Moyai

A self-hosted, multi-harness background coding agent from BerriAI, the company behind the LiteLLM gateway. Assign it a task from a browser or Slack; it edits code, runs tests, and opens a pull request, with mid-task corrections and resumable sessions (saved files and conversation history). Routed here 2026-10-07 via Telegram.

Harness-as-a-setting

Each session defaults to the Claude Agent SDK with prompt caching on, but can switch per session to Hermes, Codex, OpenCode, Deep Agents, or Tool Loop — all sharing the same isolated workspace, tools and permissions. Model choice is the same kind of knob: 100+ providers through LiteLLM (OpenAI, Anthropic, Google, Bedrock, Mistral, DeepSeek, xAI, …), changeable mid-conversation, with spend tracked per teammate. Harness and model are both treated as configuration, not identity — the strongest statement yet in this corpus that a coding agent's product is the orchestration shell around the loop, not any one loop implementation.

A sixth shape of multi-harness ownership

nemoclaw (NVIDIA, runtime), Scout (Microsoft, identity), kimi-claw (Moonshot, hosting + model) and personal-jarvis (community, UI convention) each gate multi-harness operation by whatever they already control. Moyai adds a fifth owner-type — a gateway company — and gates by neither a sandbox policy nor an identity system but by credential isolation: provider API keys live on the server and the sandbox that runs the agent's code never sees them. That is a narrower guarantee than NemoClaw's routed-inference control plane (which can refuse a model call outright) — Moyai's sandbox simply has nothing to steal, rather than being policed.

The cost claim, against this spoke's existing Devin entry

BerriAI states it moved its own internal agent from devin to Moyai and cut costs by about 79% — $101,872 for Devin over 31 days against roughly $21,700 for Moyai over the same period. This is the first head-to-head cost figure between two products in the spoke's autonomous-coding- agent cluster, and it is entirely self-reported: one party's internal usage, on tasks it chose, 31 days, no stated methodology for what counted as a comparable task on either side. Devin's own distinguishing claim is duration (8+ hours unattended); nothing here says whether Moyai's tasks ran as long or did the same kind of work, so the comparison may be pricing two different jobs.

What is open

No independent usage data, no published license visible on the repository page despite a launch post titled "Moyai is now open source," and the default deploy target (Modal) bills compute that keeps running until the app and sandboxes are stopped from the Modal dashboard — a cost footgun the docs name but do not default away from. 94 stars, 12 forks, 223 commits at ingest.

Related

devin · nemoclaw · microsoft-scout · kimi-claw · personal-jarvis · agentic-coding-harness · agent-orchestration · ai-gateway