Wikidesk Server

A companion server for LLM-wiki that turns named wiki repos into shared knowledge services for multiple AI coding agents. wikidesk doesn't care how your wiki is organized, what agent runs the research (Claude Code, Pi, OpenCode, Codex, etc.), or what prompts you use -- each configured wiki name just needs a derived repo directory holding the actual wiki content (wiki-{name}/wiki/, except default uses wiki/wiki/).

The primary workflow is simple: agents read local wiki mirrors directly as knowledge bases. Mirrors default to wiki-{name}/, except default uses wiki/, and clients can override them. On top of that, wikidesk optionally provides a research tool that lets agents request new research -- dispatching a dedicated research agent to investigate the question, update the server wiki repo, and return an answer. Whether your agents can trigger research or only read is up to you, controlled by your agent rules.

How it works

1. Agents read configured local mirror directories for existing knowledge
2. When an agent needs new research, it submits a question to /wiki/{name}/mcp or /wiki/{name}/api/research
3. The server queues the question for that wiki and spawns its configured research agent
4. The research agent investigates the question, potentially creating or updating pages under the server wiki repo
5. The answer is returned with [[wikilinks]] resolved to the client's local mirror path
6. Agents sync their local wiki copies automatically

Agent rules

Configure your agents to use the wiki by adding rules to your CLAUDE.md, AGENTS.md, or equivalent:

## Wiki

* The configured wiki mirror directory contains a knowledge base on <your topics>.
  Consult it before making decisions in these areas.
* Do not modify wiki files directly.

## Wiki

* The configured wiki mirror directory contains a knowledge base on <your topics>.
  Consult it before making decisions in these areas.
* Do not modify wiki files directly.
* When the wiki doesn't cover a topic you need, use that wiki server's `research` MCP
  tool to request investigation. Poll `get_result` until the task completes,
  then sync your local wiki copy.

Automatic wiki sync (client-server mode)

In client-server mode, local mirror directories need to stay in sync with the server. You can automate this using your agent harness's lifecycle hooks to run wikidesk sync at the start and end of each session.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "wikidesk sync"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "wikidesk sync"
          }
        ]
      }
    ]
  }
}

Security

The research agent runs with full permissions. The agent_command typically includes flags like --dangerously-skip-permissions (Claude Code) or equivalent settings that grant the child agent unrestricted system access. This is intentional — the research agent needs to read and write wiki files AND query random websites — but it means the research agent or the server spawning it must run inside a sandbox. Of course you can allow-list specific tools, websites etc. but in general case you want the LLM-wiki agent to freely browse the internet.

Recommended approaches:

• Docker/Podman: Mount only the wiki repo and config into the container.
• bubblewrap (bwrap): Minimal Linux sandboxing with filesystem and network restrictions.

I'm using a custom Nix+bubblewrap-based sandboxing tool (not yet released) for development.

Server setup

1. Install wikidesk-server

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ilya-epifanov/wikidesk/releases/latest/download/wikidesk-server-installer.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://github.com/ilya-epifanov/wikidesk/releases/latest/download/wikidesk-server-installer.ps1 | iex"

cargo install wikidesk-server

2. Set up your LLM-wiki

Follow the LLM-wiki setup instructions to create and configure each wiki repo. Name each repo wiki-{name} next to config.toml; the special wiki name default uses repo directory wiki. wikidesk requires each repo to contain a wiki/ content root. Keep wikidesk prompt templates outside that content root so the research agent cannot edit its own prompt.

3. Create a configuration file

See config.example.toml for all options.

