Fast AI coding for your table.
About 69% for SWE‑bench Verified.

Inside 1.7.0 too: Night Shift fit try one task many times and keep di best attempt (/night add --best-of N), one machine fit run plenty Signal bridges, and a client fit choose between two nodes for one network without e lose di one wey e know. Release notes and downloads.

Install

$ curl -fsSL https://github.com/dreamference/ragon/releases/latest/download/install.sh | bash

$ ragon-admin server start      # e go download di model once (~20 GB), then e go start serve am
$ cd ~/your-project && ragon    # na so e be

> irm https://github.com/dreamference/ragon/releases/latest/download/install.ps1 | iex

Try dis ones first

Or put work for tonight queue with /night add <task>: Ragon go work through di queue while you dey sleep, each task for im own git branch, and you go review di branches with coffee for hand.

$ ragon "explain this repository to me like I'm new on the team"
# reads the code index (ragon-code outline, def, refs, callers) instead of grepping,
# then walks you through the layout, the entry points and how they connect

$ ragon "the tests are failing: find out why and fix them"
# runs the tests, reads the failure, edits the code and runs the tests again,
# asking before anything risky

> /night add --best-of 3 add input validation to the signup form, with tests
# tonight: three attempts, each in its own git worktree; the one the chooser picks
# is committed on its own branch, and nothing is merged or pushed
$ ragon-admin night status      # the timer, the window and the queue
> /night apply <id>            # in the morning: bring the branch back as uncommitted changes

Wetin you go get

Ragon 1.7.0 desktop app put Ask and Work for one window. Work na full workbench for agent sessions: composer wey get files, skills and commands, review pane to stage, revert and commit, worktrees, terminal tabs, and file tree and file tabs beside di thread. For your oda computers, e be client of your GB10, wey e dey find for di local network.

$ ragon worktree on   # new sessions start in their own worktree
$ cd ~/your-project && ragon
> /airgapped on      # no network for any command it runs
> /worktree apply    # bring the changes back to your checkout

$ sudo apt install ./ragon_1.7.0_arm64.deb

$ ragon-admin mcp            # stdio MCP for JetBrains and VS Code
$ ragon signal setup         # Note to Self only, off until linked
$ ragon chat telegram setup  # topics and the Mini App

Why na Ragon

SWE-bench Verified, wey we measure for one DGX Spark

On 2026-10-08, Ragon 1.5.1 agent with Qwen3.8-27B solve 68 out of 100 SWE-bench Verified tasks for one DGX Spark, offline: fixed sample of 100 tasks, one attempt per task, default mode. If you scale am to all 500 tasks, na estimated ≈69% (95% band 63.9–73.4%). Na our own measurement, for our own machine and for a subset, no be leaderboard entry.

Read am as company, no be ranking. For di same 100 tasks Ragon solve as many as Qwen3-Coder-480B and GLM-4.6, models wey big pass am many times, but every row dey use different agent, different hardware and, for di leaderboard rows, di official x86 test images, while our own run for community arm64 images. For refine mode, one first session go study di task without changing anything, then new session go do am: e solve one task pass di default (69 against 68, each side win tasks wey di oda one lose), for about 2.2 times di time per task, so e remain opt-in (ragon --refine).

1. Ragon: 100 tasks, one attempt each; one DGX Spark (GB10, 128 GB), model dey serve locally, no cloud; we measure default mode for 2026-10-08, refine mode for 2026-10-09. Estimate from di SWE-bench experiments per-task data.
2. 500 tasks, one attempt; dem no talk di hardware. Source: experiments, 20250805_openhands-Qwen3-Coder-480B-A35B-Instruct. Di same model with mini-SWE-agent: 57/100, 55.4%.
3. 500 tasks, one attempt, dem no talk which agent; dem no talk di hardware. Source: experiments, 20250930_zai_glm4-6.
4. 500 tasks, one attempt; dem no talk di hardware. Source: experiments, 20260217_mini-v2.0.0_kimi-k2-5-high.
5. 500 tasks, one attempt; dem no talk di hardware. Source: experiments, 20260217_mini-v2.0.0_deepseek-3-2-high.
6. 500 tasks each; dem no talk di hardware. Four submissions: OpenHands (80/100, 77.6%), live-SWE-agent (80/100, 79.2%), Sonar Foundation Agent (77/100, 79.2%), mini-SWE-agent, high (79/100, 76.8%). One earlier mini-SWE-agent run score 74/100, 74.4%.
7. 500 tasks, one attempt; dem no talk di hardware. Source: experiments, 20251120_livesweagent_gemini-3-pro-preview. Di same model with mini-SWE-agent: 72/100, 74.2%.
8. Di best published 20–40B model for di leaderboard. 500 tasks; test-time scaling dey use two or more attempts per task; dem no talk di hardware. Sources: test-time scaling, single attempt.
9. 500 tasks, one attempt; dem no talk di hardware. Source: experiments, 20251209_mini-v1.17.2_devstral-small-2512.
10. GitHub own figure, April 2025: "agent mode achieves a pass rate of 56.0% on SWE-bench Verified with Claude 3.7 Sonnet". Dem no talk number of tasks, attempts and hardware; e no dey for di leaderboard. Source: GitHub blog.
11. Anthropic dey publish SWE-bench Verified for im Claude models, wey dem run for simple scaffold of dia own (bash tool and edit tool), no be for di Claude Code product. Sources: Claude Code, Anthropic SWE-bench method.
12. Cursor dey publish results for im own benchmark, Cursor Bench, no be SWE-bench Verified. Source: Cursor blog, Composer.

All di leaderboard numbers: swebench.com and di SWE-bench/experiments repository; "same 100 tasks" na each submission published per-task results wey dem limit to di tasks wey Ragon run.

Paper

FAQ