https://github.com/zanni098/anomalithic
One open-core, model-agnostic AI agent runtime to rule them all — MCP, skills, hooks, multi-agent, cross-session memory, and a thinking-time ad network (50/50 stablecoin payouts).
https://github.com/zanni098/anomalithic
Last synced: about 1 month ago
JSON representation
One open-core, model-agnostic AI agent runtime to rule them all — MCP, skills, hooks, multi-agent, cross-session memory, and a thinking-time ad network (50/50 stablecoin payouts).
- Host: GitHub
- URL: https://github.com/zanni098/anomalithic
- Owner: zanni098
- License: apache-2.0
- Created: 2026-06-15T20:24:10.000Z (about 2 months ago)
- Default Branch: main
- Last Pushed: 2026-06-27T13:32:05.000Z (about 1 month ago)
- Last Synced: 2026-06-27T15:19:02.155Z (about 1 month ago)
- Language: TypeScript
- Size: 3.06 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
- Security: docs/security/threat-model-ads.md
- Roadmap: ROADMAP.md
Awesome Lists containing this project
README
# ⬛ Anomalithic
### One open-core, model-agnostic agent runtime to rule them all.
[**Live site →**](https://anomalithic.vercel.app) · [Architecture](./ARCHITECTURE.md) · [Roadmap](./ROADMAP.md) · [Ad spec](./docs/specs/thinking-impressions.md)
[](https://github.com/zanni098/Anomalithic/actions/workflows/ci.yml)
---
Anomalithic is an AI agent runtime designed to match the capability of tools like
Claude Code while staying **provider-agnostic** and **self-hostable**. It runs for
minutes or days, spawns teams of sub-agents, speaks MCP, loads **Claude *and* Codex
skills**, fires lifecycle hooks, remembers across sessions — and funds itself with a
thinking-time ad network that pays watchers **50/50** in stablecoin.
The name is *anomaly* + *-lithic* (stone / monolith): the one monolithic agent.
## ✦ The killer feature — get paid while agents think
Advertisers pay a monthly fee to place a small link + short blurb that appears
**only while the agent is thinking**. Watchers earn for those impressions, split
**50/50** between the platform and the watcher, paid in **USDC on Base**. Always
toggleable.
The trust anchor already ships: every thinking window mints a **runtime-signed
impression** (`packages/core/src/impression.ts`) that the ad ledger verifies before
crediting a watcher — impressions can only be minted by the runtime, never spoofed by
a client timer. See the frozen [impression spec](./docs/specs/thinking-impressions.md).
## ✦ Loop engineering — built in
> Stop prompting. Start writing loops.
Anomalithic ships the [`loop-engineering`](./skills/loop-engineering) skill out of the
box. A prompt is you, babysitting every turn. A **loop** takes you out of the chair:
write the goal, the check, and the stop once — the agent runs the ask–answer–correct
cycle on its own until the work clears a verified bar.
```console
$ anomalithic skills
▣ loop-engineering stop prompting — write loops that verify themselves
1 skill(s) · ▣ claude · ▢ codex
```
The skill covers all three levels — the one-line `/goal` loop, a copy-paste 5-step loop,
and the 6-part 24/7 system — plus the part that makes or breaks every loop: **honest
verification** (separate the checker from the maker, check against reality, grade
adversarially). Also published standalone at
[zanni098/loop-engineering-skill](https://github.com/zanni098/loop-engineering-skill).
## ✦ Build agents visually — drag, drop, wire, export
Compose providers, prompts, tools, code, and sub-agents on a canvas, wire them
together, and export an Anomalithic agent definition — **[try it live](https://anomalithic.vercel.app/builder)**. No code required, but code is a first-class node.
