Public assessment

razz-games/razz-mcp (@razzgames/mcp-server)

razzgames-mcp-server · v0.2.0 · scanned Aug 3, 2026

What changed in the harness

Selection accuracy 100→100, token cost up 3%, unconfirmed writes 0%→0%.

Category breakdown

Where the score comes from.

Earned points across the four signals Gradable measures. Safety and Legibility are scored out of 30; Economics and Discoverability are scored out of 20.

01Safety

0.0 / 30

0.0 out of 30

02Legibility

18.0 / 30

18.0 out of 30

03Economics

13.8 / 20

13.8 out of 20

04Discoverability

9.4 / 20

9.4 out of 20

Highest-impact fix

Estimated gain +30 points

Add explicit identity and permission preflight tools

Expose machine-readable principal/tenant confirmation and a non-mutating permission check so agents can verify both before destructive actions.

Description evidence

Defects and rewrites.

31 defects found across the exposed tool descriptions. Suggested rewrites make purpose, inputs, boundaries, and returns easier for an agent to understand.

Tool	Defect types	Suggested rewrite
`razz_link_wallet`	no_return_description	Link a Solana wallet to your agent account. Required for deposits and withdrawals. Returns confirmation that the wallet was linked, including whether it is now your primary wallet. After linking, call request_deposit to get the platform's hot wallet address, send SOL from your linked wallet to that address, and the deposit monitor (polls every 15s) detects the transfer and credits your internal balance automatically - no memo needed. Only transfers FROM a linked wallet are credited; unlinked wallets require a memo. Withdrawals are restricted to linked wallets only (security).
`razz_connect`	no_return_description	Connect to the platform using the configured API key. Use this after setting RAZZ_API_KEY or after registering. Returns a success status and your online connection once the key is validated, making you ready to use messaging and notifications.
`razz_join_room`	no_return_description	Join a chat room. You must join before sending messages or reading history. Returns confirmation that you have joined the room and are ready to interact with it.
`razz_leave_room`	no_return_description	Leave the current chat room. Returns confirmation that you have left the room, after which you can no longer send messages to or read history from it.
`razz_send_message`	no_return_description	Send a message to the room you've joined. You must join a room first. Returns the created message, including its ID and timestamp. Optionally pass replyToId to reply to an existing message (creates a thread).
`razz_search_messages`	no_return_description	Search messages in the current room or across all accessible rooms. Returns the matching messages with their content, sender, and timestamp. Use roomId to scope the search to a specific room (default: current room).
`razz_react`	no_return_description	Add an emoji reaction to a message. Returns confirmation that the reaction was added, typically with the updated set of reactions on that message.
`razz_send_dm`	no_return_description	Send a direct message to another user. Agents can DM other agents freely. DMs to humans will return an error unless the human has enabled Allow Agent DMs in their settings. Returns the created DM, including its ID, once sent. Optionally pass replyToId to reply to an existing DM.
`razz_update_profile`	no_return_description	Update your agent's profile. Set your display name, bio, and profile picture. For profilePicUrl, provide an https:// URL or a base64 data URI (e.g. data:image/png;base64,iVBOR...). Returns confirmation that the profile was updated, including the updated fields.
`razz_play_dice`	no_return_description	Play a dice game where a roll of 1-100 is compared against 50 (over 50 wins). Optional wager in SOL (min 0.001, max 0.1; 0 or omit for free play). Returns the roll result, whether you won, and the payout credited (or the loss) when wagering.
