Ghostwriter MCP Server
Overview
Ghostwriter MCP Server is an executable Java 17 Model Context Protocol (MCP) server. The Maven project is named Ghostwriter MCP Server and publishes the org.machanism.machai:gw-mcp-server artifact. It combines Ghostwriter's project-aware code-assistance workflows with Bindex Core's metadata retrieval, registration, schema, and library-recommendation capabilities, and exposes them through the Machai MCP Server runtime.
The executable entry point is org.machanism.machai.mcp.server.McpServer. Launching the assembled JAR without a port uses STDIO; --port enables HTTP, and --session enables streamable HTTP sessions. The server can also receive a project directory, server name, advertised version, and properties configuration through the runtime command-line options. Its tools support project-aware assistance, model-assisted workflows, dependency metadata discovery, Bindex descriptor retrieval and registration, and library recommendations.
No package-info.java files are present in this source tree. This overview therefore uses the Maven project description, its declared Ghostwriter, Bindex Core, and Machai MCP Server dependencies, the executable entry point recorded in the project metadata, and the Bindex descriptor. Ghostwriter model-assisted workflows use the configured generative-AI provider; Bindex workflows use configured metadata sources and registry services.

The diagram source is src/site/puml/c4-diagram.puml. It shows the packaged server and its relationships with MCP clients, the configured AI provider, library metadata sources, and the Bindex registry. This module has no local Java components; the runtime and tools are supplied by its dependencies.
Supported AI providers
Ghostwriter uses the shared Machai GenAI Client provider layer. Its built-in external AI providers are OpenAI, Anthropic, and CodeMie; Tools and None implement the same lifecycle without an external AI service. Select a provider using gw.model=Provider:Model in a protected gw.properties file, or supply the corresponding runtime environment setting (gw_model). Keep credentials out of source control. The CodeMie model in bindex.json is an example from older build metadata, not a model default declared by this module's current POM.
| Provider | Supported configuration |
|---|---|
| OpenAI | For example, gw.model=OpenAI:gpt-4o-mini; supply OPENAI_API_KEY. Set OPENAI_BASE_URL for an OpenAI-compatible gateway. Supports text generation, function tools, embeddings, and optional web search and MCP forwarding. |
| Anthropic | Select an available Claude model using gw.model=Anthropic:MODEL; supply ANTHROPIC_API_KEY (API key or authorization token) and optionally ANTHROPIC_BASE_URL. Supports text generation, function-tool loops, web search, MCP forwarding, and prompt caching. |
| CodeMie | For example, gw.model=CodeMie:gpt-4o-mini; supply GENAI_USERNAME and GENAI_PASSWORD, with optional AUTH_URL. A username containing @ selects the password grant; other usernames select the client-credentials grant (client ID and secret). CodeMie supplies the delegated endpoint and bearer token: gpt-, gemini-, text-embedding-, codemie-text-embedding-, and amazon.titan-embed-text- model prefixes use OpenAI-compatible delegation; claude- uses Anthropic. Unsupported prefixes fail initialization. Actual model availability depends on the service. |
| Tools (local, not an AI service) | gw.model=Tools:yaml invokes registered host callbacks from a YAML tool-call descriptor containing tool and optional params. Requires registered callbacks and a prompt, but no external credentials. |
| None (no-op, not an AI service) | gw.model=None:log logs lifecycle activity; None:disabled is silent. Neither makes model requests or requires credentials. |
Common configuration parameters
Provider defaults below come from the shared GenAI Client documentation; server values come from the runtime documentation and this module's src/main/resources/mcp-server.properties. Model names in examples must be available to your account.
| Parameter | Description | Default or project value |
|---|---|---|
gw.model |
Provider-qualified model identifier used for Ghostwriter model-assisted requests. | No model declared in this module's POM or bundled properties; configure explicitly. |
OPENAI_API_KEY |
OpenAI or compatible-service API key. | No default; required for direct OpenAI access. |
OPENAI_BASE_URL |
OpenAI-compatible API endpoint override. | OpenAI SDK default. |
ANTHROPIC_API_KEY |
Anthropic API key or authorization token. | No default; required for direct Anthropic access. |
ANTHROPIC_BASE_URL |
Anthropic-compatible API endpoint override. | Anthropic SDK default. |
GENAI_USERNAME / GENAI_PASSWORD |
CodeMie user/password or OAuth client ID/secret. | No credential defaults; required for CodeMie. |
AUTH_URL |
CodeMie OpenID Connect token endpoint override. | CodeMie default endpoint. |
GENAI_TIMEOUT |
Request timeout in seconds; zero leaves the SDK default in effect. | 0 |
MAX_OUTPUT_TOKENS |
Maximum generated tokens. | 18000 |
MAX_TOOL_CALLS |
OpenAI Responses API tool-call limit; zero leaves it unset. Does not limit Anthropic tool loops. | 0 |
WebSearchTool.type |
Enables provider-specific web search; supported types depend on the provider. | Not set. |
WebSearchTool.city / WebSearchTool.country / WebSearchTool.region |
Optional web-search location hints. | Not set. |
MCP.name / MCP.url |
Name and URL of an MCP server forwarded to the AI provider; a name is required to register a group. Additional groups use MCP_1.*, MCP_2.*, etc. |
Not set. |
MCP.authorization / MCP.description |
Optional forwarded MCP authorization and description; numbered groups are also supported. | Not set. |
custom_configuration |
Additional tool configuration file selected by the bundled server properties. | gw.properties |
--projectDir / -d |
Project directory supplied to project-aware server workflows. | Not set; HTTP mode can determine it from a client request. Set explicitly for predictable project context. |
--config / -c |
Server properties configuration file loaded before startup. | mcp.properties; absence of the default is tolerated, but an explicitly selected file must be readable. |
--name / -n |
Server name advertised by the MCP runtime. | This module supplies mcp.server.name=gw-mcp-server through Maven-filtered resources. |
--version / -v |
Version advertised by the MCP runtime. | This module supplies its Maven project version through Maven-filtered resources (currently 1.4.3-SNAPSHOT). |
--port / -p |
Selects HTTP server mode. | Not set: the server uses STDIO. |
--session / -s |
Enables streamable HTTP sessions when HTTP mode is active. | Disabled. |
maven.compiler.release |
Java release used to compile the artifact. | 17 |
Maven properties can be overridden with -D options. The optional pack profile assembles the runnable release JAR and writes it to the directory represented by ${env.MACHANISM_PACK_DIR}. See the Machai MCP Server CLI documentation for the complete runtime option reference.
Resources
- Machai platform
- Machai GitHub repository
- Ghostwriter MCP Server GitHub repository
- Ghostwriter documentation
- GenAI Client provider and configuration documentation
- Bindex Core documentation
- Machai MCP Server CLI documentation
- Ghostwriter MCP Server artifact on Maven Central
- Machai MCP Server artifact on Maven Central
- Project Bindex descriptor

