Reasonix
esengine (Huahui Yu) · Reasonix: Open-Source, DeepSeek-First Coding Agent for Terminal, Desktop, Browser, and Editor
Reasonix is a free, MIT-licensed coding agent that runs on your own computer and uses the model provider you configure. It reads your project, edits files, and runs commands and tests from the Reasonix Studio desktop app, a terminal UI, a local browser UI, or VS Code and other ACP editors. It suits developers who want a low-cost agent tuned for DeepSeek's prefix cache, with approvals, a sandbox, and rewind built in.
Best for
Developers, especially DeepSeek users, who want a free, open-source coding agent they run themselves, with the choice of desktop, terminal, browser, or editor and real controls over approvals, sandboxing, and rewind
Not ideal for
Teams that need a vendor-backed product with support and compliance commitments, anyone who wants a hosted cloud agent, and users who prefer a slow-moving, stable tool over one that changes weekly
Who it's for
Developers who want a free, open-source coding agent with their own model keys, especially those using DeepSeek models
Capabilities
- One local engine behind four surfaces: the Reasonix Studio desktop app, a full-screen terminal UI, a browser UI served from your machine (`reasonix web` / `reasonix serve`), and editors via a VS Code extension or ACP (`reasonix acp`)
- Reads the project, edits files, and runs commands and tests, with headless `reasonix run "task"` for scripts and automation
- Cache-first context handling for DeepSeek: deterministic, append-only context, with stale tool output pruned before summary compaction
- DeepSeek preset built in, plus any OpenAI-compatible or Anthropic Messages-compatible endpoint and presets for Kimi, Qwen/DashScope, GLM/Z.ai, MiniMax, MiMo, StepFun, OpenCode Go, Vercel AI Gateway, NVIDIA NIM, Ollama Cloud, and more
- Optional planner and executor models in separate sessions, for example a stronger model planning while a cheaper one edits
- Ask, Auto, and YOLO approval postures, configurable allow and deny rules, and a Plan mode that waits for your approval before starting work
- Sandboxed commands in the Read only and Workspace write presets (Seatbelt on macOS, bubblewrap on Linux, restricted tokens and Job Objects on Windows), failing closed if the sandbox cannot start
- Snapshot-based checkpoints taken before each edit-tool change, so you can rewind code, conversation, or both by turn, or fork from an earlier point, without touching git
- MCP client for stdio and Streamable HTTP servers that also reads an existing Claude Code-style `.mcp.json`, plus Markdown skills, custom slash commands, and hooks
- Parallel sub-agents (1 to 8 tasks at once through `spawn_agents`), long-running `/goal` objectives with optional token budgets, and project memory in `REASONIX.md` or `AGENTS.md`
- Remote SSH hosts, a Web Studio page on reasonix.io that relays to your signed-in desktop Studio with end-to-end encryption, and an Android build that works against your servers over SSH
- Extension Protocol v1 sidecars and a Go SDK for plugins that add model providers, runtime hooks, and UI, plus a plugin that lets DeepSeek Harness hand tasks to Reasonix as a subagent
Limitations
- No hosted agent: Reasonix runs on your own machine, and you need an API key or account with a model provider, which you pay directly
- The cache-hit and cost-saving figures come from the project itself, and the cache-first design pays off most with DeepSeek's prefix caching rather than with every provider
- Maintained by an individual developer and community contributors, with no company, support contract, or documented security certification behind it
- Very fast release cadence: the 2.x line ships several releases a week and defaults change between versions, while the website's command reference notes that parts of it still describe the older 1.x line
- Rewind covers only files changed by its edit tools, not side effects of shell commands
- Full access and YOLO modes remove sandboxing or approval prompts, and code extensions run outside the sandbox with full trust
- `reasonix serve` starts without authentication by default, so you have to turn on token or password auth before exposing it beyond localhost
- Release notes lead with Chinese, and much of the community support runs through QQ and Discord
Use cases
- Running long coding sessions on DeepSeek models while keeping input-token costs low through prefix-cache reuse
- Fixing a bug or adding a feature in a repository from the Studio desktop app, approving each edit and command
- Scripting `reasonix run` in a CI job or batch workflow with a read-only or workspace-write permission preset
- Using one agent from the terminal, the desktop app, and VS Code with the same configuration, provider keys, and memory
- Working on a remote development box over SSH, or checking on a desktop session from a phone browser through Web Studio
- Delegating coding subtasks from DeepSeek Harness to Reasonix through its ACP subagent plugin
Our take
Reasonix is one of the most complete open-source options for developers who want to code with DeepSeek models. Its loop is designed so long sessions keep reusing DeepSeek's prefix cache, and that is a real cost lever even if the exact savings are the project's own figures. It has also grown beyond its DeepSeek-only start: any compatible endpoint works, and the desktop app, terminal, browser UI, and editor integration share one engine. The safety design is more serious than many community agents, with OS-level sandboxing, Plan mode, and per-turn rewind. The tradeoffs are governance and pace. It is one maintainer plus a community, not a company, and the 2.x line changes several times a week, so pin a version if stability matters and read the release notes before upgrading.
Who should use it
Developers who use DeepSeek models and want to keep long sessions cheap, open-source users who want a desktop app as well as a terminal agent, and tinkerers who want to bring their own provider and extend the agent with MCP, skills, or Go extensions.
Who should skip it
Organizations that need vendor support, security attestations, or a hosted agent, developers committed to Claude or OpenAI subscription plans, and anyone who wants a tool whose behavior stays fixed between monthly updates.
