Name: OpenWalrus
Author: OpenWalrus

Question 1

What is OpenWalrus?

Accepted Answer

An open-source composable agent runtime written in Rust. One 8MB daemon — inference is builtin, memory, search, and gateways are community extensions. Run one binary, swap any piece.

Question 2

What models does it support?

Accepted Answer

OpenWalrus connects to OpenAI, Claude, DeepSeek, Grok, Qwen, Kimi, Ollama, and any OpenAI-compatible endpoint. Two API standards (OpenAI and Anthropic) cover all providers. Configuration is a single TOML block per provider.

Question 3

Is my data private?

Accepted Answer

Yes. Everything runs on your machine. No data leaves unless you send it to a model API. No telemetry, no cloud sync.

Question 4

Which platforms are supported?

Accepted Answer

macOS (Apple Silicon and Intel) and Linux (x86_64 and ARM). The install script auto-detects your platform.

Question 5

How much does it cost to run?

Accepted Answer

OpenWalrus itself is free and open source. Bring your own model provider — connect any cloud API or use Ollama for local models. No markup, no metering, no usage-based billing from walrus.

Question 6

How does agent memory work?

Accepted Answer

The memory extension provides graph-based memory backed by LanceDB with three tables: entities, relations, and journals. Six tools (remember, recall, relate, connections, compact, distill) let the agent build and traverse a knowledge graph. Identities and recent journals are auto-loaded into the system prompt at session start.

Question 7

What are extensions?

Accepted Answer

Extensions are the protocol that makes walrus composable. Inference is builtin; memory, search, and gateways are community extensions. All use the same protocol. Extensions declare capabilities at startup: Tools, BuildAgent, BeforeRun, Compact, EventObserver.

Question 8

What are skills?

Accepted Answer

Skills define agent behavior and are installable in one command from the hub. Three tiers: bundled (ship with walrus), managed (community hub), and workspace (your own). Extensions handle runtime-level integration; skills handle prompt-level behavior. Both are composable.

Question 9

Can agents delegate work to other agents?

Accepted Answer

Yes. The task registry lets agents spawn tasks, check status, and await results. Tasks have five states (queued, in_progress, blocked, finished, failed), concurrency limits, timeouts, and an approval flow. Run walrus ps to see live task state.

Agents without the bloat.

Install. Attach. Done.

One command to extend.

Swap any piece. Break nothing.

What you get. What you don't.

Frequently asked questions