Kibi MCP

What kibi-mcp MCP server does

The kibi-mcp MCP server exposes Kibi’s requirements and traceability workflow to coding agents. Kibi turns natural-language product intent into a structured, branch-local knowledge base containing requirements, scenarios, tests, semantic facts, and links between code symbols and the requirements they implement. It then checks whether that model remains coherent as code changes.

The knowledge base is intended to be maintained during normal feature, fix, and refactoring work. Requirements can be checked for complete semantics and scenarios, scenarios can be connected to tests, and proof-bearing tests can retain fresh execution evidence. The system also uses typed properties, predicates, and Prolog rules to identify contradictions, unsupported invention, and incomplete knowledge.

How it works

Kibi keeps state in the repository’s .kb/ directory. Git branches receive separate knowledge-base snapshots, so feature context stays associated with the branch where it was created. A detached checkout can answer reads from a read-only snapshot; writes wait until a branch is checked out. A branch created without Kibi’s Git hooks is reported as uncompiled, with kibi sync suggested to prepare it.

The agent can bootstrap a repository by identifying trusted sources such as issue trackers, wikis, and specifications through available connectors. Kibi drafts a plan with source citations, reports conflicts and open questions, and waits for approval before writing. The MCP workflow includes kb_plan_bootstrap and kb_apply_plan for planning and applying that process, along with kb_status and kb_check for inspecting and validating the knowledge base.

Setup and configuration

Kibi requires Node.js 22 or newer. Install kibi-core, kibi-cli, and kibi-mcp as development dependencies, then run kibi init through the project’s package-manager runner. Initialization creates the .kb/ layout and installs Git hooks used to keep it synchronized. The command preserves existing knowledge and schema files and does not overwrite hooks it did not create.

On Linux x64 or arm64 systems using glibc 2.28 or newer, and on macOS, SWI-Prolog is bundled. Other environments, including Alpine with musl and native Windows, require SWI-Prolog 9.0 or later on PATH; Windows is expected to use WSL. Connect the project-local MCP server to the coding-agent host, then load the kibi-bootstrap skill before bootstrapping.

Tools and capabilities

The kibi-mcp MCP server supports these documented MCP operations in the supplied material:

• kb_status reports knowledge-base state.
• kb_check validates the knowledge base and reports issues.
• kb_plan_bootstrap creates a proposed bootstrap plan.
• kb_apply_plan applies an approved plan.

Kibi includes built-in symbol handling for TypeScript and JavaScript. The optional kibi-plugin-treesitter adds offline extraction for Python, Go, Rust, Java, C#, PHP, C, C++, Bash, Ruby, and Terraform/HCL. An experimental UI plugin can add checks for React or Angular component design requirements when its markers are present.

Limitations and notes

The project is marked Beta. Kibi does not infer product knowledge during initialization; an agent or human must provide trusted intent sources and approve the bootstrap plan. External connectors are used by the agent to read source systems, but no specific connector is included in the supplied material.

Knowledge-base state remains local in .kb/, and the Kibi packages send no external telemetry or analytics. The optional UI plugin checks agreed design patterns and markers; it does not replace pixel, color, or spacing linting. On unsupported platforms, installing SWI-Prolog separately is required.