MelisAIEngineClaude

Anthropic Claude provider for the Melis AI engine — package melisplatform/melis-ai-engine-claude.

Purpose

MelisAIEngineClaude is the concrete Anthropic Claude provider for the MelisAI suite. It extends the engine's MelisAIEngineModelService contract and supplies Claude-specific HTTP client setup, the Anthropic Messages API payload shape, file handling, tool/function mapping and response parsing. The module is stateless — it owns no database tables; all runtime state lives in the engine's own tables.

The engine selects this provider automatically at runtime when the active model's company contains Anthropic (and this module is installed).

Enable it

Add to config/melis.module.load.php after the AI engine:

php

return [
    'MelisAIEngine',
    'MelisAIEngineClaude',
];

Dependency: melisplatform/melis-ai-engine ^5.3. The Anthropic API key is stored per model in the Melis AI backoffice (the model's mapk_* key row) — not in a config file or environment variable.

Key services

Service alias	Role
`MelisAIEngineModelClaudeService`	The Claude provider. Extends `MelisAIEngineModelService` and implements the Anthropic Messages API contract. Built by the service manager with `['modelId' => …, 'agentId' => …]` factory options.

Notable methods on MelisAIEngineModelClaudeService:

Method	Role
`setClient()`	Builds a `Laminas\Http\Client` POSTing to the configured Claude Messages URL. Sets headers `Content-Type/Accept: application/json`, `x-api-key` (from the model's API key), `anthropic-version: 2023-06-01`, and an `anthropic-beta` header enabling prompt-caching and web-fetch betas. Uses a long timeout to accommodate tool-use loops.
`getMessageKey()`	Returns `'messages'`.
`addToolsToPayload()`	Reads agent functions via `getAgentFunctions()`, strips internal engine keys (`mcp`, `type`, `server`, `module`, `max_silence`, `operation_timeout`) via `sanitizeToolForClaudeAPI()`, appends tools in Anthropic shape (`name`, `description`, `input_schema`), appends the built-in `web_fetch` tool, and deduplicates by name.
`constructContent()` / `addContentToPayload()`	Maps role `model` to `assistant`; builds Anthropic content blocks (`{type:text}`, `{type:image, source:{url
`sendCustomAI()`	Injects `model` (`mam_generative_model`) and `max_tokens`; POSTs JSON with retries on 5xx; parses `content[]` into text and `tool_use` blocks; dispatches each tool (MCP via `MelisAIEngineMcpService::invokeTool`, otherwise built-in); appends results as `{type:tool_result, tool_use_id, content}` in a `user` message; loops until `stop_reason === 'end_turn'` or a safety hop-cap is reached. Enables prompt caching.
`processFiles()` / `processContextFiles()`	Converts uploads per the model's `mam_file_upload_mode` (default: `embed`): DOCX/XLSX → extracted text, images/PDF → base64, or via `uploadDocument()` when internal upload is enabled.
`setPromptTokenCount()` / `setResponseTokenCount()` / `setTotalTokenCount()`	Read `responseData['usage']` fields (`input_tokens`, `output_tokens`, plus cache-read/creation counters); total = input + output.
`getAllowedMimetypes()`	Returns allowed upload MIME types from the Claude config block.
`continueConversation()`	Rehydrates message history from saved state and calls `sendCustomAI()` again for multi-hop tool-use continuation.

Tool / function-calling

Engine tools already use Anthropic's input_schema shape, so addToolsToPayload() primarily sanitises and forwards them. A model tool call arrives as {type:'tool_use', name, id, input}; the service extracts name / input / id, routes via MelisAIEngineMcpService::isMcpTool() (MCP invokeTool vs. built-in function), and returns the result as {type:'tool_result', tool_use_id:<id>, content:<json>} inside a user message, then continues the loop. The id field correlates call↔result as required by the Anthropic API.

Configuration

Provider settings live under:

config['plugins']['melisaiengine']['datas']['AI']['Claude']

Key	Purpose
`api_url`	The Anthropic Messages API endpoint URL.
`allowed_mimetypes`	MIME types accepted for file uploads.

The API key is not read from config or environment — it comes exclusively from the model's DB key row (mapk_*).

Database tables

This module owns no tables. All conversation and usage state is stored in the engine's tables (melis_ai_conversation_state, melis_ai_daily_usage).

Key files

Concern	Path
Provider service	`vendor/melisplatform/melis-ai-engine-claude/src/Service/MelisAIEngineModelClaudeService.php`

MelisAIEngineClaude ​

Purpose ​

Enable it ​

Key services ​

Tool / function-calling ​

Configuration ​

Database tables ​

Key files ​

See also ​