Sakuraseisai AI Sales Roleplay is an AI-assisted sales-training integration developed by Sakuraseisai Co., Ltd. Its existing five-tool MCP endpoint supports Guest Training with fictional sales situations, temporary session controls, seller-turn evidence and limited practice feedback. Guest mode is enabled on the Tokyo endpoint for MCP-compatible clients; no Sakuraseisai account or user-supplied API key is required by the server for the Guest workflow. This is not an approved or directory-listed Muse connector. Muse directory support for credential-free installation remains unverified.

1. Developer and Service Overview

Developer: Sakuraseisai Co., Ltd.
Product: Sakuraseisai AI Sales Roleplay
Primary use: Fictional B2B sales practice and human capability development
Tested languages: Japanese (ja) and English (en)
Support: info@sakuraseisai.com

The MCP service manages the fictional scenario, temporary training state and seller evidence; it does not run its own conversational buyer model. A private Bearer-authenticated Muse connector successfully completed a three-turn Japanese roleplay with Muse acting as Buyer on November 3, 2026. The host-generated coaching commentary is distinct from the MCP’s synthetic ledger-template feedback; neither constitutes a certified skills score.

2. Integration Status and Hosted Endpoint

Connection type: Existing remote Streamable HTTP MCP
Hosted endpoint: https://muse-mcp.sakuraseisai.com/mcp
Transport: HTTPS MCP/JSON-RPC; use a compliant Streamable HTTP MCP client and the negotiated protocol headers
Current Tokyo image: 0.4n.0-guest-rc1
Five-tool contract identifier: 4l1-private-trial-mcp-v1 (inherited from the original candidate)
Environment: Tokyo guest-enabled staging candidate; synthetic scenario pump-v1 only
MCP serverInfo.name: sakuraseisai-private-trial-mcp

Verified: real Muse private Bearer connector tool calls and three seller turns; remote MCP SDK tests; Tokyo public Guest initialize, tools/list, three-turn practice, session isolation and invalid-Bearer rejection. A separate direct, unauthenticated client in the Muse workspace successfully initialized and discovered all five tools. Not verified: creation or publication of a no-auth connector through the official Muse Connector Directory; account-specific per-tool OAuth in Muse.

The endpoint is an MCP service, not a browser web page. Ordinary browser GET requests are not the training workflow. With Guest enabled, a compliant client may initialize without an Authorization header; malformed or invalid supplied Bearer tokens are still rejected with HTTP 401. Existing valid Bearer test credentials remain accepted separately.

3. Supported Practice Workflow

  • Prepare: Call coach_training_prepare with fixture_id: "pump-v1" and language ja or en; creates a pending practice session.
  • Start: Call coach_training_start with the exact current localized Start label from the session controls. The session must be active before seller turns can be recorded.
  • Buyer brief: Call coach_training_buyer_brief to retrieve buyer persona, context, and factual boundaries for the host AI.
  • Respond: Call coach_training_respond to commit each seller utterance exactly as supplied; the tool does not rewrite seller wording and does not generate buyer natural-language dialogue.
  • End: Call coach_training_end with the required mode (and label when applicable) to end the session and return evidence-bounded qualitative feedback.

The connector does not itself produce the buyer’s conversational natural-language response through an external paid language-model provider. A Japanese three-turn private Muse-hosted roleplay has already demonstrated that the host can act as Buyer; official directory onboarding and broader reliability remain separate review questions.

4. Available MCP Tools

The same five tools remain deployed in 0.4n.0-guest-rc1; their original business contract identifier remains 4l1-private-trial-mcp-v1. Fields were reconciled against live tools/list in the original 4L-1 SDK verification; Guest adds a separate access profile without changing these five tool contracts. Tool results are returned as MCP CallToolResult with both structuredContent (JSON object) and a text content item containing the same JSON.

Common successful tool-result envelope fields (product output schema): candidate_version = 4k2h-host-native-v1; simulation_only = true; real_model = false; muse_verified = false (a legacy product-envelope flag; not a statement that the real Muse host test was never performed); buyer_architecture = MUSE_HOST_NATIVE; persistence = BLOCKED_NOT_IMPLEMENTED (no personal-scenario or transcript persistence); state ∈ COMMITTED | FAILED_NO_COMMIT; replayed (boolean); plus tool-specific data.