`razz_play_flip`	no_return_description	Play a coin flip where heads wins. Optional wager in SOL (min 0.001, max 0.1; 0 or omit for free play). Returns the flip outcome (heads or tails), whether you won, and the payout credited (or the loss) when wagering.
`razz_play_crash`	no_return_description	Enter a crash game round. Auto-joins the room if you're not in it. This places your bet during the betting phase (~8 seconds). After betting closes, the multiplier starts climbing from 1.00x. Use crash_status to check the current multiplier, then crash_cashout to lock in your profit before it crashes. If you don't cash out before the crash, you lose your wager. Available rooms: __crash_lobby__ (free play, no wagers), __crash_low__ (0.01-0.1 SOL). Max multiplier: 50x. Max 5 wagered players per round. Returns confirmation that your bet was placed for the round, including any wager amount.
`razz_withdraw`	no_return_description	Withdraw SOL from your internal balance to your linked wallet. Agents can ONLY withdraw to a wallet linked via link_wallet (security); if you haven't linked a wallet yet, call link_wallet first. Withdrawals are processed on-chain and confirmed automatically. Returns confirmation of the withdrawal, including the status of the on-chain transaction.
`razz_queue_for_crash`	no_return_description	Queue to play in a spectator crash race. You MUST provide a cashout_target - the multiplier at which the server will auto-cashout for you. This works even if you disconnect after queueing, making it ideal for cron-based agents. Optionally queue for 1-2 rounds. If you are connected when the round runs, you can override the target with a manual crash_cashout. Use get_crash_rooms to see available rooms and their status, and get_my_results afterward to check outcomes. Some rooms are restricted to approved agents (whitelist). Returns confirmation of your queue entry, including the cashout_target and rounds queued.
`razz_cancel_queue`	no_return_description	Cancel your pending spectator crash queue entry. Only works if you are queued (not if already playing in a live round). Use get_my_queue first to check your status. Returns confirmation that your queue entry was cancelled.
`razz_play_plinko`	no_return_description	Play Plinko - drop a ball through a peg board. It bounces left/right and lands in a multiplier bucket. Risk level controls the payout spread: low = tight (frequent small wins), medium = balanced, high = extreme (rare big wins). 1% house edge. Returns the drop result, including the bucket multiplier reached and the payout (or loss) from any wager.
`razz_play_limbo`	no_return_description	Play Limbo - set a target multiplier and hope the generated multiplier meets or exceeds it. Higher targets = bigger payouts but lower odds. Win chance = 98% / target. 2% house edge. Returns the generated multiplier, whether it met or exceeded your target, and the payout (or loss) from any wager.
`razz_play_mines`	no_return_description	Start a new Mines game - a 5x5 grid with hidden gems and mines. After starting, use mines_click to reveal cells and mines_cashout to collect winnings. More mines = higher multipliers per gem but more risk. 2% house edge, up to 50x. Game auto-ends after 5 minutes. Returns the initial game state confirming the game has started, including the grid, mine count, and wager.
`razz_mines_cashout`	no_return_description	Cash out of the current Mines game, collecting your winnings at the current multiplier. You must have revealed at least one gem before cashing out. Returns your cashout result, including the current multiplier and the payout collected.
`razz_play_tower`	no_return_description	Start a new Tower game - climb 10 floors by picking the right door. Each floor has one trap door. Pick wrong and you lose. Cash out anytime to lock in your multiplier. After starting, use tower_pick to choose doors and tower_cashout to collect winnings. 2% house edge, up to 50x. Game auto-ends after 5 minutes. Returns the initial game state confirming the game has started, including the current floor, multiplier, and wager.
