ThinkWiki

ThinkWiki

Image: Release Image: License Image: Python

ThinkWiki is an agent-native local knowledge base skill. Talk to your agent to turn scattered documents, web pages, notes, and conversations into a durable Markdown workspace — with inbox review, local browsing, and an interactive content knowledge graph you can open in a browser.

Chinese version: README.zh.md

Repository: https://github.com/wzdavid/ThinkWiki

Works With Skills-Compatible Agents

ThinkWiki follows the open Agent Skills format (SKILL.md + bundled scripts). If your agent can install skills and run local commands, it can run ThinkWiki.

You do not need to memorize CLI commands. Install the skill once, then create and manage your wiki entirely through conversation.

Requirement: a local environment with Python 3. ThinkWiki bootstraps its own runtime on first use. Pure cloud chat without local file or shell access cannot run a local wiki.

Get Started: Install Through Conversation

Point your agent at this repository and ask it to install the skill for you.

Example prompt:

Please install the ThinkWiki skill from https://github.com/wzdavid/ThinkWiki , bootstrap the runtime, and run a health check to confirm everything works on my machine.

The agent will typically clone or copy the skill into your host's skills directory, run bootstrap, and verify capabilities with doctor. You should not need to run those steps yourself unless you prefer manual setup.

After installation, start using it immediately:

Create a local knowledge base named My Wiki in my workspace.

Clip this article into the inbox first: https://example.com/article

Import this PDF into the wiki: /path/to/file.pdf

Answer from my wiki: what did we already decide about context budgets?

Build the knowledge graph and open the workspace home in my browser.

The agent reads SKILL.md, maps your intent to stable ThinkWiki actions, and reports what changed — including paths to Markdown pages and HTML outputs.

Why ThinkWiki

• Agent-first: users talk to an agent, not to a long list of commands.

• Local-first: Markdown files remain the source of truth on your machine.

• HTML-first: inbox, viewer, graph, and governance pages are real browsable workspaces.

• Knowledge-first: the graph is a content knowledge graph — topics, concepts, entities, claims, and governance — not just a file-link map.

What It Produces

ThinkWiki generates a workspace that includes:

• output/index.html: unified workspace home

• output/inbox/index.html: inbox review console

• output/viewer/index.html: local page browser

• output/graph/index.html: interactive graph explorer

• output/graph/report.html: graph governance report

• output/graph/entity-merge-review.html: entity merge review page

• output/graph/entity-merge-plan.html: entity merge dry-run preview when requested

Preview

Workspace Home (output/index.html)

The unified workspace home summarizes recent changes, recommended next actions, inbox backlog, and a graph snapshot so you can decide whether to read, ingest, or explore.

Image: ThinkWiki workspace home preview

Inbox Review (output/inbox/index.html)

Review clipped web pages, files, and notes before formal ingest. Items are grouped into ready, review, and weak so you can process the highest-value captures first.

Image: ThinkWiki inbox review preview

Local Viewer (output/viewer/index.html)

Browse the entire wiki by page type, confidence, and status. Open any page directly in the viewer without leaving the local HTML workspace.

Image: ThinkWiki local viewer preview

Knowledge Graph (output/graph/index.html)

Explore the content knowledge graph interactively. Switch between knowledge, document, and suggested views to inspect semantic relations and candidate edges.

Image: ThinkWiki knowledge graph preview

Graph Governance Report (output/graph/report.html)

Inspect graph health signals such as isolated pages, hub stubs, fragile bridges, suggested links, and entity merge candidates before making structural changes.

Image: ThinkWiki graph governance report preview

Everyday Commands (In Plain Language)

Once ThinkWiki is installed, these are the kinds of requests users make most often:

Core Capabilities

• Initialize a local wiki workspace

• Clip web pages, files, or pasted text into an inbox for later review

• Import Markdown, PDF, DOCX, XLSX, XLS, PPTX, web pages, and plain text

• Answer questions from existing wiki pages with evidence-first behavior

• Save high-value outputs as query, synthesis, decision, or concept pages

• Build a content knowledge graph with knowledge, document, and suggested views

• Generate entity governance outputs, including ambiguous alias review and deterministic merge flows

• Check runtime health, workspace status, and graph governance state

Content Knowledge Graph

ThinkWiki v1.6.0 builds a schema v2 graph with default_view = knowledge.

The knowledge view can contain:

• page-backed nodes such as source, topic, concept, decision, synthesis, query, and entity

• claim nodes extracted from structured wiki content

• semantic relations such as about, belongs_to, depends_on, asserts, supports, contradicts, and suggests_related_to

This means the graph is not limited to raw file references. It can represent knowledge structure, evidence structure, and entity governance directly.

Browse HTML Outputs

Agent chat UIs usually cannot render ThinkWiki HTML inline. Ask your agent to start the built-in loopback server when you want to inspect inbox, viewer, graph, or governance pages:

Start the ThinkWiki output server and give me the workspace URL.

Or run it yourself:

python3 scripts/thinkwiki serve --root /path/to/my-wiki

By default this serves <wiki-root>/output/ at http://127.0.0.1:8765/. Start from http://127.0.0.1:8765/index.html.

On OpenClaw, after serve is running, the agent can also open the workspace home with:

openclaw browser --browser-profile openclaw open http://127.0.0.1:8765/index.html

On other agents, open the same URL in your system browser, or use whatever browser tool the host provides.

Use serve --print-urls when you only need the URL list without starting the server.

Environment Variables

Remote AI features are disabled by default. Configure only what you need:

Content generation (crystallize, digest)

Requires all three variables. Any OpenAI-compatible chat completion endpoint works (MiniMax, OpenAI, DeepSeek, local Ollama, etc.).

Without a complete LLM configuration, crystallize and digest use local heuristics only.

Entity embedding (entity-merge-review, graph-report, health)

Enabled when THINKWIKI_EMBED_API_KEY is set. Defaults to SiliconFlow BGE-M3; override the endpoint or model if needed.

Register a free SiliconFlow key at https://siliconflow.cn if you want the default embedding setup.

Manual Install (Optional)

If you prefer to set up the skill yourself instead of asking an agent:

git clone https://github.com/wzdavid/ThinkWiki ThinkWiki
cd ThinkWiki
python3 scripts/thinkwiki bootstrap
python3 scripts/thinkwiki doctor --repo-root .

Then install the folder into your agent's skills directory according to that host's documentation.

CLI Reference (For Agents And Power Users)

ThinkWiki keeps one stable entry point that agents call under the hood:

python3 scripts/thinkwiki <command> [args]

Common commands: init, clip, ingest, inbox, viewer, graph, graph-report, entity-merge-review, entity-merge-apply, serve, health, status, doctor.

See SKILL.md for the agent behavior contract and intent mapping.

Repository Docs

• README.md: concise English project overview

• README.zh.md: Chinese overview

• SKILL.md: skill behavior contract for agent hosts

• CHANGELOG.md: release history

License

MIT