Alma: Independent Developer's AI Agent Harness Hits Global Top 3, Cuts Costs 16%

AI infrastructure company Runta recently published its FrontierHarness v1.1 benchmark, evaluating 21 agent harness configurations including Claude Code, Codex CLI, and OpenCode on task completion ability and cost. On the November 11 leaderboard, Codex CLI placed first, Pi second, and Alma — a project by independent developer yetone (author of the Magpie open-source project) — took third place.

In the benchmark, Alma completed each successful task at an average cost of $0.80, compared to $0.94 for Claude Code, a 16% reduction. Yetone, nicknamed "Ye Tiandi" by the community, has stated he dislikes such personality-cult labels.

Core Design Focus

Alma provides standard agent capabilities — file read/write, terminal execution, web search, Skills, MCP, and multi-model support — but concentrates on three areas: lowering model invocation cost, enabling cross-session long-term memory, and simplifying multi-agent management for a single user.

1. Separate Tool Model Configuration

Alma allows a dedicated "Tool Model" for tool calls, avoiding the expense of routing every simple operation through the primary (often more costly) model. This design reduces token overhead for routine actions.

2. Embedding-Based Long-Term Memory

Unlike context-window-only approaches, Alma extracts key information from historical conversations, converts it to semantic vectors, and stores it. When a new task starts, Alma rewrites the query via AI, retrieves semantically similar memories using a similarity threshold, and injects the results as context for the main model.

3. Herdr Integration for Multi-Agent Orchestration

Alma recently integrated Herdr, bringing Claude Code, Codex, Pi, and other coding agents into a single chat interface. Users can assign frontend work to Claude Code, backend to Codex, and code review to Pi without juggling multiple terminal windows. Sessions are auto-categorized as Working, Done, or Needs You; the Needs You state surfaces only when an agent requires human confirmation, letting developers prioritize interventions.

4. Mobile Remote Access via Alma Relay

Alma extends multi-agent management to iOS and Android. After launching agents on a desktop, users can monitor progress, answer prompts, and approve commands from a phone. Mac users also get a remote-desktop view with click, scroll, and input support. To traverse NAT/firewall boundaries, yetone open-sourced Alma Relay, a relay service deployable via Docker or Cloudflare Workers. The relay uses end-to-end encryption by default, so the server only forwards ciphertext.

Hands-On Walkthrough

Alma runs on Windows, macOS, Linux, iOS, and Android. After installing and configuring model APIs (optionally via Magpie), the UI presents projects/sessions on the left, the agent conversation in the center, and an integrated tool panel (files, diffs, commits, terminal, browser, tools) on the right. Tool invocation can be fully automatic or manually toggled per capability (web scrape, search, glob, ripgrep, read, patch, write, shell, delegate).

The settings include a professional agent team with preset roles — Designer, Product Manager, Developer, Researcher, Operator, Planner — and support recursive delegation among child agents. A Team feature lets multiple Alma instances join a shared workspace via connection addresses.

Overall, Alma does not reinvent basic file/terminal tools but integrates them with cost-aware model routing, persistent semantic memory, unified multi-agent UX, and secure mobile remote control, lowering the barrier to running concurrent AI coding agents.