Firebird MCP Server

Firebird MCP Server 🔥

🌐 Language / Idioma: English | Português (Brasil)

Open-source Model Context Protocol (MCP) server for Firebird Database, developed for free and independent community use with support for all major releases (Firebird 2.5, 3.0, 4.0, and 5.0).

Built according to the open industry standards for MCP servers (stdio / JSON-RPC 2.0), ensuring turnkey compatibility with any AI client or IDE: Google Antigravity IDE, Cursor, Claude Desktop, VS Code (with Cline / Roo Code / Copilot extensions), Windsurf, Zed, and CLI tools.

---

🎯 Key Highlights

1. Pure Python Wire Protocol (Zero DLL Dependencies):

• Connects directly over TCP (default port 3050) using pure Python (firebirdsql).
• Requires no native C/C++ compilation and eliminates the need to install fbclient.dll (Windows) or libfbclient.so (Linux) on the host running the AI client or IDE.
• Fully compatible with legacy authentication (Legacy_Auth in FB 2.5) as well as modern pluggable SRP authentication in FB 3.0, 4.0, and 5.0.

1. Multi-Database Support per Project:

• Allows a single project workspace to declare and manage multiple databases (e.g., main, audit_logs, staging), switching dynamically via chat prompts or tool arguments.

1. Workspace Auto-Discovery (`.firebird-mcp.json`):

• Automatically detects database configuration files placed at the root of your project or repository.

1. Execution Plan & Performance Analysis (`PLAN`):

• Compiles queries to extract the Firebird optimizer's execution plan (PLAN) without scanning entire datasets, detecting unindexed full-table scans (NATURAL SCAN) and memory/disk sorting bottlenecks (SORT).

1. Strict Query Safety & MVCC Best Practices:

• fb_query enforces read-only operations (SELECT), guarding against accidental data modification.
• Commits read transactions immediately upon completion, preventing **OAT (*Oldest Active Transaction*)** stalls, garbage collection starvation, and runaway .fdb database file growth.

1. Broad System Catalog Mapping:

• Comprehensive introspection of Firebird system catalog tables (RDB$), supporting types across versions: BOOLEAN (FB3+), INT128, DECFLOAT(16/34), TIMESTAMP WITH TIME ZONE (FB4+), BLOB SUB_TYPE TEXT/BINARY, and index selectivity (RDB$STATISTICS).

---

🗄️ Multi-Database & Per-Project Configuration

Real-world production codebases (such as enterprise ERPs in Delphi, Laravel, or Python) frequently require:

• Different databases per project/workspace.
• Multiple databases within the same project (e.g., operational database ERP_DATA.FDB and audit trail AUDIT_LOGS.FDB).

Where and How to Configure Databases:

Firebird MCP provides two flexible organization patterns:

Option A (Recommended Default): Central Global Catalog (~/.firebird-mcp.json)

Instead of scattering .firebird-mcp.json files across every repository, maintain a single central catalog in your user home directory (~/.firebird-mcp.json on Linux/macOS or C:\Users\YourUser\.firebird-mcp.json on Windows).

In this catalog, you register all databases on your machine and use the "projects" property to specify which projects/repositories use each database:

{

"default": "erp_main", "databases": { "erp_main": { "host": "localhost", "port": 3050, "database": "/var/data/ERP_DATA.FDB", "user": "SYSDBA", "password": "masterkey", "charset": "WIN1252", "description": "Primary ERP operational database", "projects": ["web_projects", "erp-desktop"], "tags": ["erp", "financial", "sales"] }, "erp_audit": { "host": "localhost", "port": 3050, "database": "/var/data/AUDIT_LOGS.FDB", "user": "SYSDBA", "password": "masterkey", "charset": "WIN1252", "description": "Secondary database for audit trails and logs", "projects": ["web_projects", "erp-desktop"], "tags": ["audit", "logs", "security"] }, "pos_store": { "host": "192.168.1.10", "port": 3050, "database": "C:\\POS\\STORE.FDB", "user": "SYSDBA", "password": "masterkey", "charset": "WIN1252", "description": "Store front checkout database", "projects": ["pos-vcl"], "tags": ["pos", "cashier"] }, "shared_staging": { "host": "localhost", "port": 3050, "database": "/data/STAGING.FDB", "user": "SYSDBA", "password": "masterkey", "charset": "WIN1252", "description": "Shared staging/test database accessible across all projects", "projects": [] } } }

Tip on Shared Databases: Databases with an empty "projects": [] array or "*" are treated as global shared databases and remain visible across all workspaces.

