MythosClaw / the autonomous agent platform
Your machine,run by aswarm of agents.
CatAgent is a local-first, Mac-native AI agent swarm. Twelve specialist agents and a 13-model mesh in one binary that runs on your hardware and bills through your own Claude subscription.
Not a chatbot.
A command center.
Cloud agents run someone else's software on someone else's servers and send your code along for the ride. CatAgent inverts that. It compiles to one Go binary, runs natively on Apple Silicon, and orchestrates a fleet of twelve specialized agents inside a single process. Each has its own persona, scrollback, and live stats.
Twelve named specialist agents in one binary, not threads or workflow nodes.
A single loopback gateway aggregating 13 models across Claude, Ollama, and OpenRouter.
First-party OAuth billing through your existing Claude subscription.
Local SQLite and JSONL memory that never leaves your machine.
A live terminal cockpit for real-time fleet observability.
/ THE INNERVOICE PIPELINE
One turn, five stages, one auditable trail.
Every request is enriched before it reaches the primary model. Each stage emits a broker event you can watch live in the cockpit.
Receive
Intent classification. Every turn is typed as question, code, command, research, trade, or chat before anything else runs.
Analyze
Sub-task decomposition with risk flags and delegation hints. The request is broken down and routed.
Trio
Parallel multi-model brainstorm. Four local Ollama models fan out at once, each with its own timeout, sorted by latency. Consensus at zero token cost.
Synth
XML system-message assembly. The Trio opinions are folded into the prompt so the primary model answers with consensus already in hand.
Mythos
Streamed Opus-class response. The enriched turn reaches the primary model and streams back, while every stage above is recorded to the broker.
Everything a swarm needs, in one binary.
Multi-agent swarm
12 named specialist agents, each with independent 500-entry scrollback, status lifecycle, and per-slot stats for turns, tokens, and cost.
Local gateway
Loopback-only OpenAI-compatible Go gateway aggregating Claude, Ollama, and OpenRouter behind one URL, with SSE streaming and a 13-model fallback chain.
InnerVoice reasoning
A transparent 5-stage enrichment pipeline that classifies, decomposes, brainstorms, synthesizes, and streams every turn.
Trio consensus
Parallel fan-out across four local models per turn, with timeouts and partial-failure handling, assembled into the prompt before the primary answers.
Persistent local memory
SQLite FTS and JSONL stores, multi-source recall, and salience-gated promotion to long-term memory. Stays on your disk, always.
Turn-chain audit
Every assistant turn sealed into a SHA3-256 chained, append-only block with a genesis anchor and full-session verification.
Live cockpit TUI
Bubble Tea v2 dashboard with virtual scroll, ANSI diff rendering, 20 sidebar sections, sparklines, and frame-timing observability.
Browser companion
LLM-driven browser automation pointed at the local gateway, with native actions for reading state and persisting observations.
Telegram and voice
Remote command over a long-poll Telegram bot with vision passthrough, plus local offline TTS. Drive the swarm from anywhere.
17 themes, OKLCH color
A true Oklab spectral rainbow and a neon hacker default, switchable at runtime.
Secret scrubber
A stdlib regex pipeline that redacts API keys, tokens, private keys, and PII at the SDK boundary before anything touches disk.
One-line install
Curl-pipe bash that builds the Go binaries and links the CLI. Apple Silicon native, no Docker or virtualenv.
The bundle nobody else ships.
Plenty of products do one of these. None do all of them in a single local binary.
Local-first, Mac-native, one binary
Runs on Apple Silicon with no Docker, no Python env, no npm. Devin needs cloud VMs, OpenHands needs Docker, CrewAI and LangGraph need Python environments.
First-party OAuth billing
Routes through your existing Claude subscription via Keychain OAuth, so interactive swarm runs cost near-zero marginal dollars. No separate API key, no per-token meter.
A real swarm, not workflow nodes
Twelve named specialist agents with personas, scrollback, and a shared 18-topic message bus, all in one process.
Transparent reasoning
The 5-stage InnerVoice pipeline pre-processes every turn and emits auditable events. No other terminal client exposes its reasoning chain this way.
