Do you agree with the draft ranking of CLI agent harnesses for T3 Code + fx on an Ubuntu VPS (2 vCPU, 3.8 GB RAM, ~2.3 GB free) using Nebius Token Factory with DeepSeek V4.1 Flash? If not, adjust and rerank the recommendations.
Partial disagreement, and a rerank. Eleven harnesses are rescored on stack fit for the stated weights — these are judgment scores, not measured benchmarks — backed by 180 product-evidence rows, 145 repository-fact rows, 15 focused fact checks, and 4 Nebius protocol findings, all scanned 2026-11-02. Mainplane and Vercel eve are out-of-category, and dsh-superagent remains an add-on, not a harness. Live project facts can change.
Codex CLI for T3-native work against Nebius; omp as the extensible second harness. Do not add both pi and omp — omp is a pi fork and their roles overlap (github.com).
Add dsh only if the DeepSeek-specific plugin/superagent path is a real use case (github.com). Kilo is the next optional open generalist (github.com).
Droid now documents BYOK and headless exec yet adds little beside fx/omp/Codex (docs.factory.ai). Claude Code speaks Anthropic Messages, not generic OpenAI (docs.anthropic.com).
| # | Harness | Score | License | Stars | BYOK | Custom base URL | ACP | MCP | Subagents | Evidence rows | Source |
|---|
Method: stack-fit judgment scores (0–100) on the user's stated weights for 11 CLI agent harnesses, not benchmark measurements. Evidence: 180 product-capability rows deduplicated to canonical products, 145 repository-fact rows (official repositories preferred, forks excluded), 15 fact checks of official Factory and T3 material and 2026 research, and 4 Nebius/model findings — scanned 2026-11-02. Capability cells are the majority reading across each product's evidence rows; dashes mean no evidence found. Mainplane, Vercel eve, and dsh-superagent were cut as out-of-category. Live project facts can change.