4.1 coach_training_prepare — Write

Purpose: Prepare a new sales-training session from the fixed synthetic fixture.
Permission: training:write
Required inputs: simulation_only (const true), request_id (string matching ^[A-Za-z0-9_-]{1,80}$), language (ja | en), fixture_id (const pump-v1).
Notes: The MCP wire schema for this candidate exposes only the pump-v1 fixture arm. Custom scenario/version preparation is not part of the published tools/list schema for this staging candidate.
Output data (high level): data.ok, pending session object including session_id (UUID), version, attempt_id, revision, localized controls.start_label / end_label / continue_label, and related session fields.
Side effect: Creates in-memory session state before training begins.
Annotations: readOnlyHint=false, destructiveHint=true, idempotentHint=false, openWorldHint=false.

4.2 coach_training_start — Write

Purpose: Explicitly start a prepared session.
Permission: training:write
Required inputs: simulation_only (const true), training_session_id (UUID), request_id, expected_version (integer ≥ 1), attempt_id (1–80 chars), revision (integer ≥ 0), label (1–120 chars).
Label rule: label must equal the current localized Start control exactly. Staging values are English Start Roleplay and Japanese ロールプレイを開始; always prefer the value returned in session controls.start_label.
Output data (high level): active-session status and host-facing practice context; may include a host-native buyer-continue cue rather than server-generated buyer dialogue.
Side effect: Transitions the session to an active training state.
Authorization note: A conversational claim such as “already approved” is not an authorization substitute.

4.3 coach_training_buyer_brief — Read

Purpose: Retrieve the currently applicable buyer persona, practice context, and evidence boundaries.
Permission: training:read
Required inputs: simulation_only (const true), request_id, training_session_id (UUID).
Output data: ok, training_session_id, attempt_id, language, mode, buyer_persona, constraints (max_main_questions=1, preserve_unknown=true, no_price_or_policy_invention=true, do_not_sell_for_seller=true, seller_evidence_requires_tool=true), unknown_fact_ids, expected_objection_types, seller_turn_count, turn_cap.
Side effect: None on the training session.
Authorization: Credentials with only training:read discover and may call this tool only; Write tools are not listed or executable for that principal.

4.4 coach_training_respond — Write

Purpose: Record one seller turn in an active practice session.
Permission: training:write
Required by inputSchema: simulation_only (const true), training_session_id, request_id, expected_version, attempt_id, revision, text (1–4000 chars), substantive (boolean).
Also required by runtime product gate: text_provenance ∈ UNVERIFIED_HOST_SUPPLIED | USER_PASTE_CLAIMED. This property appears in the published schema enum but is enforced as mandatory by the business layer (SELLER_TEXT_PROVENANCE_REQUIRED if omitted). It is a host claim, not platform verification that Muse captured the utterance verbatim.
Output data (high level): acknowledgment or rejection of the turn; on success, updated session/version state and the recorded seller turn text. Successful responses also include seller_evidence_trust (echo of provenance), verbatim_capture: "NOT_PLATFORM_VERIFIED", and an evidence note. buyer_turn is not generated by this connector.
Side effect: Appends a seller turn. Calls after session end must be rejected.
Seller-text integrity: Tokyo remote SDK tests confirmed codepoint-exact retention of Japanese and English seller text supplied through the tool arguments. That does not prove Muse will pass each original user utterance unchanged. If host-supplied text cannot be treated as verbatim evidence, do not present paraphrases as exact quotations; a proposed host-side degradation label used in internal trial planning is LIMITED_EVIDENCE_NOT_VERBATIM, but that string is not asserted here as an automatically emitted server feedback code.

4.5 coach_training_end — Write

Purpose: End an active training session and request evidence-bounded feedback.
Permission: training:write
Required inputs: simulation_only (const true), training_session_id, request_id, expected_version, attempt_id, revision, mode ∈ request | confirm | now | continue.
Conditional: when mode is confirm or continue, label is required and must match the current End or Continue control; when mode is request or now, label must not be supplied. Staging End/Continue labels are English End Roleplay & View Feedback / Continue Roleplay and Japanese ロールプレイを終了してフィードバックを見る / ロールプレイを続ける; prefer values from session controls.
Output data (high level): ended-session status and qualitative feedback based on confirmed session evidence. The staging candidate does not provide calibrated numerical competency scores.
Side effect: Transitions the session to the ended state.