# config.toml
bind_address = "127.0.0.1:1238"

[[wikis]]
name = "rlhf" # derives ./wiki-rlhf and /wiki/rlhf
description = "RLHF, preference optimization, DPO/PPO/GRPO/RLOO, reward modeling, and alignment training methods."
prompt_template = "prompts/rlhf.md"

# SECURITY: This command runs UNSANDBOXED by default.
# See the Security section -- always run the server in a container.
agent_command = ["claude", "-p", "$PROMPT", "--dangerously-skip-permissions"]

4. Start the server

wikidesk-server --config config.toml

5. Run as a daemon (optional)

To keep the server running across reboots:

mkdir -p ~/.config/systemd/user

cat > ~/.config/systemd/user/wikidesk.service << 'EOF'
[Unit]
Description=wikidesk research server
After=network.target

[Service]
Type=simple
WorkingDirectory=%h/wikidesk
ExecStart=%h/.cargo/bin/wikidesk-server --config %h/wikidesk/config.toml
Restart=on-failure
RestartSec=5

[Install]
WantedBy=default.target
EOF

systemctl --user daemon-reload
systemctl --user enable --now wikidesk
journalctl --user -u wikidesk -f  # check logs

cat > ~/Library/LaunchAgents/com.wikidesk.server.plist << EOF
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
  "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key>
  <string>com.wikidesk.server</string>
  <key>ProgramArguments</key>
  <array>
    <string>$HOME/.cargo/bin/wikidesk-server</string>
    <string>--config</string>
    <string>$HOME/wikidesk/config.toml</string>
  </array>
  <key>WorkingDirectory</key>
  <string>$HOME/wikidesk</string>
  <key>RunAtLoad</key>
  <true/>
  <key>KeepAlive</key>
  <true/>
  <key>StandardOutPath</key>
  <string>$HOME/Library/Logs/wikidesk.log</string>
  <key>StandardErrorPath</key>
  <string>$HOME/Library/Logs/wikidesk.log</string>
</dict>
</plist>
EOF

launchctl load ~/Library/LaunchAgents/com.wikidesk.server.plist
tail -f ~/Library/Logs/wikidesk.log  # check logs

$action = New-ScheduledTaskAction `
  -Execute "$env:USERPROFILE\.cargo\bin\wikidesk-server.exe" `
  -Argument "--config $env:USERPROFILE\wikidesk\config.toml" `
  -WorkingDirectory "$env:USERPROFILE\wikidesk"

$trigger = New-ScheduledTaskTrigger -AtLogOn

$settings = New-ScheduledTaskSettingsSet `
  -AllowStartIfOnBatteries `
  -DontStopIfGoingOnBatteries `
  -RestartCount 3 `
  -RestartInterval (New-TimeSpan -Seconds 10)

Register-ScheduledTask `
  -TaskName "wikidesk" `
  -Action $action `
  -Trigger $trigger `
  -Settings $settings `
  -Description "wikidesk research server"

docker run -d \
  --name wikidesk \
  --restart unless-stopped \
  -v /path/to/wikidesk:/etc/wikidesk \
  -p 1238:1238 \
  wikidesk-server --config /etc/wikidesk/config.toml

Logging

wikidesk logs to stderr and defaults to info. Use RUST_LOG for more detail:

RUST_LOG=wikidesk_server=debug,warn wikidesk-server --config config.toml

For systemd, add Environment=RUST_LOG=wikidesk_server=debug,warn under [Service]. For launchd, add the same variable to the plist environment.

Consumer workspace setup

HTTP paths are:

• GET /wiki -- list configured wikis
• POST /wiki/{name}/api/research -- submit research over HTTP; request may include local_path for wikilink rendering
• POST /wiki/{name}/api/sync -- sync a local mirror
• /wiki/{name}/mcp -- MCP endpoint for that wiki

There are two ways for agents to consume the wiki. Choose one.

Client-server mode (recommended)

Each agent machine runs wikidesk, which communicates with the server over HTTP. The client syncs configured local mirror paths automatically after each research request.

