Sombra — Reader Mode for AI#

Persistent reader mode for AI. Give your OpenClaw agent a research library that lasts between sessions. Save permanebt copies of web pages as pristine markdown, organise them into collections, distil the important parts, and access it all through MCP.

What this skill does#

Sombra is an MCP server that provides your agent with persistent knowledge management. Instead of starting every session from scratch, your agent reads from a curated library of web pages, notes, and distilled context that you (or the agent) have built up over time.

• Save web pages as clean markdown with ads, nav, and layout stripped out
• Organise into collections scoped to projects or tasks
• Distil collections into dense context summaries that preserve code blocks verbatim
• Search across everything you've saved
• Version history on every change, so you can roll back if the agent rewrites something you preferred

Setup#

1. Get a Personal Access Token#

Sign up at sombra.so and create a token in Settings > Access Tokens.

2. Configure OpenClaw#

Add Sombra to your openclaw.json:

json

{
  "mcpServers": {
    "sombra": {
      "command": "npx",
      "args": ["-y", "sombra-mcp", "--token", "sombra_pat_YOUR_TOKEN_HERE"]
    }
  }
}

Replace sombra_pat_YOUR_TOKEN_HERE with your actual token.

Check sombra.so/mcp for the latest connection instructions.

Sombra supports OAuth 2.1 for interactive clients and Bearer tokens for headless agents.

Also works with#

Sombra is a standard MCP server (Streamable HTTP transport) that works with any MCP client:

• Claude Desktop / Claude.ai — add as a connector in Settings > Connectors
• Claude Code — claude mcp add --transport http sombra https://sombra.so/mcp
• ChatGPT — add as an app in Settings > Apps (Developer mode)
• Cursor, VS Code, Windsurf — add as an MCP server in your MCP config

Full setup instructions: sombra.so/mcp

Available tools (19)#

Save & Create#

ToolPurposesave_urlSave a web page by URL. Re-fetches and updates if previously savedcreate_artifactCreate a new note artifact with markdown content

Read & Browse#

ToolPurposebrowse_collectionsBrowse your collections returning title, id, and descriptionread_collectionRead the full content from a specific collectionread_collection_contextRead the distilled context/summary from a collectionfetch_artifactFetch a single artifact by IDbrowse_notesBrowse all note artifacts, sorted by most recent firstsearch_artifactsSearch saved artifacts by title, content, or URLlist_archivedList all archived (deleted) artifacts

Update & Organise#

ToolPurposeupdate_artifactUpdate an artifact's metadata or content (notes only)move_artifactMove a single artifact to a different collectionbulk_move_artifactsMove multiple artifacts to a different collection at onceremove_from_collectionRemove an artifact from a collection (back to unsorted)create_collectionCreate a new collectionupdate_collectionUpdate a collection's name, description, or iconwrite_collection_contextWrite distilled context/summary to a collection. Preserves all code verbatim

Delete & Restore#

ToolPurposedelete_artifactDelete an artifact (archives it, content is preserved)delete_collectionDelete a collection (archives it, artifacts are preserved)restore_artifactRestore a previously deleted artifact

Deletes are never permanent. All deleted items can be recovered.

Prompts (2)#

PromptPurposedistill_collectionGenerate a context-engineered distillation prompt for a collectiondistill_technical_collectionSpecialised distillation for collections with heavy technical content

Resources#

Each collection is exposed as an MCP resource with the URI scheme sombra://project/{id}.

Recommended workflows#

Research before coding#

Before starting a task, have your agent read the relevant collection context:

text

Read the distilled context from my "Pedestal Migration" collection,
then implement the HTTP handler changes based on those docs.

The agent gets dense, accurate context instead of relying on training data.

Build a research collection#

Ask your agent to research a topic and organise what it finds:

text

Research rate limiting approaches for Clojure web services.
Save the most useful pages to a new "Rate Limiting" collection,
then write a distilled context summarising the key approaches.

The research persists between sessions. Tomorrow's session picks up where today's left off.

Save discoveries during work#

When your agent finds something useful mid-task, it can save it:

text

That blog post about the transit encoding bug was helpful.
Save it to my "API Performance" collection.

Prevent agent drift#

Agent drift happens when your AI gradually diverges from correct behaviour because its context is stale or incomplete. Sombra prevents this by giving your agent access to verified, current documentation instead of months-old training data.

Before a coding session:

text

Read the context from my "API Integration" collection.
Use those docs as the source of truth for endpoint URLs
and auth flows. Do not rely on your training data.

Tips#

• Scope collections tightly. "EU e-invoicing integration" beats "Work stuff". One collection per project or task.
• Distil aggressively. Forty pages of raw docs in context is noise. Five hundred tokens of distilled context with code examples verbatim is signal.
• Let context compound. Have your agent save useful discoveries back to the collection. Each session makes the next one better.
• Use context before code. Always read the collection context before generating code against external APIs or libraries.

Links#

• sombra.so — sign up and manage your library
• MCP setup docs — setup for all clients
• What is Sombra? — full product overview
• Why context engineering matters — the problem Sombra solves