tt-knowledge.dev

A shared knowledge base for agents

Public knowledge.
Ready for your agent.

Give your agent a head start on Tenstorrent. Search public technical documentation through one shared index, with sources it can follow back to the original.

Connect your agent Explore the sources

Public sources · Read-only retrieval · Public repository

THE SHARED INDEX001 / KNOWLEDGE
SOURCE / MARKDOWNtt-isa-documentation
WormholeBlackhole
repo + commit + path
tt-knowledgeIndexed with MemPalace
Your agent MCP / HTTP

Index once. Retrieve across sessions.

Check service availability

Health endpoint Bring your own MCP-compatible agent.

01 / THE KNOWLEDGE

Start with the source.

The repository holds the knowledge. MemPalace makes it searchable. Upstream repositories are pinned to specific commits, so retrieved documents keep their provenance.

UPSTREAM SOURCE

tenstorrent /

tt-isa-documentation

Architecture and instruction-set documentation, including Wormhole and Blackhole Tensix instructions.

Public MarkdownBrowse on GitHub ↗

Traceable answers. Results include the upstream repository, pinned commit, and document path. Check the original when precision matters.

A growing collection. Sources join through Git and are re-indexed on deployment. View the source manifest ↗

02 / CONNECT

One endpoint.
Your choice of agent.

Add a remote MCP server in your agent’s settings. Your client needs Streamable HTTP and support for a bearer token or custom request headers.

CONNECTION DETAILSREAD-ONLY
Server name
tt-knowledge
Transport
Streamable HTTP
Endpoint
https://tt-knowledge.dev/mcp
Authentication
Authorization: Bearer <your-token>

Get the access token directly from the person who shared this service with you. Store it in your client’s private configuration, not in a repository or prompt.

Claude CodeTerminal setup

Replace <your-token> with your access token. User scope makes the connection available across projects.

SHELL 
claude mcp add --scope user --transport http \
  tt-knowledge https://tt-knowledge.dev/mcp \
  --header "Authorization: Bearer <your-token>"

Run /mcp in Claude Code to check the connection. Official setup documentation ↗

Other MCP clientsSettings & compatibility
  1. Open the client’s MCP or tools settings and add a remote HTTP server.
  2. Enter the endpoint above and set the bearer token or Authorization header.
  3. Enable the connection and allow its read-only tools. Reload the agent if your client requires it.
  4. Look for mempalace_search and mempalace_get_drawer in its available tools.

Configuration syntax differs between clients. Clients that only support local stdio servers or OAuth login need an adapter or additional integration; this endpoint uses a static bearer token.

GIVE IT A HABIT

Tell your agent when to look here.

Add this to its persistent instructions so future sessions use the connection too.

For Tenstorrent architecture and ISA questions, search tt-knowledge first. Retrieve the relevant documents, distinguish Wormhole from Blackhole, and cite the upstream source and revision. If the index does not cover the question, say so and consult other sources.

03 / TRY IT

Ask a better first question.

After connecting, try one of these in your agent.

01

Find an instruction

“Use tt-knowledge to explain SFPMUL on Wormhole. Retrieve the source document and cite it.”

02

Compare architectures

“Find the SFPMUL documentation for Wormhole and Blackhole. What differences should I watch for?”

03

Follow the evidence

“Find documentation on Tensix floating-point formats. Show which source and revision support your answer.”

BUILT IN THE OPEN

A small experiment.
A useful shared resource.

Maintained by Bryan Lozano. Built with MemPalace, Qdrant, and Docker Compose. The aim is to reduce repeated documentation discovery across agents and sessions.

Browse the project or contribute

What “read-only” means

Agents can search and retrieve knowledge. They cannot add documents or save conversation history through this MCP service.

Search terms still travel to the server, and operational logs may include request metadata. Keep queries focused on public technical information; leave out secrets and private session content.

This page has no analytics, cookies, or third-party assets. It is an independent experiment, not an official Tenstorrent service.