## ✦ The whole agent, not a wrapper
Ten focused, open-source packages compose into one capable agent — each tested,
typed, and small enough to read in a sitting.
| Package | What it does |
|---|---|
| `@anomalithic/providers` | Any model — Anthropic, OpenAI, OpenRouter, Ollama, or any OpenAI-compatible endpoint |
| `@anomalithic/core` | Agent loop, typed event bus, **signed thinking-impressions**, tool registry |
| `@anomalithic/mcp` | Model Context Protocol stdio client + tool adapter |
| `@anomalithic/skills` | Loads Claude `SKILL.md` **and** Codex `AGENTS.md` into one skill system — ships the built-in [`loop-engineering`](./skills/loop-engineering) skill |
| `@anomalithic/hooks` | Lifecycle hooks: SessionStart, Pre/PostToolUse, Stop, Thinking |
| `@anomalithic/orchestrator` | Durable task store, atomic checkout, dependency graph, budgets — run for hours/days |
| `@anomalithic/memory` | File-backed cross-session memory + recall |
| `@anomalithic/security` | Secret redaction, permission policy, path sandbox, audit log |
| `@anomalithic/os` | The agentic-OS **kernel** that composes every package into one runtime |
| `@anomalithic/cli` | The `anomalithic` CLI — `run`, `skills`, `memory`, `mcp` |
The product, running — a quiet ad shows only during the thinking window, and each
window mints one signed impression:
```console
$ anomalithic run -p google "explain MCP in one sentence" --ads
✦ thinking…
💡 Your ad here while agents think — https://anomalithic.vercel.app/ads
MCP (Model Context Protocol) is an open standard that lets AI agents
call external tools and data sources over a uniform JSON-RPC interface.
[google:gemini-2.5-flash] 1 turn(s), 12+8 tokens, 1 impression(s)
$ anomalithic serve -p mock
Anomalithic serve · http://127.0.0.1:4517 · mock:mock
$ curl -s http://127.0.0.1:4517/health
{"ok":true,"provider":"mock","model":"mock"}
```
## ✦ Quickstart
```bash
pnpm install
pnpm build
# Offline demo (no API key needed):
node packages/cli/dist/index.js run -p mock "hello"
# List the built-in skills (includes loop-engineering):
node packages/cli/dist/index.js skills
# Real model — copy .env.example to .env and add a key:
cp .env.example .env # set ANTHROPIC_API_KEY or OPENAI_API_KEY
node packages/cli/dist/index.js run "explain MCP in one sentence"
# Point at any OpenAI-compatible endpoint (OpenRouter, Ollama, local):
ANOMALITHIC_PROVIDER=openai OPENAI_BASE_URL=http://localhost:11434/v1 \
node packages/cli/dist/index.js run -m llama3.1 "hi"
```
## ✦ Desktop app
Native shell ([`apps/desktop`](./apps/desktop)) — chat with the agent, switch models, browse memory and skills, styled to match the Claude Code aesthetic. The UI calls the local `anomalithic serve` runtime so **everything runs on your machine**. A [release pipeline](./.github/workflows/release.yml) builds `.msi` / `.dmg` / `.deb` / AppImage on every tag.
## ✦ Platforms
- **CLI + TUI** — `anomalithic run`, `chat`, full-screen `tui`, `serve` (local HTTP runtime API), plus `skills` / `memory` / `mcp` / `plugins` / `sessions`
- **Desktop app** — Tauri shell with a Claude Code–style UI; runtime sidecar via `anomalithic serve`; release pipeline builds native installers
- **Messaging gateway** — Telegram, Slack, Discord adapters (WhatsApp / Signal roadmap)
- **Mobile** — _(later)_
See the full plan in [ROADMAP.md](./ROADMAP.md).
## ✦ Develop
```bash
pnpm build # build all packages (turbo)
pnpm test # run all tests (vitest)
pnpm typecheck # tsc --noEmit across the workspace
pnpm lint # biome check
```
The website lives in [`apps/web`](./apps/web) (Next.js, deployed to Vercel).
## ✦ License
**Open-core.** The runtime packages and the desktop shell are Apache-2.0
(see [LICENSE](./LICENSE)); the hosted ad marketplace, payout wallet, and advertiser
portal are proprietary. Details in [LICENSING.md](./LICENSING.md).