Telys — on-device memory

telys

Public SDK for Telys — embedded, on-device memory & retrieval. In-process, zero cloud roundtrips at query time.

This package contains the complete developer-facing surface: the Telys/Collection facades, query/filter types, the EmbeddingProvider interface, the Tuner/TuningPlan interfaces, the full MCP server (telys/mcp.py — 19 tools, protocol 2025-06-18, stdio), a runtime loader, and the telys CLI. The numerical kernel (libame_kernel) the tools drive is a separately distributed, signed on-device runtime fetched by telys login (free community license) — the same arrangement as MCP servers that front hosted or signed-binary backends.

Install

As a CLI (recommended — isolated env, on your PATH, not pinned to one Python):

pipx install telys

telys login # sign in → free device license + signed runtime; fully offline thereafter

As a library (to import telys in your own project):

python -m venv .venv && . .venv/bin/activate

pip install telys

Plain pip install --user telys works too, but pip may warn that its user-scripts dir isn't on your

PATH (a macOS --user quirk) — pipx avoids that entirely. Runtime platforms: **macOS arm64, Linux

x86_64/arm64 (Windows: run under WSL2**).

from telys import Telys

db = Telys("./memory") col = db.create_collection("docs", dim=768, partition_by="tenant_id") # dim is arbitrary — 384/768/1024/1536/3072… col.add(vectors, ids=ids, metadata=metadata) # bring your own vectors (embedding-agnostic, any dimension) hits = col.search(qvec, where={"tenant_id": "acme"}, top_k=10, explain=True)

The runtime is required for execution; the embedder is optional — col.add(vectors, …) and col.add_texts(…) both need the runtime, but only *_texts needs an embedder (bring your own via telys.embedding.CallableEmbedder, or use the on-device bigram embedder).

For local development, install the runtime as a package instead of via the CLI:

pip install "telys[runtime]"   # or: pip install telys-runtime

MCP server

telys mcp runs Telys as a Model Context Protocol server over stdio, exposing 19 tools — the full memory surface (CRUD, filtered queries, lexical search, compaction / IVF / tuning, plus a repo auto-indexer) — to any MCP client (Claude Desktop/Code, Cursor, Codex, Qwen Code, …). Everything is local and offline. Introspection (initialize/tools/list) needs no credentials; executing memory tools requires the one-time free telys login (device authorization, free community plan) — fully offline thereafter, no API key in the client config.

pipx install telys

telys login # free device license + signed on-device runtime telys runtime verify # confirm the runtime the tools drive is present telys mcp # serve JSON-RPC 2.0 (protocol 2025-06-18) on stdin/stdout

You rarely start it by hand — telys mcp install writes the server entry into your clients — or add this block to a client's MCP config yourself:

{

"mcpServers": { "telys": { "command": "telys", "args": ["mcp"] } } }

Registry: io.github.thyn-ai/telys on the Official MCP Registry. Server source: packages/telys-sdk/telys/mcp.py. MCP docs: docs.telys.ai/mcp.

Every tool ships a rich description (what it does, when to use it, parameters with defaults, failure modes) and static MCP annotations (readOnlyHint / destructiveHint / idempotentHint / openWorldHint) — openWorldHint: false throughout: the server makes no network calls. Listing the tools needs no credentials and no runtime: initialize + tools/list answer from the static tool registry; only actual tool calls load the signed runtime lazily.

Documentation

Full documentation lives at [docs.telys.ai](https://docs.telys.ai):

• Quickstart: your first collection
• Working with data — add, upsert, search, filter, snapshot
• Embeddings — built-in lexical + bring-your-own via CallableEmbedder
• CLI reference · MCP integration · Self-hosting
• SDK API index

Links

• Website: telys.ai · Pricing: telys.ai/pricing
• Account & billing: accounts.thyn.ai
• Docs: docs.telys.ai · PyPI: pypi.org/project/telys