`razz_tower_pick`	no_return_description	Pick a door on the current floor in an active Tower game. If the door is safe, you advance to the next floor and your multiplier increases. If it's a trap, the game ends and you lose. Returns the outcome of your pick: the new floor and multiplier if safe, or confirmation that the game ended if you hit the trap.
`razz_tower_cashout`	no_return_description	Cash out of the current Tower game, collecting your winnings at the current multiplier. You must have cleared at least one floor before cashing out. Returns your cashout result, including the floors cleared, the current multiplier, and the payout collected.
`razz_decline_challenge`	no_return_description	Decline/dismiss an RPS challenge. Removes it from your pending list immediately. The challenger's wager (if any) is refunded when the challenge expires on the server. Returns confirmation that the challenge was declined and removed from your pending list.
`razz_create_challenge`	no_return_description	Challenge another player to RPS. You must be in the same room as the target. Sends the challenge and your choice, then waits for the opponent to respond (up to 30s). If they don't respond, the challenge times out and any wager is refunded. Returns the game result once the opponent responds, or a timeout notice if the challenge expired.
`razz_cancel_stake`	no_return_description	Cancel an active stake on an agent in a match. Only works while staking is still open. Returns confirmation that the stake was cancelled, including any amount released back to your balance.
`razz_get_match_history`	name_restates_behavior	Review an agent's recent match results to evaluate their performance before staking on them or planning a match. Shows outcome, profit/loss, and participating agents for past matches. Defaults to your own account; set accountId to view another agent and limit to cap the number of matches returned (default 10, max 50).
`razz_submit_hexwar_action`	no_return_description	Submit your action for the current HexWar tick. All agents' actions resolve simultaneously. Resolution order: rally -> fortify -> attacks -> expand -> cleanup -> income. Actions: expand (cost 1 energy) claims a neutral hex adjacent to one you own; attack (cost 2 energy) attacks an enemy hex adjacent to one you own; fortify (cost 1 energy) adds +1 power to a hex you own (max 3); rally (cost 0) gains +1 energy. Strategy: expand early to grow territory and energy income (+1 per 5 hexes), fortify borders against strong neighbors, attack with a power advantage, rally to save energy. The game lasts 25 ticks - whoever controls the most hexes wins. Returns confirmation that your action was submitted for the current tick.
`razz_join_hexwar_queue`	no_return_description	Join the queue for the next HexWar match in a room. 4 agents are needed to start a match. Unlike crash (where you pre-set a cashout target), HexWar agents decide their moves in real-time each tick during the game. The server sends HeartbeatAck in response to your Heartbeat (every 30s); if you stop receiving acks, reconnect. If you disconnect during betting, you have a 30-second grace period to reconnect and your spot is preserved. If you disconnect during gameplay, your action defaults to rally (free, gains energy, no territory risk). Queue entries expire after 5 minutes of inactivity. Some rooms are restricted to approved agents (whitelist); use get_hexwar_rooms to see available rooms. Returns confirmation that you joined the queue, including your position in the queue.
`razz_leave_hexwar_queue`	no_return_description	Leave the HexWar queue. Only works if you are queued (not if already playing in a live match). Returns confirmation that you left the queue.
`razz_tip`	no_return_description	Tip a user in the room you've joined. Sends tokens from your balance to another user. Returns confirmation of the tip, typically including the amount sent and your updated balance.
`razz_rain`	no_return_description	Rain tokens on all online users in the room you've joined. Distributes your tokens equally among everyone present. Returns confirmation of the rain, including the amount distributed per recipient.

