mtg-mcp

mtg-mcp

mtg-mcp gives an LLM grounded Magic: The Gathering card, deck, provider, and statistical evidence. The LLM makes deckbuilding decisions.

Version 0.9.0 is a fresh stdio server design. Its full all profile in remote mode exposes 97 tools, one mtg://server/capabilities resource, and no prompts. The focused default profile exposes 33 tools in read-only mode and 55 in local or remote mode. It does not migrate 0.8.x data or tool schemas.

Status

The evidence-first rewrite is complete. 0.9.0 is the first stable release of the new server.

The packaged acceptance run before the Spellbook addition made 88 safe tool calls. It uses saved test data for the two Scryfall card-data download tools and the two Playgroup write tools because a live cleanup route is unavailable. Scryfall rollback needs Scryfall to publish another data version; until then, the check reports that it is waiting. Commander Spellbook has its own limited, read-only live check. See live acceptance.

Start the server

Bootstrap the repository, then run the MCP smoke test:

.\bootstrap.ps1
task smoke:mcp

On Linux or macOS:

./bootstrap.sh
task smoke:mcp

Bootstrap installs Mise when needed. Mise installs Task and PowerShell from mise.toml and the .NET SDK from global.json in an isolated tool directory. After bootstrap, run task <command> normally. The Taskfile invokes Mise for .NET; if a fresh shell cannot find Task, activate Mise or temporarily use mise exec -- task <command>.

Committed Mise and NuGet lock files keep setup repeatable. Version pins live in mise.toml, global.json, Directory.Packages.props, and dotnet-tools.json. Use task deps:check to inspect available updates. Use task deps:update only when you intend to review the related pin and lock-file changes.

Use mtg-mcp as the MCP command. The default invocation is equivalent to:

mtg-mcp --mode=local --toolsets=default

Use --smoke for a one-shot configuration check. It does not start an MCP session.

task install also starts and stops an isolated, read-only server with the Scryfall toolset enabled. That catches packaged SQLite and Scryfall startup problems that --smoke cannot see.

Choose a mode

Modes control authority.

local is the default. A Scryfall cache miss can need a local write to record the response and rate-limit timing. In read-only, that miss returns local-write-required before any HTTP request.

Commander Spellbook reads are available in all three modes. They never change a deck or make a write request to the source. Successful responses use a short-lived local cache so repeated exact requests do not create needless provider traffic.

Choose toolsets

Toolsets control which tools the session can see. They do not grant permission to write.

Available toolsets are decks, scryfall, stats, archidekt, playgroup, and spellbook. Selection is fixed for the MCP session.

Read mtg://server/capabilities to see the active mode, toolsets, tool counts, credential state, package version, and local data state. It does not contact a provider.

Configure the server

Pass configuration on the command line, in environment variables, or in an optional mtg-mcp.json in the working directory. Command-line values win over environment values. Environment values win over JSON.

The default data root is the platform application-data directory under mtg-mcp/v0.9. All MCP processes that use the same root reuse decks.db and scryfall.db. Reads do not create the root.

Configure Archidekt

Use MTGMCP__ARCHIDEKT__USERNAME and MTGMCP__ARCHIDEKT__PASSWORD, or create:

~/.mtg-mcp/archidekt.json

The provider origin is fixed. Authentication output never contains credentials, account identity, or a credential path.

Configure Playgroup

Use MTGMCP__PLAYGROUP__API_KEY, or create:

~/.mtg-mcp/playgroup.json

The file contains one apiKey property. The provider origin is fixed to the official public API.

Work with local decks

The decks toolset supports:

• revisioned deck, entry, zone, category, and provider-binding changes;
• atomic ordered batches;
• guarded backup, restore, and delete;
• native JSON and generic text import and export;
• manual import and export files for Archidekt and Moxfield;
• exact Scryfall identity reconciliation;
• tag-rule grouping with repeatable results; and
• on-curve estimates based on card data and caller-supplied land rules.

Every change to an existing deck requires expectedRevision. The local store is not tied to one import format. Validation checks structure, not Commander legality, card quality, or strategic fit.

Use deck_identity_reconcile_preview before apply. Resolution uses printing ID, set and collector number, Oracle ID, then exact name. It never uses fuzzy matching or selects an arbitrary printing.

Use deck_category_rules_validate, deck_category_rules_preview, and deck_category_rules_apply to evaluate caller-owned tag rules. Rules can be inline or use the transparent common-v1 preset. The MCP does not decide what a category means.

Estimate mana on curve

Use deck_on_curve_estimate to sample how often one mainboard card can be cast by a chosen turn. Supply the saved deck ID and current revision, the target entry ID, one rule for each eligible land, a turn limit, and optionally a seed and sample count. The result includes the exact installed Scryfall facts used, the caller rules, a replay seed, a 95% Wilson interval, miss counts, and up to two sample traces.

