# Just the DAEMON.

CrabTalk is an open-source agent daemon written in Rust. Five megabytes, zero children, nothing hidden — just the plumbing every agent app rebuilds, written once and shared.

Install · one line

`curl -sSL https://crabtalk.ai/install | sh`

00 · Install

Fig. 00

**Fig. 00 · Architecture.** A cross-section of the daemon. One long-lived process in the centre; clients (CLI, bots, scheduled jobs) attach from outside; the model lives across a clean boundary in a separate project called crabllm.

> “Attach your apps, bring whatever tools you like.”

[Chapter 01](/content/oss/crabtalk#chapter-01/index.html)

## Attach, and speak.

One command launches the daemon and drops you into a session. No server to provision, no config to author, no dashboard to sign into. The binary is five megabytes; it holds nothing back.

Every message you send is visible in the stream. Every tool call, every event, every byte. If you want to build your own client — a Telegram bot, a scheduled cron, a dashboard — attach the same way this terminal does.

CONNECTION · oneattached

$ crabtalk

...daemon started

\> what can you help me with?

I can help you write code, search the web,

manage files, and remember context

across topics.

01 · Live

**Fig. 01 · An connection.** The terminal client attaches to the daemon on startup. Type a message; the reply streams back over the same connection. Every other client — bot, cron, dashboard — does exactly this.

[Chapter 02](/content/oss/crabtalk#chapter-02/index.html)

## Five pillars

Four guarantees the daemon makes, drawn out on this sheet. Crashes don’t propagate. Topics route themselves. Memory stays precise. Coordination scales. Each one a plate; each plate a sentence; the sheet reads in any order you like.

01 / 05

In most AI systems, one tool crashes and the whole chat crashes with it.

CrabTalk runs every tool, every command, every extension in its own space. **One breaks; the rest keep going.**

**Fig. 01.** Broken parts, stable core. A central body holds; satellites move on their own orbits, breaking apart and re-forming without disturbing the centre.

02 / 05

**Fig. 02.** Topic lanes. Messages travel along their own subject channels — work in the work lane, home in the home lane — without per-window juggling.

Other AI makes you open a new chat per subject.

CrabTalk sorts by topic itself — work stays in the work channel, home switches itself, **no window-juggling**.

03 / 05

67DD5C66F3A8861EA635D2CF0E49E0ED488AA790A54CF3C9CA01D068B718B752FD246F987B04B9EE329E5DCCF8219F9F586EC057D94052D7ECFDDB3229CE2175E0FD5F28BFDE756148DEFF6CA944B2679F31D43AB9A0DF59099E0DB751EDC88228460EE1B73D6C1204BB7BA3C0D11AC8FE805EBA097B679C12B1B5C72F080EB9E57C0891F3F512D7C4F9C8239408B331B321086F81A0FCD64C4178D6E2302F8FB6A773B10D4542D396D5F9D253FB710E7389F568

**Fig. 03.** Indexed retrieval. A bound archive built on the same internals engineers trust for their databases — fast lookup that doesn’t slow as it grows.

AI memory usually forgets fast, or slows to a crawl as it grows.

We ported the `internals of SQLite` to the recall layer — **accurate, fast to query, compact on disk**.

04 / 05

**Fig. 04.** At scale. Hundreds of agents in coordinated motion — a conductor facing a full orchestra, each section playing its own line on cue.

Other systems may cap around `10` concurrent agents.

CrabTalk is designed for coordinateing `10,000`. **A conductor facing a full orchestra**, not a small ensemble.

05 / 05

**Fig. 05.** One conversation, two voices. A guest agent speaks directly into the shared history — tagged, not delegated — then the primary resumes on the next turn.

Sometimes a second agent should speak into an ongoing conversation — a reviewer, a translator, an expert.

CrabTalk lets a guest read the shared history, add a tagged message, and step back. **The primary picks up where it left off.**

06 · End

## Just the daemon.

5 MB · zero children

[Read the spec](https://crabtalk.github.io/crabtalk/) [github.com/crabtalk](https://github.com/crabtalk/crabtalk)