Coding Harnesses

Every AI coding agent installed across the fleet — how they differ, what models they reach, and where the free quota lives.

Updated: 2026-06-05 Machines: hcm · fue · muc · sin Maintained in agentic-coding/tools/ai-inventory/harnesses.html

⚠️ June 18, 2026 — Gemini CLI shutdown

Google stops serving Gemini Code Assist + Gemini CLI for individual/Pro/Ultra tiers on 2026-06-18. Everything moves to Antigravity + Antigravity CLI. Our gemini --yolo alias and ai-launcher option 11 break that day — migrate before then.

Where the free quota lives

Amp Free (fue/muc) — free daily allowance; deep mode runs GPT-5.3-Codex. Recently reshaped: ads removed, allowance paused for inactive users — log in to re-verify.
Droid / Factory (all) — included Factory quota, default gpt-5.2; ran GPT-5.3-Codex to tbench rank 6.
Antigravity (hcm, app + agy CLI) — free model picker incl. Claude Opus 4.6 (Thinking), Claude Sonnet 4.6 (Thinking), Gemini 3.1 Pro (Low/High), Gemini 3.5 Flash (Low/Med/High), GPT-OSS 120B. Free Opus 4.6 — effectively replaces the lost Copilot premium lane. Google-account sign-in (agy to log in), weekly limits, unlimited tab completions.
Qwen Code (hcm) — 2,000 free requests/day via Qwen OAuth.
Copilotpaid lane GONE. The wetheflywheel account (which held the 1,500 premium-req/mo sub) was DELETED 2026-06-20. Only thomas-prommer remains, on Copilot Free (GPT-5 mini / Haiku, 0 premium interactions, ~200 chat + 2,000 completions/mo, resets monthly). No GPT-5.3-Codex / Sonnet / Gemini-Pro lane until a new sub is set up. The copilot alias no longer pins a token.
Devin (hcm) — 10 free ACUs trial on Devin Cloud; local CLI itself is free to run.
Gemini CLI (hcm) — free tier until June 18 shutdown.
OpenRouter kimi-k2:free — zero-cost Kimi via aid-kimi-free.

Harness comparison

Harness Type Machines Version Auth / billing Models What makes it different
Claude Code Terminal agent hcm fue muc sin 2.1.163 2× Claude Max subs (Max 20x + Max 5x, cs switcher) Opus 4.8 (1M ctx), Sonnet 4.6, Haiku 4.5 Primary harness. Deepest scaffolding: skills, hooks, agent teams, workflows, cron. Everything else is overflow/experiment.
Codex CLI Terminal agent hcm muc 0.133.0 OpenAI API key (usage-billed; no sub) GPT-5.4, GPT-5.3-Codex OpenAI's own agent. Aliases: codex (full-auto), ccodex (danger-full-access + web search). Codex Plus sub ($20) evaluated, not bought.
OpenCode Terminal agent (multi-provider) hcm 1.3.17 Per-provider: OpenRouter (default), Copilot OAuth, Moonshot Default z-ai/glm-5.1 / z-ai/glm-5-turbo via OpenRouter; via aliases: Copilot Sonnet 4.6 (oc-gh), Kimi K2.5 (oc-kimi), GPT-5.4 (oc-gpt54), Grok 4 (oc-grok) The model-router: one harness, every key. How we burned the 300 Copilot premium requests on Opus 4.6. Cheap overflow lane = GLM-5.1 (now metered via OpenRouter — Z.ai coding plan expired 2026-06-07).
Copilot CLI Terminal agent hcm 0.0.411 Copilot Free onlywetheflywheel (paid sub) deleted 2026-06-20; thomas-prommer is the sole account, 0 premium req/mo Free tier: GPT-5 mini, Haiku. No GPT-5.3-Codex / Sonnet / Gemini-Pro / Opus (all were premium-only) Premium lane dead with the account. copilot alias no longer pins a token — runs on thomas-prommer Free. oc-gh shares the same (now free) pool.
Droid (Factory) Terminal agent hcm fue muc 0.74.0 Factory account — included free quota GPT-5.2 (default), GPT-5.3-Codex, Claude, Gemini Terminal-Bench 2.0 rank 6 (GPT-5.3-Codex). Shares our commands/agents via ~/.factory symlinks. droid --auto high.
Amp (Sourcegraph) Terminal agent + editor ext fue muc 2026-06-04 build Amp Free (free daily allowance) or credits Deep mode = GPT-5.3-Codex; Oracle reasoning = GPT-5.x The "free codex quota" harness. Auto-upgraded by upgrade-coding-agents.sh since 2025-12. Free allowance recently reshaped (paused for inactive users) — re-login to verify.
Devin CLI Terminal agent + cloud VM hcm 2026.5.6 Cognition account; 10 free ACUs (cloud), local CLI free Cognition's own model stack (Devin / SWE-family) Two modes: local terminal agent (codename "chisel") + cloud sessions that open PRs from a VM. Yolo: DEVIN_PERMISSION_MODE=dangerous (set globally). Tested against domain-ops#1.
Gemini CLI Terminal agent hcm 0.29.0 Google account — free tier Gemini 3 Pro / Flash EOL 2026-06-18. Replaced by Antigravity CLI. Aliases gemini (yolo), gemini-auto.
Antigravity (+agy) IDE (VSCode fork), multi-agent hcm IDE 1.21.9 / agy 1.107 Google account — free plan (weekly limits, 3×'d twice in May) Gemini 3 Pro, 3.5 Flash; unlimited tab completions IDE-first agent manager with browser integration; agy just launches the IDE. Used for detailed code reviews when other quota ran out.
Qwen Code Terminal agent (Gemini CLI fork) hcm 0.10.3 Qwen OAuth — 2,000 req/day free Qwen3-Coder family Biggest raw free allowance in the fleet. qwen --yolo in ai-launcher.
Aider Terminal pair-programmer hcm 0.86.2 BYO key (OpenRouter default) Default glm-5.1; aliases for Kimi, Sonnet, o3, o4-mini, kimi-k2:free Leanest overhead, git-native diffs, repo-map. Terminal-Bench pickup (Apr 2026) for cheap-model overflow work.
Goose (Block) Terminal agent + extensions hcm 1.29.1 BYO key — provider never configured Any (Anthropic, OpenAI, OpenRouter, Ollama…) Extension/MCP-heavy architecture, subagent "summon". Second Terminal-Bench pickup (Apr 2026); still unconfigured — finish or remove.
Hermes (Nous) Assistant gateway (channels + agent) hcm muc 0.13.0 Credential pool: Anthropic OAuth ✓, Copilot gh-token, MiniMax key muc default: MiniMax-M2.7; supports Nous Portal, Kimi, GLM, OpenRouter, Gemini, xAI Not a coding CLI — the OpenClaw successor: WhatsApp/Slack/Discord channels, kanban, cron, self-improving skills. See Hermes section below.
OpenClaw Assistant gateway hcm 2026.4.2 Legacy — migrated to Hermes (May 11). Archive with hermes claw migrate cleanup.