Multi-model consensus as a primitive
The Trio brainstorm orchestrates multiple models before the primary answers. Most competitors lock to a single model.
Live cockpit observability
A real-time interactive terminal dashboard embedded in the agent binary, not a read-only trace viewer.
Never exfiltrates
The Castle Rule is enforced at the code level. No telemetry, no vendor logging, no data leaving the machine.
Forensic-grade audit
A SHA3-256 turn chain gives you a tamper-evident record that SaaS agents do not have.
One animal.
Twelve minds.
A coordinated swarm that thinks, builds, and verifies in parallel, and lives entirely on your machine.
/ HOW IT COMPARES
CatAgent vs. the field.
Where it runs is the difference everything else follows from.
| Dimension | CatAgent | Others |
|---|---|---|
| Where it runs | On your Mac, natively on Apple Silicon, as a single Go binary. | Operator, Devin, and Claude Managed Agents run entirely in vendor cloud. OpenHands and LocalAGI need Docker. CrewAI and LangGraph run in Python environments you provision. |
| Setup | One curl-pipe command builds the binaries and links the CLI. No Docker, no virtualenv, no npm. | Cloud products need accounts and billing. OpenHands needs a Docker pull. CrewAI and LangGraph need pip and Python. |
| Memory | Persistent local SQLite FTS and JSONL with salience-gated promotion. Never leaves the machine. | Cloud agents store state on vendor servers. AutoGen is in-RAM only. LocalAGI memory is largely ephemeral. |
| Observability | Live interactive cockpit TUI embedded in the binary. 12 agent slots, InnerVoice strip, broker traffic, cost, frame timing. | Most ship dashboards in a separate cloud console or read-only trace viewers. |
| Multi-model | 13-model fallback mesh with a named Trio consensus across multiple models before the primary answers. | Operator and Devin lock to their own models. Most frameworks offer no built-in multi-model consensus. |
| Browser, voice, Telegram | All three as a unified control plane. | Operator is browser-only, Devin is chat-ops only, browser-use is browser-only. None combine all three. |
| Privacy | Castle Rule enforced in code, never exfiltrates a file, loopback-only, no telemetry, secrets scrubbed. | Cloud agents send code and context to vendor servers by definition. |
| Cost and billing | Runs on your existing Claude subscription via first-party OAuth, near-zero marginal cost. | Operator around 200 USD per month, Devin around 70 to 220+ USD per month, frameworks bill per-token via your API keys. |
| Extensibility | Any OpenAI-compatible tool plugs into the local gateway with no config change. | Frameworks need code and infra, cloud products only extend within vendor SDKs. |
On your Mac, natively on Apple Silicon, as a single Go binary.
Operator, Devin, and Claude Managed Agents run entirely in vendor cloud. OpenHands and LocalAGI need Docker. CrewAI and LangGraph run in Python environments you provision.
One curl-pipe command builds the binaries and links the CLI. No Docker, no virtualenv, no npm.
Cloud products need accounts and billing. OpenHands needs a Docker pull. CrewAI and LangGraph need pip and Python.
Persistent local SQLite FTS and JSONL with salience-gated promotion. Never leaves the machine.
Cloud agents store state on vendor servers. AutoGen is in-RAM only. LocalAGI memory is largely ephemeral.
Live interactive cockpit TUI embedded in the binary. 12 agent slots, InnerVoice strip, broker traffic, cost, frame timing.
Most ship dashboards in a separate cloud console or read-only trace viewers.
13-model fallback mesh with a named Trio consensus across multiple models before the primary answers.
Operator and Devin lock to their own models. Most frameworks offer no built-in multi-model consensus.
All three as a unified control plane.
Operator is browser-only, Devin is chat-ops only, browser-use is browser-only. None combine all three.
Castle Rule enforced in code, never exfiltrates a file, loopback-only, no telemetry, secrets scrubbed.
Cloud agents send code and context to vendor servers by definition.
Runs on your existing Claude subscription via first-party OAuth, near-zero marginal cost.