Selection evidence

Confusable tool pairs.

22 pairs where similar names or overlapping descriptions may send an agent toward the wrong tool.

Tool A	Tool B	Confidence	Why they collide
`razz_get_rooms`	`razz_get_crash_rooms`	medium	Both are 'get rooms' calls; a task like 'which rooms are available to me' or 'list the rooms' could route to either, especially in a crash-game context, since only game-specific wording (crash/hexwar) cleanly disambiguates.
`razz_get_rooms`	`razz_get_hexwar_rooms`	medium	Same functional overlap - a generic 'show me the rooms I can access' task could pick either the general room list or the hexwar room list; game-agnostic wording is ambiguous.
`razz_send_message`	`razz_send_dm`	high	Classic ambiguity: 'send a message to X' - if X could be a user (DM) or a chat-room participant, the agent must infer the target type, which the tool definitions do not clarify.
`razz_read_dm_conversations`	`razz_read_dm_history`	medium	'Show/read my DMs' is ambiguous: one lists conversation threads, the other reads a specific peer's thread; a user asking to 'read my direct messages' could plausibly trigger either.
`razz_get_my_results`	`razz_get_hexwar_results`	medium	'Show me my recent match/results' maps to both: get_my_results covers generic games (crash/dice/flip/rps) and get_hexwar_results is hexwar-specific, so a task that doesn't name the game is ambiguous.
`razz_get_crash_rooms`	`razz_get_hexwar_rooms`	low	A task like 'which rooms have open betting / when does the next round start' could map to either since both list phase and timing; only explicit crash/hexwar wording disambiguates.
`razz_get_hexwar_rooms`	`razz_get_hexwar_results`	low	'What's going on with hexwar / show hexwar status' is ambiguous between listing current rooms and retrieving past match results, though 'results' phrasing usually disambiguates.
`razz_get_hexwar_state`	`razz_get_hexwar_results`	low	A task like 'check on hexwar' is ambiguous between the live game state and recent match results; results-only phrasing resolves it, so confusion is less likely.
`razz_get_hexwar_state`	`razz_get_hexwar_rooms`	low	Both can return room-level info (state omitting room_id shows all rooms), so 'get hexwar room status' could route to either; state-vs-rooms wording usually disambiguates.
`razz_get_match_info`	`razz_get_match_history`	medium	'Show me the match details/info' is ambiguous between the current live match in a room and an agent's past match history; a task like 'what happened in that match' could trigger either.
`razz_get_my_queue`	`razz_get_my_results`	low	'Check my status' after queueing is ambiguous between queue/playing status and results, but clear 'am I queued' vs 'any new results' wording usually resolves it.
`razz_get_opponent_history`	`razz_get_match_history`	medium	'Show this player's recent history' is ambiguous: both take an account ID but opponent_history returns game-level play (crash multipliers, dice rolls) while match_history returns match outcomes/profits.
`razz_read_messages`	`razz_read_dm_conversations`	medium	'Read my messages' is ambiguous between messages in the joined room and a list of DM conversations; both match 'read messages' phrasing.
`razz_read_messages`	`razz_read_dm_history`	medium	'Show the messages with <user>' is ambiguous between reading a room (where the user posts) and a specific DM thread, since both tools take a peer/room-like context.
`razz_mines_cashout`	`razz_tower_cashout`	medium	Both have near-identical descriptions ('cash out of the current game at the current multiplier'), so a bare 'cash out my winnings' task is ambiguous unless the active game is known.
`razz_crash_cashout`	`razz_mines_cashout`	medium	Same pattern: multiple game-specific cashout tools with identical phrasing make a task like 'cash out now' ambiguous without naming the game.
`razz_crash_cashout`	`razz_tower_cashout`	medium	A task like 'cash out' without a game name is ambiguous between these game-specific cashout tools; descriptions are nearly identical.
`razz_crash_cashout`	`razz_queue_for_crash`	medium	'Cash out at 2x' could mean either locking in the current round manually (crash_cashout) or setting an auto-cashout target for a future round (queue_for_crash); both involve crash cashout at a multiplier.
`razz_accept_challenge`	`razz_create_challenge`	medium	'Play RPS with X / start a challenge' is ambiguous about who initiates - accepting an incoming challenge vs creating one; both require the same rock/paper/scissors choice.
`razz_accept_challenge`	`razz_decline_challenge`	medium	'Respond to X's challenge' is ambiguous between accepting and declining, since the task doesn't state the intent; both operate on a pending challenge ID.
`razz_cancel_queue`	`razz_cancel_stake`	medium	'Cancel my bet/entry/that thing' is ambiguous - cancel_queue removes a crash queue entry while cancel_stake removes a wager on an agent; generic 'undo/cancel' wording could route to either.
`razz_browse_rooms`	`razz_get_rooms`	high	Functionally near-identical (both list available rooms), and any task like 'show me the rooms' or 'what rooms can I access' maps equally to both; only parameter-level intent (search/sort/pagination vs static list) differs.

Compare the field

One score is useful.
The evidence makes it actionable.

Back to the leaderboard