Hermes × Kimi — where that connection stands

Hermes (NousResearch hermes-agent) is our personal-assistant gateway, not a coding harness: it owns the WhatsApp/Slack/Discord/iMessage channels, a kanban, cron jobs, and self-improving skills. It migrated from OpenClaw on 2026-05-11 and runs on hcm + muc (muc is the live gateway).

Live on muc today

  • Default model: MiniMax-M2.7 (provider minimax, MINIMAX_API_KEY)
  • Credential pool: Anthropic OAuth (verified ✓), GitHub Copilot token (via gh), claude_code provider
  • Reasoning effort medium, 60 max turns, gateway timeouts tuned for channel use

The Kimi lane

  • Hermes ships first-class providers kimi-coding (KIMI_API_KEY) and kimi-coding-cn (KIMI_CN_API_KEY) — wired in config, ready to activate
  • We hold prepaid Moonshot credits; plan of record (Apr 2026): consume them first, then fall back to OpenRouter Kimi
  • KIMI_API_KEY + MOONSHOT_API_KEY re-added to BAO (secret/library/ai) on 2026-06-06 and verified live against api.moonshot.ai — all machines pick it up via the env.shared cron
  • Zero-cost fallback that works today: aid-kimi-free → OpenRouter kimi-k2:free

Model access matrix

Model familyReachable throughCost lane
Claude Opus 4.8 / Sonnet 4.6 / Haiku 4.5Claude Code (2 Max subs) · Hermes (Anthropic OAuth) · Antigravity free (Opus 4.6 + Sonnet 4.6, Thinking) — Copilot premium-req lane gone (wetheflywheel deleted 2026-06-20)Subscription + Antigravity free
GPT-5.4 / GPT-5.3-CodexCodex CLI (API key) · Amp deep mode (free) · Droid (Factory quota) · oc-gpt54 — Copilot CLI premium lane gone (wetheflywheel deleted)Metered via API; free via Amp/Droid
Gemini 3 Pro / 3.5 FlashAntigravity (free, weekly) · Gemini CLI (until 06-18) · Hermes (GOOGLE_API_KEY)Free tier
GLM-5.1 / GLM-5-turbo (Z.ai)OpenCode default · Aider default — both via OpenRouter · Hermes zai provider (API key)OpenRouter metered — the cheap overflow lane. Z.ai coding plan expired 2026-06-07 (unused since April), not renewed
Kimi K2.5 (Moonshot)oc-kimi · aid-kimi · Hermes kimi-coding — key re-export pending · kimi-k2:free via OpenRouterPrepaid credits (unconsumed) / free tier
MiniMax M2.7Hermes on muc (default)API key
Grok 4oc-grok (XAI_API_KEY) · Hermes xai providerAPI key
Qwen3-CoderQwen Code CLIFree (2,000/day)
DeepSeek + 100s moreOpenRouter key → Aider, OpenCode, Goose, HermesMetered, cheap
Local (LLaVA, Whisper, Qwen via Ollama)Ollama on hcm — /video skill etc.Free, offline

Copilot AI-credit pricing (HISTORICAL — paid picker retired 2026-06-20)

The wetheflywheel account holding this paid AI-credit plan was deleted 2026-06-20. None of these premium models are reachable now — Copilot is Free-tier only (GPT-5 mini / Haiku). Table kept for reference if a new sub is set up.

1 AI credit = $0.01. Prices per 1M tokens; "credits/request" assumes a typical agentic turn of ~20K input + 2K output, uncached. Source: GitHub models & pricing, June 2026 billing.

ModelInputCached inOutput~Credits/requestvs cheapest
GPT-5 mini$0.25$0.025$2.00~0.9
GPT-5.4 mini$0.75$0.075$4.50~2.42.7×
Claude Haiku 4.5$1.00$0.10$5.00~3.03.3×
Gemini 3.5 Flash$1.50$0.15$9.00~4.85.3×
GPT-5.3-Codex ← default for real work$1.75$0.175$14.00~6.3
Gemini 3.1 Pro (≤200K)$2.00$0.20$12.00~6.47.1×
GPT-5.4 (≤272K)$2.50$0.25$15.00~8.08.9×
Claude Sonnet 4.6 / 4.5$3.00$0.30$15.00~9.010×

Operational notes