Operator around 200 USD per month, Devin around 70 to 220+ USD per month, frameworks bill per-token via your API keys.
Any OpenAI-compatible tool plugs into the local gateway with no config change.
Frameworks need code and infra, cloud products only extend within vendor SDKs.
Compared against: OpenAI Operator, Devin, Claude Managed Agents, CrewAI, LangGraph, OpenHands, Pheron Agent, LocalAGI. Pricing estimates from public sources as of mid-2025.
Watch the swarm come online.
/ INSTALL - ONE LINE
The whole swarm on your Mac. One command.
Bootstraps gateway, CatAgent CLI and the cockpit TUI. macOS · prereq-checked · no config. Inspect the script before you run it - always.
+ MYTHOSCLAW CHROME EXTENSION →/ SECURITY & PRIVACY
Local-first isn't a setting. It's the architecture.
CatAgent is built so your data has nowhere to go but your own disk. The gateway and the OAuth bridge both refuse to bind anything but loopback, so they have no network surface to leak through. There is no telemetry, no vendor logging, and the Castle Rule is enforced at the code level. The Mac never exfiltrates a file.
Every secret is scrubbed at the SDK boundary before it can reach memory, logs, or storage. Anthropic, OpenAI, GitHub, AWS, GCP, Discord, and Telegram tokens, SSH and Solana keys, and PII all get replaced with opaque markers. Every assistant turn is sealed into a SHA3-256 chained audit block anchored to a genesis hash, so the whole session can be verified for tampering.
Loopback-only gateway and bridge
No network surface, enforced at startup.
Castle Rule
Never exfiltrates a file, enforced in code, zero telemetry.
Secret scrubber
Redacts keys, tokens, private keys, and PII before disk.
SHA3-256 turn-chain
Append-only, tamper-evident, fully verifiable.
First-party Keychain OAuth
No raw API keys in the hot path.
Stays on your machine
Memory, audit, and config never leave.
Questions, answered.
One command. Run the curl-pipe installer on Apple Silicon and it checks prerequisites, clones the repo, builds the Go gateway and cockpit binaries, syncs dependencies, and links the catagent CLI to your path. No Docker, no Python virtualenv. Launch the cockpit and the twelve-agent swarm comes online.
macOS on Apple Silicon. The installer auto-checks for git, bun, go, and uv, then builds everything from source. The idle gateway runs in roughly 10 to 20 MB of RAM, light enough to leave running all day.
A 13-model fallback mesh. By provider prefix it routes to local Ollama models, to OpenRouter when you opt in, or through the OAuth bridge to Claude (Opus, Sonnet, Haiku). The Trio stage fans out across four local models in parallel, and cost-gated switches pick the right model per turn automatically.
Yes. CatAgent is local-first by architecture, not by toggle. The gateway and OAuth bridge bind to loopback only and have no network surface. The Castle Rule, never exfiltrate a file, is enforced in code. No telemetry, no vendor logging. Memory, audit chains, and config all stay on your disk.
Near-zero marginal cost for interactive use. Requests route through your existing first-party Claude subscription via Keychain OAuth, so you do not pay per token for interactive swarm runs. Local Ollama models in the Trio stage cost nothing. OpenRouter is optional.
CatAgent ships as source you build locally. The one-line installer clones the repository and compiles the Go binaries on your own machine, so you can read exactly what runs. Check the GitHub repository for the current license.
macOS on Apple Silicon is the target and the installer is explicitly macOS-only. The binary is compiled natively for Apple Silicon. Other platforms are not currently supported.
Twelve named specialist agents run inside one binary, each with its own scrollback and live stats, communicating over an 18-topic message bus. Every turn passes through the 5-stage InnerVoice pipeline (Receive, Analyze, Trio, Synth, Mythos). You watch it live in the cockpit TUI.
Run the swarm on your own machine.
CatAgent is local-first, Mac-native, and yours. One line installs it. Your Claude subscription powers it. Nothing leaves your hardware.
curl -fsSL catagent.org/install.sh | bash