Option B: Local Project Configuration (.firebird-mcp.json)

If you prefer a repository to isolate its own credentials, place a .firebird-mcp.json file in the project root. The MCP checks for a local project file first; if present, it has absolute precedence over the global catalog.

---

How the Agent Interacts with Databases:

1. Automatic Workspace Context (CWD):

• When you open a workspace (e.g., /home/lucas/web_projects), the MCP automatically resolves the active project and filters databases matching that project.
• Single Database in Project: The MCP auto-selects the database without asking unnecessary questions.
• Multiple Databases in Project (e.g., main and audit): For generic commands (e.g., *"run a select on orders"*), the MCP halts and asks which database to use.

1. Direct Investigation Targeting (Zero Friction):

• If you mention the database in your prompt (e.g., *"investigate bug on audit database"* or *"check table X on pos"*), the agent/MCP resolves the alias immediately and connects without asking confirmation questions.

1. Smart Approximate Search:

• Switch databases using fb_use_database("audit") — the MCP matches exact names, substrings (e.g., erp_audit), and associated tags.

1. ⚠️ Rigid Safety Rule: Mandatory Double Confirmation for DDL and Data Manipulation (DML):

• Structural commands (CREATE, ALTER, DROP, RECREATE) and data mutation commands (INSERT, UPDATE, DELETE, EXECUTE PROCEDURE) are NEVER executed directly without prior explicit user consent.
• The assistant must present the exact SQL/DDL statement, explain its technical purpose, detail the impact (affected rows/tables, lock risks, irreversibility), and request confirmation.
• The MCP server blocks execution with an error unless the confirmed=True parameter is passed following the user's explicit affirmative response.

---

🛠️ Available MCP Tools

Database & Connection Management

Schema & Metadata Introspection

Performance Analysis & Execution

---

💻 Installation & Execution

🐳 Option 1: Docker (Gold Standard — Recommended for Windows, macOS & Linux)

Running via Docker eliminates all local Python runtime and dependency requirements:

# Clone and build the image locally:

git clone https://github.com/lucasgabriel0802/firebird-mcp.git cd firebird-mcp docker build -t firebird-mcp:latest .

Universal IDE Configuration via Docker (Stdio Mode):

Configure your IDE to launch the container on-demand. This configuration works seamlessly on Windows, macOS, and Linux:

{

"mcpServers": { "firebird": { "command": "docker", "args": [ "run", "-i", "--rm", "-v", "${workspaceFolder}:/workspace", "-w", "/workspace", "--add-host", "host.docker.internal:host-gateway", "firebird-mcp:latest" ] } } }

Why this is the Gold Standard:

- -i: Keeps stdin/stdout open for JSON-RPC without allocating a pseudo-TTY (no -t to avoid byte corruption).

- --rm: Cleans up the container automatically when the IDE closes.

- -v "${workspaceFolder}:/workspace": Mounts your current project for automatic .firebird-mcp.json discovery.

- --add-host host.docker.internal:host-gateway: Resolves your host machine's IP (e.g. Firebird running on port 3050 in Windows, macOS, or Linux) using host.docker.internal.

---

🚀 Option 2: Docker Compose (Background Service & Batteried Testing DB)

A. Run as an HTTP/SSE Background Service

If you prefer running the MCP server as a permanent background service on port 8000:

docker compose up -d firebird-mcp

Connect via SSE endpoint: http://localhost:8000/sse.

B. Full Local Testing Environment (with Firebird 5.0 Database)

Spin up both the MCP server and a ready-to-te