Plugins & integrations

Kravn's capabilities are plugins — nothing is hard-wired. That keeps the core small and lets you extend the catalog and govern every call without forking the product.

There are two kinds:

• mcp-server plugins — provide an in-process MCP server (tools/resources/prompts) that Kravn exposes and composes like any other upstream. This is how the native corporate integrations work.
• hook plugins — Apigee-style interceptors that transform tool-call requests, results and the advertised tool list at each lifecycle junction. This is how governance pipelines are built.

Built-in plugins ship in the console's Plugins marketplace — search, filter, enable, and configure them. Credential-bearing plugins document the exact permissions they need right in their setup screen.

Native integrations

Talk to the corporate systems you already run — no separate MCP server to operate. Each is a native mcp-server plugin; you enter credentials once and its tools join the catalog.

Integration	What it does
SharePoint	Search, browse document libraries, and read documents (Word/PDF/Excel/text extracted to text) over Microsoft Graph (app-only).
Microsoft Teams	Find people, find/list/read chats (scopable by a date range), list teams/channels and read channel posts, over Microsoft Graph (app-only).
Jira	Query and read issues via the Jira REST API.
Confluence	Search and read pages.
Code Interpreter	Run Python in a WASM sandbox (Pyodide — no host filesystem or network) to read and transform attached files, e.g. complete a spreadsheet and return it.

Credentials for these are stored encrypted at rest, and access is governed by the same team model as everything else — you compose their tools into a restricted MCP endpoint for the teams allowed to use them.

Governance & content pipelines

Hook plugins are composed into pipelines — an ordered chain per lifecycle junction. A global pipeline runs for all traffic; each endpoint can add its own overlay that can only add steps, never remove or bypass a global one. Built-in hooks include:

Hook	What it does
Secrets Redactor	Detects and strips private keys, cloud/API tokens, JWTs and credentials from results before they reach the model.
PII Tokenizer	Replaces emails, IPs, credit cards (Luhn-checked) and phone numbers with stable, deterministic tokens so the model reasons consistently without seeing the real values.
Prompt-Injection Guard	Flags/neutralizes indirect prompt injection in tool output ("ignore previous instructions", role-tag spoofing, exfiltration directives) — the #1 MCP-specific risk.
Content Safety Filter	Lexicon-based self-harm / violence / hate detection with redact or annotate, and optional request blocking.
Deny List Filter	Block requests and/or redact results by phrase or regex.
SafeHTML Sanitizer	Strips common XSS vectors from HTML results (script/iframe/on*/javascript:).
Audit / Compliance Logger	Writes a tamper-evident, hash-chained audit record for every tool call.

Fewer tokens, cleaner context

Several plugins exist specifically to shrink what reaches the model — lower cost, better focus:

• HTML → Markdown turns verbose HTML tool/resource results into clean Markdown.
• Document extraction in the integrations renders DOCX to structured Markdown (headings, lists and — the real win — tables), and PDF/Excel to text, instead of shipping raw binary.
• TOON Encoder re-encodes uniform JSON arrays into a compact tabular notation, cutting 30–70% of tokens.

Writing your own

The plugin contract lives in @kravn/plugin-sdk. A plugin is a plain object (the default export of an ES module) with a manifest and either a server (for mcp-server) or hooks (for hook). See PLUGINS.md in the repository for the full guide.

---

Next: Security & compliance.