt by name; prompt arguments are never read.

rate_limiter.state is upstash (Upstash answering), degraded (Upstash configured but failing: requests are limited by the in-memory fallback, never allowed through) or memory (Upstash not configured). Failure reasons carry the HTTP status or error, never a client IP. basmala_prefix_mismatch counts ayah-1 texts from the upstream that started with neither recorded basmala form and were served unchanged. It is an aggregate count only; the log line for it carries no verse reference.

One-time setup on Fly

fly secrets set MCP_STATS_TOKEN=$(openssl rand -hex 32)
fly deploy

The server writes one thing to the /data volume that fly.toml mounts: the daily totals, one file per UTC day in /data/usage (MCP_USAGE_DIR), written every 5 minutes and when Fly stops the machine. LOG_DIR (/data/logs) has been unused since WO#360.

Data sources

The du'a and Names data are snapshots of the main app's data files, so any update to them lands here on the next npm run bundle-data (the deploy workflow runs it). A published package carries the snapshot taken when it was built.

Out of scope for v1

Per the Phase 3 planning doc:

• Semantic search / search_islamic_guidance — needs a Gem-reviewed eval set. (find_verses is an index lookup, not semantic search.)
• explain_islamic_concept — directly conflicts with CLAUDE.md's "AI never gives spiritual/fiqh advice" rule until an Architect-level ruling is made.
• get_pastoral_guidance — Gem 3's framing depends on AskSakina-controlled surface; cannot ship via MCP without a separate review.
• check_halal_ingredient — blocked on WO#61 restoration.
• HMAC signing enforcement — v1 accepts but ignores the X-Sakina-App-Id header.

Deployment

Self-contained at publish time — npm run bundle-data snapshots the du'a corpus, the 99 Names, and the canonical safety modules from the main AskSakina monorepo into data/ and src/safety/_synced/. The compiled dist/ plus the bundled data/ directory have no runtime dependency on the main app.

bundle-data runs automatically via a prebuild script (WO#138), so npm run build is self-contained on a fresh checkout — no manual bundling step required. The src/safety/_synced/ directory IS committed to git so any drift from the main-app safety modules shows up in git diff between commits.

docker build does NOT run prebuild (it only copies the mcp-server/ directory, not the monorepo). Run npm run bundle-data on the host before docker build so data/ is populated. The same applies if you build from a fresh clone — npm install && npm run bundle-data before docker build.

CI deploy (canonical — WO#249)

All MCP releases go through the MCP Server Deploy GitHub Actions workflow (.github/workflows/mcp-deploy.yml) — no developer machine holds flyctl or npm auth. The Architect triggers it from the repo Actions tab (workflow_dispatch) with two boolean inputs:

• deploy_fly (default true) — build + flyctl deploy to sakina-mcp.fly.dev.
• sync_mirror (default true) — push the built package to the public mirror (github.com/yserrag/asksakina-mcp) and tag v<version> there. The mirror's own workflow then publishes @asksakina/islamic-knowledge-mcp to npm with trusted publishing and provenance (WO#444); this repository holds no npm token.

The job runs the full Gem 10 runbook on main: npm ci → npm run bundle-data (load-bearing) → flyctl deploy → poll /health → live Pickthall byte-check on 2:255 and 9:129 → live Dhun-Nun check → rate-limiter probe (warn only) → live publish guards → agent evals (warn only; after the guards since WO#386, so the blocking guards get the rate-limit budget first) → mirror sync and tag (the mirror publishes). The byte-check runs the MCP initialize handshake + tools/call get_quran_verse against the live endpoint: "save Him" (Pickthall) passes; "except Him" (Saheeh International) or an unreachable /mcp fails the job before the mirror sync, dumps flyctl status + logs, and prints the manual rollback commands. 9:129 must read "There is no God save Him", which only the bundled corrected text carries. Secrets FLY_API_TOKEN and ASKSAKINA_MCP_DEPLOY_KEY live in repo settings and are never echoed.

Automated rollback-to-previous-image is the Gem 10 end-state but is a deferred follow-up: the current live image is itself broken, so there is no known-good image to auto-restore yet. The first runs are fail-loud + manual rollback only.

