Open Textbook
← All agents Open in the workspace →
сообщество Works from Russia Terminal

OpenCode

MIT license, no vendor account involved; with a local model, data never leaves the server.

Verified: Sep 13next price check: Dec 12

Price
your own key or OpenCode Go $10 a month
License
MIT
Model
любой провайдер, включая локальный
model under the hood
–
OpenCode: switch providers with a config edit
OpenCode: switch providers with a config edit

Facts with sources

  • In the ServerFlow benchmark across 16 tasks, the shells cluster together; OpenClaw and OpenCode held up more consistently than the rest; in one of seven runs the agent read 280K tokens and produced no working result. Source (in Russian)
  • A local server with an H100 costs upward of 6.5 million rubles; payback at these volumes exceeds 160 years. Source (in Russian)
  • Electricity is about 42% cheaper than the same tokens via API, but that doesn’t offset the hardware cost. Source (in Russian)

Strengths

  • Open source, 75+ providers and free with your own keys; the $10/month Go plan gives roughly $60 of usage by reviewers' estimates.
  • Bryan Software: about 7k tokens of baseline overhead before you type, against roughly 33k for Claude Code, with noticeably better caching.
  • For Jimmy Garcia the free tool transcribed 14 minutes into 103 phrases, cut an EDL and rendered 1080p in about four hours.
  • Plan and build modes, custom agents via config, LSP feeding compiler errors back to the model, and resumable sessions.

Weaknesses

  • Anthropic blocked subscription sign-in in January 2026: Claude only works through an API key, billed per token on top of any subscription.
  • Jimmy Garcia's agent could not read images and never checked its render: 30 seconds of black on a 4:07 video, captions covering the picture.
  • Bryan Software: on simple tasks 3 to 6 times more tokens than Pi, with users reporting 12k to 24k tokens for a plain 'hi' versus 4k.
  • Jimmy Garcia's pre-production files (title, description, hashtags) came out empty, so the published video went up with placeholders.

Example from practice

Edit and publish a video through an existing pipeline

Jimmy Garcia gave OpenCode the same skills, production rules and prompt as Claude Code: turn three screen recordings and two voice tracks into a finished, published video. Running NVIDIA's Nemotron 3 Ultra in the cloud at no extra charge, OpenCode transcribed 14 minutes of speech into 103 timestamped phrases, built a narration script from his own words, cut an EDL, graded the clips, rendered 1080p and mastered the audio to -14.1 LUFS, all in just over four hours. But it could not read image files and never looked at its own output: the published 4:07 video had 30 seconds of black at the end, an empty 26-second end card and captions covering most of the frame. His verdict: he would switch for free, but it is not there yet.

source: Jimmy Garcia III (Claude Code), 'I Pay $100 a Month for Claude Code. I Gave Free OpenCode the Same Job.'

What reviewers say

Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.

Who it is for

  • Team of up to fifteen people without a security sign-off
  • Company with a ban on external services

Models in this agent

Score is from our manager-task benchmark; models without a score are not measured yet.

Kimi K3 8.95DeepSeek V4 Flash 7.45GLM 5.3Qwen 3.8 27BGPT-6 AstraClaude Fable 5.1

How to brief an agent like this

The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.

Open in the workspace