5. Authentication and Access Requirements

Current server access model: Guest Training is enabled on the Tokyo MCP endpoint. Anonymous requests without an Authorization header can initialize, discover, and use all five tools with the fixed synthetic pump-v1 scenario. No Sakuraseisai SaaS sign-in or user-supplied API key is required for this server-side Guest flow. This does not establish that the Muse Directory can install a credential-free MCP connector.

  • Guest mode: Unauthenticated session-bound temporary practice only; no user account owner, Supabase database access, personal Scenario list/get/save, or durable transcript archive. The server enforces isolation and resource limits.
  • Existing Bearer mode: The previously tested private Muse connector continues to support independent synthetic Bearer grants. Bearer credentials use scopes such as training:read and training:write and their existing configured validity rules; no Bearer token is publicly published here.
  • Invalid token behavior: A request that actually supplies an invalid/expired Bearer credential returns HTTP 401 and never silently downgrades into Guest.
  • Protocol session: Use the returned MCP-Session-Id, send notifications/initialized, and maintain the session’s server-side identity binding. Unrecognized or expired sessions are rejected.
  • Muse platform compatibility: Credential-free initialization and discovery were directly tested from a Muse workspace client, but the private credential-registration workflow has no documented auth_scheme: none option. Official directory no-auth acceptance, hosted user installation and per-tool/incremental OAuth remain unverified.
  • Personal Scenario functions: Not part of these five tools. A later account-enabled release would require trustworthy Clerk OAuth identity, permissions and explicit user action; it is not presently offered in Muse.

6. Proposed Tool Permissions and User Approval

The following classifications describe the actual tool side effects and are proposed for Muse review. They have not yet been approved or assigned by Meta:

  • coach_training_prepare — Write (creates session state).
  • coach_training_start — Write (activates training).
  • coach_training_buyer_brief — Read (retrieves context only).
  • coach_training_respond — Write (records a seller utterance).
  • coach_training_end — Write (ends session and generates feedback).

Approval requirements are controlled by the host platform and the service’s authorization rules. The connector does not bypass human-in-the-loop approval controls. In particular, recording each seller turn is a state-changing Write operation; the frequency and usability of approval prompts during continuous conversation remain subject to real Muse testing.

7. Data Flow, Retention, and Privacy

For the Guest workflow, the MCP service processes a fictional scenario identifier and constraints, Seller text submitted in tool calls, temporary training-state and evidence metadata, and protocol session identifiers. It does not require or derive a Sakuraseisai SaaS account identity. Existing Bearer test flows remain scoped to their own synthetic grants.

  • Temporary practice: The five tools hold training evidence in process memory. Guest training uses bounded temporary sessions and expiry/cleanup; process restart also clears in-memory sessions. This is not a guarantee that Meta/Muse chat history is deleted.
  • No personal database access: The current five tools do not read or write Supabase personal Scenarios, training history, CRM data or customer records. Guest access to personal Scenario functionality is denied.
  • Host-side conversational AI: Muse provides Buyer natural-language dialogue in the host chat; the MCP path does not call a separately billed external buyer-model API. The server’s current feedback is a synthetic ledger template, not validated skills certification.
  • Operational logs: Infrastructure may process request/diagnostic metadata separately from temporary practice state. Tokyo nginx logs are configured for daily logrotate with 14 rotations; Docker json-file logging has no verified size/retention cap in the last read-only assessment. Do not infer a common fixed deletion period for all logs.
  • Data minimization: Users should submit fictional or appropriately anonymized practice information and avoid real customer personal data, credentials, payment information or trade secrets.

Refer to our English Privacy Policy and its Muse-specific addendum when published. Muse/Meta independently governs its host-side conversations, account controls and platform data retention. Sakuraseisai does not claim deletion authority over the Muse chat history or third-party copies.

