pod
Other formats agents might prefer:
markdownjsonllms.txt

Agent? You probably want markdown or json, or Pod over MCP.

Reported issues for open-zk-kb

Pod holds 16 of 16 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to open-zk-kb.

Most discussed

feat: Docker deployment + remote MCP server support

Overview

Support running open-zk-kb as a centralized MCP server in Docker, allowing multiple clients to connect to a single shared knowledge base over the network.

Motivation

Currently, each client spawns its own local stdio MCP server with its own vault. This means:

  • Multiple vaults if using multiple clients (Claude Code, Cursor, OpenCode, etc.)
  • No way to share knowledge across machines
  • No cloud/NAS deployment option

A centralized Docker deployment would enable:

  • Single vault sh

Read the thread · 2026-04-05 · closed · 4 comments

Add ownership model: "Server computes, agent judges"

Overview

Codify a formal ownership model that defines what the MCP server should compute vs what the calling agent should decide vs what skills/docs should guide vs what the plugin should inject. Add this policy to AGENTS.md, docs/architecture.md, and skill instructions so it's enforceable — not just advisory.

This is foundational to every future feature decision (e.g., #77's "who finds related notes?", #86's "who surfaces domain notes?", #88's "what does the plugin own?").

Motivati

Read the thread · 2026-04-25 · closed · 3 comments

Add OpenCode plugin for guaranteed runtime hooks (session.created injection)

Overview

Ship @open-zk-kb/opencode-plugin as a sibling package to the MCP server. The plugin uses OpenCode's lifecycle hooks to provide runtime guarantees that an MCP server cannot: automatic context injection at session start, deterministic project-from-CWD detection, and survival of operating context across compaction.

The MCP server stays untouched. Other clients (Claude Code, Cursor, Windsurf, Zed) continue using the existing skill/managed-block delivery mechanism.

Motivation

The

Read the thread · 2026-04-25 · closed · 2 comments

Web-based vault viewer — browse knowledge base in browser via Quartz or Kiln

Overview

Add a web-based vault viewer as an alternative to Obsidian for browsing the knowledge base. Serve the vault locally as a website with graph view, search, backlinks, and wikilink resolution — inherently read-only, no Obsidian installation required.

knowledge-open                      # Opens in Obsidian (current, default)
knowledge-open --viewer web         # Opens in browser (this issue)

Motivation

The Obsidian experience (#118) is the primary vault viewer. But a web-b

Read the thread · 2026-05-01 · closed · 1 comment

Behavioral guidance delivery for Cursor and Zed

Overview

Cursor and Zed currently receive only the MCP server registration during install — no behavioral guidance is delivered. This is a gap surfaced by deep research on #93's ownership policy. Verified against CLIENT_CONFIGS in src/setup.ts:52-93: neither client has a skillPath or agentDocsPath.

The result: agents in Cursor/Zed see KB tools (knowledge-search, knowledge-store, knowledge-maintain, knowledge-ingest) but have no instructions on when to use them, **how

Read the thread · 2026-04-25 · closed · 1 comment

Add knowledge-open MCP tool to launch vault in Obsidian

Overview

Add a knowledge-open MCP tool that opens the vault in Obsidian. Detects whether Obsidian is installed, launches it pointing at the vault directory, and falls back to a download link if not installed. Gives users a one-command way to visually browse all their captured knowledge.

Motivation

The KB captures knowledge across projects via MCP tool calls, but there's no way for users to browse it outside of knowledge-search. The vault is already Obsidian-compatible by design

Read the thread · 2026-04-25 · closed · 1 comment

Add index and log note kinds for per-project navigation and history

Overview

Add two new note kinds — index and log — automatically maintained per project. The index is a deterministically-rebuilt catalog of all project notes grouped by kind; the log is an append-only chronological record of project events. A new knowledge-overview tool surfaces both in one call.

These two notes are core to the LLM Wiki pattern (Karpathy gist) and serve as the natural Obsidian entry points for any project's vault: open the index for navigation, open the log for hi

Read the thread · 2026-04-25 · closed · 1 comment

Add domain note kind for project operating manuals

Overview

Add a new domain note kind that captures the operating manual for a project — the agent's role, scope, conventions, named workflows, and boundaries. One per project, always surfaced in search results when the project filter matches (regardless of relevance ranking).

This is the equivalent of a project-specific AGENTS.md content, living inside the KB as a first-class note rather than as an external file. Lets the KB carry domain operating context — not just captured events.