Install wikidesk

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ilya-epifanov/wikidesk/releases/latest/download/wikidesk-installer.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://github.com/ilya-epifanov/wikidesk/releases/latest/download/wikidesk-installer.ps1 | iex"

cargo install wikidesk

Configure and use

# Print a prompt for your coding agent to configure this repo's AGENTS.md/CLAUDE.md.
wikidesk agent setup http://your-server:1238 rlhf rust-notes
# With local path overrides: wikidesk agent setup http://your-server:1238 audio:wiki ml:wiki-ml

export WIKIDESK_SERVER_URL="http://your-server:1238"
export WIKIDESK_WIKIS="rlhf,rust-notes"
# Optional local mirror override syntax: name:relative/path, e.g. audio:wiki,ml:wiki-ml
# Local paths must be relative slash paths with no . or .. components.

# Submit a question to one wiki and sync it
wikidesk research -w rlhf "What is RLHF and how does it relate to DPO?"

# Sync all configured wikis, or one with -w/--wiki
wikidesk sync
wikidesk sync -w rlhf

wikidesk sync refuses to write into an existing directory unless it already has wikidesk's .gitignore marker. Move existing files out of the target path before the first sync.

Mount/symlink mode

Agents connect to the server directly via MCP. The wiki directory is mounted or symlinked into each agent's workspace for read access.

Caution: Ensure agents have read-only access to the wiki. Writing directly bypasses the server's research workflow and causes conflicts with concurrent research agents.

Configure MCP

Add wikidesk to your agent's MCP configuration:

claude mcp add wikidesk-rlhf --transport http http://your-server:1238/wiki/rlhf/mcp

Or add it manually to .mcp.json:

{
  "mcpServers": {
    "wikidesk-rlhf": {
      "type": "streamable-http",
      "url": "http://your-server:1238/wiki/rlhf/mcp"
    }
  }
}

The server exposes two MCP tools:

• research -- Submit a research question. Returns a task_id.
• get_result -- Poll for the result of a research task. Optionally pass local_path to render wikilinks for a non-default local mirror path.

Mount the wiki (read-only)

ln -s /path/to/wiki-rlhf/wiki ./wiki-rlhf

# Export on the server (add to /etc/exports):
#   /path/to/wiki-rlhf/wiki  agent-host(ro,no_subtree_check)

# Mount on the agent machine:
sudo mount -t nfs -o ro server-host:/path/to/wiki-rlhf/wiki ./wiki-rlhf

docker run ... -v /path/to/wiki-rlhf/wiki:/workspace/wiki-rlhf:ro ...

podman run ... -v /path/to/wiki-rlhf/wiki:/workspace/wiki-rlhf:ro,Z ...

# Requires Developer Mode or elevated prompt
New-Item -ItemType SymbolicLink -Path .\wiki-rlhf -Target C:\path\to\wiki-rlhf\wiki

# From within WSL2, the Windows filesystem is at /mnt/c/
ln -s /mnt/c/path/to/wiki-rlhf/wiki ./wiki-rlhf

Configuration reference

Runner types

wikidesk supports three runner types for executing research agents. These keys go inside each [[wikis]] table:

runner = "generic"
agent_command = ["claude", "-p", "$PROMPT", "--dangerously-skip-permissions"]

runner = "stream-json"
agent_command = ["claude", "-p", "$PROMPT", "--output-format", "stream-json", "--dangerously-skip-permissions"]

runner = "acp"
agent_command = ["claude-agent-acp"]

# Create settings file in the wiki repo
mkdir -p wiki-rlhf/.claude
cat > wiki-rlhf/.claude/settings.json << 'EOF'
{
  "permissions": {
    "defaultMode": "bypass"
  }
}
EOF

TODO

• Add simple UI for monitoring research request queues
• Manage multiple wikis, expose at different base HTTP contexts
• Add an optional jj workflow with isolated research workspaces and async Git remote sync
• Support Claude's streaming-json output mode, ACP for better progress monitoring

See also

• llmwiki-tool -- a companion CLI for wiki maintenance: fixing broken links, renaming pages with reference updates, detecting orphans, and linting against configurable rules

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.