8. Errors, Session State, and Recovery

  • Guest initialization: Without an Authorization header, MCP-compliant initialization can succeed when Guest Training is enabled; absence of a Bearer credential is not itself an error in that mode.
  • Invalid/expired Bearer: HTTP 401; never reclassified as Guest. When Guest is administratively disabled, anonymous initialization also returns 401 AUTH_REQUIRED.
  • Insufficient permissions or attempted personal Scenario access: Refused; Guest has no Supabase/owner identity or personal Scenario rights.
  • Session errors: Missing/expired MCP session, stale version, wrong confirmation label, or Seller response outside the ACTIVE state are rejected rather than treated as committed.
  • Capacity: Anonymous requests can encounter HTTP 429 or protocol errors when guest session and request ceilings are reached; retry with a fresh valid session after backing off.
  • Body/headers: Requests above the configured body/header limits are refused (for example HTTP 413 or 431).
  • Evidence caution: The service retains submitted Seller tool text, but does not cryptographically establish whether Muse passed the original user utterance verbatim. Do not claim fabricated exact quotations or certified scores.
  • Unknown outcome: A failed tool call is not proof of a successful Write. Check returned state/version and avoid blind duplicates.

9. Limits and Current Restrictions

  • Release status: Guest mode is enabled on the publicly reachable Tokyo MCP server. This is not approval or availability through the official Muse Connector Directory.
  • Scenarios: Current five-tool MCP accepts the fictional pump-v1 fixture only. No personal saved Scenario access, custom scenario loading, transcript storage or certified skill scores.
  • Languages: Japanese and English have been exercised in synthetic tests; real Muse Bearer-hosted three-turn Japanese roleplay has also been demonstrated.
  • Guest capacity: The guest service enforces bounded global sessions, initialization rates, requests and session lifetime. Exact operational ceilings may change and do not guarantee capacity per Muse user.
  • Resource protection: Container resources and temporary session memory are capped. Service availability and throughput are not guaranteed.
  • Usage and abuse controls: Guest sessions are temporary and rate-limited. Global caps may affect legitimate users; no claim of verified platform-specific identity or guaranteed per-IP nginx throttling is made.
  • Payment and model costs: These tools do not process payments and current Guest MCP execution does not call an external paid Buyer AI API; Muse host usage may be subject to its own applicable terms, usage rules or fees.
  • Directory authentication: Whether Meta accepts unauthenticated MCP as a listed connector, or supports OAuth only on later personal Scenario actions, has not been verified.

10. Reviewer Test Procedure (Guest-First)

For the initial five-tool synthetic training release, the server can be tested by a compliant remote Streamable HTTP MCP client without an Authorization header. The official Muse Directory’s ability to establish an equivalent credential-free user connection is pending platform confirmation; do not represent this client test as directory acceptance.

  • Connect to https://muse-mcp.sakuraseisai.com/mcp over HTTPS; send initialize without an Authorization header; retain the MCP-Session-Id and send notifications/initialized.
  • Call tools/list and confirm exactly five named training tools.
  • Call coach_training_prepare with simulation_only: true, a unique valid request_id, language: "ja" and fixture_id: "pump-v1".
  • Use the returned controls and session version to call coach_training_start; then call coach_training_buyer_brief.
  • Run three Japanese Seller turns, each through coach_training_respond with the matching IDs/version and text_provenance; inspect the evidence records.
  • Call coach_training_end in request mode, then confirm mode with the exact returned localized End label; confirm ENDED and the synthetic practice feedback.
  • Negative checks: invalid Bearer must remain HTTP 401; a Guest cannot read/save personal Scenarios or cross into another training session; stale/wrong state and capacity errors must be handled.
  • In a real Muse host test, separately observe Write approval prompts, seller-text fidelity and Buyer behavior. The earlier Bearer private connector passed Japanese three-turn roleplay; official no-auth Connector Directory onboarding remains an open review question.

Existing reviewer-specific Bearer access may still be used for authenticated regression testing. Secrets are never posted on this public documentation page, and Reviewer credentials are not a prerequisite for the server-side Guest path unless required by Meta’s actual review process.

11. Technical Support and Documentation Maintenance

Sakuraseisai Co., Ltd.
Technical and privacy inquiries: info@sakuraseisai.com
Website: https://sakuraseisai.com/
Contact: https://sakuraseisai.com/en/contact/

The operator updates this page when deployment, permissions, testing evidence or actual data processing changes. Significant tool or authentication changes may require a fresh platform review; none are represented here as already approved by Meta.

Document status: Guest-first Tokyo candidate; technical documentation prepared for Muse Connector review
Runtime image: 0.4n.0-guest-rc1; inherited five-tool contract identifier 4l1-private-trial-mcp-v1
Muse real-host Bearer three-turn test: November 3, 2026
Guest long-running acceptance: November 4, 2026
Document revised: November 4, 2026
Official Muse directory approval: Not obtained / not verified