Introduction
Anymo is a local-first desktop for running, watching, and steering agent swarms, with any model, under your rules. Build your self-driving codebase.
Anymo runs goal-driven agents on your machine. A Rust kernel executes every run as an append-only event log, a desktop shell shows the swarm live, and every tool call passes through one policy engine with exactly three permission modes and no bypass. It works with any OpenAI-compatible endpoint: local model servers, routers, proxies, or hosted APIs.
Anymo is currently v0.2.0-beta.1. APIs, schemas, and storage formats are not yet stable. See the beta release notes for what changed and what is known to be limited.
Get started
One command verifies your environment, scaffolds a demo workspace, and runs an offline mock demo. No API key, no network.
Quickstart
From a fresh clone to a running swarm in one command, then a tour of every major feature.
Desktop app
Chat, approvals, the live Hive view, goals, skills, Dreams, and automations, no commands required.
Connect a model
Point Anymo at Ollama, LM Studio, a router, a proxy, or any hosted OpenAI-compatible API.
Safety and permissions
Three permission modes, one policy gate for every tool call, and an audit trail by construction.
Self-learning
Runs distill into memories, skills, and a knowledge graph. You review everything before it sticks.
CLI reference
Every anymo verb, its flags, and one realistic example, from setup to swarms to the app-server.
Why Anymo
- Local-first and event-sourced. Runs, memory, skills, and keys live on your machine. Every run is an append-only event log you can replay step by step after a crash, a restart, or a month. The UI is a projection of the log, not the other way around.
- Approval discipline with no bypass. Permission modes are exactly
auto,manual, andplan. Even inauto, high-risk actions such as shell, browser, desktop control, and MCP tools always stop and ask. There is no "dangerously skip permissions" flag anywhere in the codebase. - Sealed secrets. In the desktop app, saved provider keys are encrypted on disk with a master secret held in the OS keychain. The headless CLI never persists a key at all. Either way, keys never appear in events, logs, artifacts, or the UI.
- Self-learning with human review. After a run, Anymo distills memory and skill proposals. Nothing is ever auto-approved: you review each proposal before it can influence future runs.
- Any model, with failover. A native streaming tool-call loop works against any OpenAI-compatible endpoint, plus an ordered provider fallback chain that takes over when the primary exhausts retries.
- Swarms that size themselves. The planner uses the smallest team that fits the goal and only scales toward the 300-agent cap when you explicitly ask for large-scale parallel work.
Explore the features
Each feature has a full guide: what it does, how to use it, and how it stays safe.
Chat and runs
Describe the work in plain language. The agent plans, acts with real tools, and shows everything it did, with Undo.
Swarm mode
A commander plus specialists on one goal, up to 300 agents, watched live in the Hive view with budgets and verification.
Approvals
Risky actions stop and ask, in every mode. Rare enough to stay meaningful, recorded forever.
Goals
Set a finish line and let the agent keep working until an independent evaluator agrees it is done.
Skills
Runs that worked become reusable slash commands and agent-callable tools, parameterized and review-first.
Dreams (memory)
The agent proposes what to remember; you approve or reject. Approved memory flows into future runs.
Automations
Scheduled agent work: nightly reviews, pollers, reports, with goal loops and no catch-up storms.
Plugins and connectors
Connect remote services over MCP with OAuth, and install bundles of skills and hooks.
Worktrees and isolation
Coding jobs run in isolated Git worktrees and land as reviewable patch sets. Your checkout stays yours.
Query Fabric
Deterministic local search over files, code, memory, skills, tools, and events, with cited evidence.
How it works
Anymo is an embedded Rust library wrapped in a thin desktop shell. The kernel owns everything important and has no UI dependency, so the entire runtime is testable headless.
| Layer | Technology | Role |
|---|---|---|
| Desktop shell | Tauri v2 + React + TypeScript | Hive graph, chat, approvals, goals, skills, Dreams, automations, settings |
| Hive Kernel | Rust, embedded library | Runs, scheduling, event log, providers, tools, policy, memory, skills |
| Protocol | Rust types with generated TypeScript | One source of truth for events, runs, agents, tools, skills |
| App-server | JSON-RPC over stdio, Unix socket, or WebSocket | Headless host for the same kernel |
| Workers | Optional pooled subprocesses | Specialized capabilities behind a typed, policy-gated local protocol |
The reasoning behind each of these decisions is recorded as architecture decision records; start with Architecture. For where every byte lives, see Storage and data; for how the agent finds things, see Query Fabric; for scripting and integrations, see the App-server API.
For agents and machines
This documentation is written for humans, agents, and tooling alike:
- Every page has a Copy page button that copies the page as clean Markdown, and a View as Markdown action for the raw source.
- Each page links to its source file in the repository, so agents can fetch the canonical version directly.
License and status
Anymo is open source under the Apache License 2.0. The project is in public beta; schemas and storage formats may change before a stable release.