Read the thread · 2026-04-25 · closed · 1 comment

Most recent

Feature: suggest-links maintenance action — server-computed link suggestions

Problem

link-health surfaces 300+ unlinked notes and 50+ one-way links, but provides no path from diagnosis to fix. The agent sees "Next Steps: [A] Add wikilinks" and has no idea which notes should link to which. Manually reading each note and guessing relationships is expensive and error-prone.

The diagnostic tools identify the problem but offer no resolution path.

Proposed: knowledge-maintain suggest-links

A read-only maintenance action that returns batched, server-computed link

Read the thread · 2026-05-21 · open · 0 comments

Enhancement: Include condensed kinds-with-fields in agent-docs injection

Description

The <!-- OPEN-ZK-KB:START --> managed block injected by injectAgentDocs() currently lists note kinds with word counts but not their key fields. Agents frequently need to call knowledge-template --kind {kind} before every store operation to get the schema.

Proposal

Include a condensed ~100-word summary of each kind's key fields directly in the injected block:

**Kinds** (with key fields):
- **personalization** (~50w) — title, content, guidance
- **decision** (~150w)

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/137) · 2026-05-11 · closed · 0 comments

### Add knowledge-mine MCP tool to bootstrap KB from existing OpenCode session history

## Overview

Add `knowledge-mine` MCP tool that backfills the KB from existing OpenCode session history (via `session_list` / `session_read` SDK APIs). Solves the cold-start problem: new users have value on Day 1, not Day 30.

## ⚠️ Blocked Until Curation UX Ships

This issue intentionally blocks on the **interactive curation review UX issue**. Reason: mining without curation pollutes the vault. A naive 30-day backfill might surface 100-300 candidate notes per user, of which maybe 10-30 are wort

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/107) · 2026-04-29 · closed · 0 comments

### Add model-aware capability detection via agent self-report

## Overview

Add an optional `model` parameter to MCP tool calls (`knowledge-store`, `knowledge-search`, `knowledge-maintain`) that lets calling agents self-report their model identity. Use this to classify capability level and gate feature richness accordingly. No user configuration required.

## Motivation

Several planned features (#76 URL ingestion, #77 auto cross-referencing, #80 richer skill instructions) return more complex responses that benefit capable models but may overwhelm weaker on

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/81) · 2026-04-23 · closed · 0 comments

### Add orphan and broken wikilink detection to knowledge-maintain

## Overview

Add two new `knowledge-maintain` actions: `orphans` (notes with no incoming or outgoing links) and `broken-links` (wikilinks pointing to non-existent notes). Both are pure SQL/filesystem checks with zero model dependency.

## Motivation

The current `knowledge-maintain` tool has `review`, `dedupe`, `stats`, and lifecycle actions — but no structural integrity checks. As knowledge bases grow, orphan notes and broken wikilinks accumulate silently. These are the two most common "wiki ro

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/78) · 2026-04-23 · closed · 0 comments

### Add knowledge-ingest tool for URL content extraction

## Overview

Add a new MCP tool `knowledge-ingest` that accepts a URL, fetches and extracts clean text content, and returns it to the calling agent for note creation via `knowledge-store`.

## Motivation

Currently, agents must rely on their own web-fetching capabilities (which vary wildly across clients) to capture external content. A server-side extraction tool provides:
- **Deterministic, consistent** content extraction regardless of calling model
- Clean markdown output stripped of nav, ads,

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/76) · 2026-04-23 · closed · 0 comments

### Package as Claude Code Plugin for Marketplace Distribution

## Summary

Package open-zk-kb as a Claude Code plugin for distribution via the official Anthropic marketplace, enabling one-command installation: `/plugin install open-zk-kb@claude-plugins-official`

## Background

Currently, users install open-zk-kb via:
- `bunx open-zk-kb setup install --client claude-code`
- Manual MCP config in `~/.claude.json`

A Claude Code plugin would provide:
- **Discoverability** — listed in official marketplace
- **One-command install** — `/plugin install open-zk-kb@

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/68) · 2026-04-02 · closed · 0 comments

### Replace CLAUDE.md injection with Claude Code skill for claude-code client

## Summary

Replace the `~/.claude/CLAUDE.md` managed-block injection for the `claude-code` client with a native [Claude Code skill](https://code.claude.com/docs/en/skills) installed to `~/.claude/skills/open-zk-kb/`. This uses Claude Code's built-in skill discovery mechanism instead of always-in-context instruction injection, saving ~340 tokens per session and aligning with the platform's native extension model.

Other clients (opencode, windsurf, cursor, zed) are unaffected.

---

## Motivatio

[Read the thread](https://github.com/mrosnerr/open-zk-kb/issues/28) · 2026-03-17 · closed · 0 comments

The remaining reports are on [the project's issue tracker](https://github.com/mrosnerr/open-zk-kb/issues).