Version 1 models only simple printed costs and caller-described single-faced lands. It does not read card text, infer land behavior, model other mana sources, choose plays, rate the deck, or predict real games.

Import and export decks

Use deck_interchange_formats to inspect preservation limits.

These provider export files do not contact a provider. Excluded entries remain in the native companion file and are omitted from provider text.

Query Scryfall

The scryfall toolset supports search, exact card lookup, collection lookup, prints, rulings, sets, catalogs, autocomplete, bulk metadata, tag evidence, immutable snapshots, and explicit card-data download operations.

The local card-data store contains All Cards, Rulings, Oracle Tags, and Art Tags. Card-data downloads are explicit. They never run at startup or in the background.

Optional: install full Scryfall card data

A full install contains card records, rulings, Oracle tags, and art tags. It does not contain card images. It is useful for local tag and card-data workflows; ordinary card lookups can use the provider and small local cache without it.

The download is multi-gigabyte and an update needs temporary working space. Start it only through the explicit full-data tools in local or remote mode. The public names scryfall_corpus_status, scryfall_corpus_sync, scryfall_corpus_rollback, and scryfall_corpus_delete are existing API names for inspecting and managing that installed card data.

Freshness policies are default, cache-only, and refresh. The default TTL is 24 hours. Immutable snapshots do not expire.

scryfall_card_collection accepts 150 ordered lookup rows. It uses local card-data hits first, deduplicates provider misses, and sends provider batches of at most

1. Results use stable cursor pagination.

For arbitrary queries, Scryfall remains the source of truth. The cache reuses only the exact same request. Card facts and community tags stay separate kinds of information.

Read Commander Spellbook evidence

Enable the spellbook toolset for three tools that return Commander Spellbook data:

• spellbook_variant_search returns one bounded page for an exact Commander Spellbook query.
• spellbook_variant_get returns one exact source variant by ID.
• spellbook_deck_combos_find sends only the commander and main entries from one exact saved deck revision, then reports the entries it skipped.

The tools preserve Commander Spellbook JSON and name the source, request, retrieval time, cache state, checksum, and limits. They do not rank results, infer card roles, or recommend a combo or card change. The cache lasts 15 minutes by default; configure it with a whole number from 1 through 1,440 minutes.

Calculate exact statistics

The stats toolset provides eight read-only tools for:

• univariate and multivariate hypergeometric probability;
• probability-by-turn tables;
• caller-defined mana-source availability;
• package and combo assembly;
• explicit mulligan schedules;
• minimum-copy and minimum-source solving; and
• deterministic deck composition summaries.

Every probability returns an exact reduced fraction and a stable decimal. The caller supplies the population, groups, turn draws, mana capabilities, and keep rules. The MCP does not infer legality, roles, or whether a result is good.

Deck-backed statistics can select entries by entry ID, zone name, or category ID. This lets an LLM use existing Archidekt groups or deterministic categories without hiding the selected cards.

Synchronize Archidekt

Enable the archidekt toolset. Use remote mode for remote writes.

The adapter supports owned deck list/get/create/delete, pull, push, diff, folders, and named snapshots. Apply operations require current revisions, fingerprints, and preview evidence. The MCP never chooses a conflict winner.

Requests start at least two seconds apart per configured account. The adapter allows at most 30 starts in 60 seconds and 150 requests per tool invocation. It stops on 403 and 429. Ambiguous mutations are not retried.

Archidekt may renumber category positions. Content equality therefore compares category identity, membership, flags, and primary assignment, not the provider-controlled numeric rank.

Read Playgroup evidence

Enable the playgroup toolset. It exposes tools for the pinned Public API 1.0.0. All modes include 14 safe read tools. remote adds two write tools.

Results preserve provider JSON and include operation, API version, retrieval time, checksums, and limitations. The adapter does not rank decks or infer quality.

Request starts are at least 250 milliseconds apart. Reads have bounded retry. Writes are single-attempt. Live acceptance does not invoke writes because the public API provides no cleanup operation.

Evidence and safety

The server labels these kinds of result separately:

• provider facts;
• provider evidence;
• exact derivations;
• parser classifications;
• heuristics; and
• sampled estimates.

Stable 0.9.0 contains no advisor prompts, intent inference, weak-card judgments, replacement recommendations, blended quality scores, or strategic simulation. Deferred work is listed in Potential Features.

Develop and verify

Use Task as the command menu:

task --list
task lint
task test
task benchmark:oncurve
task surface:report
task coverage
task pack
task smoke:process
task smoke:mcp
task release:tool-smoke

Normal tests are deterministic and offline. Live tests require explicit opt-in. task benchmark:oncurve runs the separate fixed 99-card Release measurement. Each production assembly must maintain at least 90 percent line coverage.

Architecture

Read North Star, Design Goals, and the rewrite guide for durable constraints.

License

Licensed under the GNU Affero General Public License v3.0 or later.