Fly.io (manual / break-glass)

Prefer the CI workflow above. These commands are the underlying mechanics, for a machine that already has flyctl auth:

cd mcp-server
fly launch              # first time only — picks up fly.toml
fly secrets set \
  UPSTASH_REDIS_REST_URL=... \
  UPSTASH_REDIS_REST_TOKEN=...
npm run bundle-data     # load-bearing: docker build copies host data/
fly deploy
fly status
curl https://<app>.fly.dev/health

Production app: sakina-mcp.fly.dev. App sakina-mcp, region lhr, internal port 3030, force HTTPS, healthcheck on /health every 30 s, auto-stop on idle, 1 shared CPU, 256 MB RAM.

Post-deploy verification and rollback

After every fly deploy, byte-check one Quran verse from the live endpoint against the locked Pickthall table (docs/wo-cert/si-pickthall-LOCKED-final.json) to confirm the translation edition is Pickthall, not Saheeh International:

curl -s https://sakina-mcp.fly.dev/mcp ... # call get_quran_verse for a known ref
# compare the returned translation to the locked Pickthall text for that ref

If the byte-check fails (SI text served, or the translation does not match Pickthall), roll back first, diagnose second. Immediately redeploy the previous image before any further investigation or fix:

fly releases --app sakina-mcp                 # find the last-good release / image
fly deploy --app sakina-mcp --image <previous-image-ref>

Do not attempt a forward fix on a live SI regression — restore the known-good image, confirm the byte-check passes on it, then diagnose the bad build offline.

Local container test

docker compose up --build
curl http://localhost:3030/health

Other targets

The package is just a Node.js HTTP server, so it runs on any platform that takes a Dockerfile or a Node process: Railway, Render, a small VPS, Cloudflare Workers (with the streamable-http transport), Vercel Functions (with a small adapter). The in-memory rate limiter falls back gracefully when Upstash isn't configured.

npm publishing

Publishing goes through the MCP Server Deploy workflow (see CI deploy) with publish_npm: true — the Architect triggers it from the Actions tab; NPM_TOKEN lives in repo settings. The manual equivalent (break-glass, needs local npm auth):

cd mcp-server
npm login                                    # to the @asksakina org
npm publish --access public

prepublishOnly runs bundle-data → build → test:tools so the published tarball always contains fresh data and a clean build. The files field whitelists dist/, data/, README.md, LICENSE, and .mcp/server.json.

Verify the published tarball with npm pack --dry-run first to see what would ship.

MCP Registry submission

mcp-server/.mcp/server.json declares AskSakina's identity for the official MCP Registry:

• Server name: com.asksakina/islamic-knowledge
• Package: @asksakina/islamic-knowledge-mcp (npm, stdio transport)
• Remote: https://sakina-mcp.fly.dev/mcp (Streamable HTTP)

Publishing flow: the MCP Registry Publish GitHub Actions workflow (.github/workflows/mcp-registry-publish.yml in the main repository, manual dispatch). It checks that server.json, package.json and npm agree, then runs:

mcp-publisher login dns --domain asksakina.com --private-key <KEY>   # or: login http
mcp-publisher publish .mcp/server.json
curl "https://registry.modelcontextprotocol.io/v0/servers?search=com.asksakina"

Domain proof for com.asksakina is an Ed25519 key: the public half in a DNS TXT record on asksakina.com (v=MCPv1; k=ed25519; p=<base64>, since May 2026) or at https://asksakina.com/.well-known/mcp-registry-auth; the private half in the repository secret MCP_REGISTRY_PRIVATE_KEY.

Smithery re-publish (after a version bump)

Three places carry the version — all bumped together:

1. package.json ("version")
2. mcp-server/.mcp/server.json (top-level version and packages[0].version)
3. mcp-server/src/server.ts (SERVER_VERSION constant — surfaces in /health and /.well-known/mcp/server-card.json)

Re-publish checklist when bumping:

cd mcp-server
npm run typecheck && npm run test:tools && npm run test:analytics
npm run bundle-data
npm publish                                       # npm registry
mcp-publisher publish .mcp/server.json            # MCP Registry
fly deploy                                        # Smithery scanner picks up the new card via /.well-known/mcp/server-card.json

Smithery auto-scans /.well-known/mcp/server-card.json and re-indexes within a few minutes — no manual Smithery API call needed.

Links

• Web: asksakina.com
• MCP showcase: asksakina.com/en/mcp
• MCP registry: registry.modelcontextprotocol.io — com.asksakina/islamic-knowledge
• npm: npmjs.com/@asksakina/islamic-knowledge-mcp
• Source: github.com/yserrag/asksakina-mcp (mirror of mcp-server/ in the main repository)
• Issues / questions: github.com/yserrag/asksakina-mcp/issues