Strengths
- Free and open source under MIT, with no reasonix.io account needed to use it locally
- Designed to keep DeepSeek's prefix cache warm across long sessions, which the project says cuts input-token cost sharply
- Not locked to DeepSeek: any OpenAI- or Anthropic-compatible endpoint works, with many provider presets
- Four surfaces on one engine: desktop, terminal, local browser UI, and editors
- Real safety controls: approval postures, Plan mode, OS-level command sandboxing, and per-turn rewind
- Very active development with frequent releases and a large contributor base
Weaknesses
- Run by one maintainer and the community, with no company or support contract
- Releases land several times a week, so behavior and defaults change often
- Cost claims are the project's own and depend on DeepSeek's caching
- Rewind does not cover changes made by shell commands
- No hosted or cloud agent; you need your own model API key
- Some documentation still mixes the 1.x and 2.x lines
Reasonix pricing
Open Source
Free
- MIT-licensed desktop app, CLI, browser UI, and editor integration
- Runs on your own machine with your own model keys
Note: Reasonix itself is free and no paid plan is documented. Model usage is billed directly by the provider whose key you add, such as DeepSeek. Signing in to a reasonix.io account is only needed for Web Studio remote access and the skills registry. Verified against reasonix.io and the GitHub repository on 2026-11-09.
Technical specs
API pricing
No hosted API. The local engine is free and scriptable through `reasonix run`, ACP over stdio, and the `reasonix serve` HTTP frontend; model usage is billed by your provider
Available models
Where Reasonix excels
Running long refactoring sessions on DeepSeek on a budget
Append-only context keeps the request prefix stable, so each turn can reuse DeepSeek's cached tokens instead of paying full input price.
Letting an agent work unattended within limits
Auto mode with deny rules, Workspace write sandboxing, and per-turn checkpoints let it run on its own while keeping file edits reversible.
Switching between desktop, terminal, and editor on one project
Studio, the terminal UI, the browser UI, and the VS Code extension run the same local engine and share configuration, provider keys, and memory.
Reasonix vs. competitors
Reasonix vs. OpenCode
OpenCode is an open-source, provider-agnostic coding agent for the terminal, desktop, and IDEs from Anomaly. Reasonix covers similar surfaces under the same license; its distinguishing choices are a context loop built to keep DeepSeek's prefix cache warm and OS-level sandboxing in its Read only and Workspace write presets.
Reasonix vs. Claude Code
Claude Code is Anthropic's commercial agent for Claude models, billed through Claude plans or the API. Reasonix is free, open source, and runs DeepSeek or any compatible model you configure with your own key, which can cost less per session depending on the model, but it does not give you Claude models through an Anthropic plan.
Reasonix vs. DeepSeek Harness
DeepSeek Harness is DeepSeek's own plugin-based agent harness and is still a developer preview. Reasonix is an independent community project, not made by DeepSeek, focused on day-to-day coding with a desktop app, terminal UI, and editor support. The two can be combined: a plugin lets DeepSeek Harness hand tasks to Reasonix as a subagent.
Reasonix vs. OpenAI Codex
OpenAI Codex pairs an open-source CLI with cloud tasks tied to ChatGPT plans and OpenAI models. Reasonix has no cloud tasks; everything runs locally against the model endpoint you choose.
Frequently asked questions
What is Reasonix?
Reasonix is an open-source coding agent that runs on your computer. You describe a task, and it reads your project, edits files, and runs commands and tests, asking for approval as often as you choose. You can use it as the Reasonix Studio desktop app, in the terminal, in a local browser UI, or from VS Code and other ACP editors.
Is Reasonix free and open source?
Yes. Reasonix is MIT licensed on GitHub (esengine/DeepSeek-Reasonix) and costs nothing to use. You pay only your model provider, such as DeepSeek, using your own API key. No paid plan is documented.
Does Reasonix only work with DeepSeek?
No longer. Early versions were DeepSeek-only, but the current releases ship DeepSeek as a built-in preset and accept any OpenAI-compatible or Anthropic Messages-compatible endpoint. Reasonix can also prefill provider entries for Kimi, Qwen, GLM/Z.ai, MiniMax, MiMo, StepFun, OpenCode Go, Vercel AI Gateway, NVIDIA NIM, Ollama Cloud, and others. Its cache-first design is tuned for DeepSeek's prefix caching.
How do I install Reasonix?
Download Reasonix Studio for macOS (Apple silicon or Intel), Windows, or Linux (.deb, .rpm, or AppImage on amd64 or arm64) from reasonix.io or GitHub Releases. For the terminal, run `npm i -g reasonix` or `brew install esengine/reasonix/reasonix`, then `reasonix setup` to add a provider key and `reasonix` to start. The VS Code extension needs the CLI installed first.
Is Reasonix made by DeepSeek? How is it different from DeepSeek Harness?
No. Reasonix is an independent project maintained by Huahui Yu (esengine) and community contributors. DeepSeek Harness is DeepSeek's own plugin-based harness, still in developer preview. Reasonix ships a plugin that lets DeepSeek Harness delegate tasks to it as a subagent over ACP.
How does Reasonix keep changes safe?
You choose an approval posture (Ask, Auto, or YOLO) and can use Plan mode so nothing starts until you approve a plan. In the Read only and Workspace write presets, commands run in an OS sandbox. A snapshot is taken before each edit-tool change, so you can rewind code or conversation by turn. Shell command side effects are not covered by rewind, and Full access or YOLO removes those safeguards.
Reasonix vs Claude Code: what is the difference?
Claude Code is Anthropic's commercial coding agent for Claude models. Reasonix is a free, MIT-licensed agent that you run with DeepSeek or another compatible provider using your own key, with a desktop app, terminal, browser, and editor surfaces on one local engine.
Integrations & fit
Alternatives to consider
Anomaly
Developers who want an open-source coding agent they can run in the terminal, a desktop window, or a browser, with any model provider and undo for agent changes
FreeAnthropic
Developers and teams who want deep codebase reasoning, multi-file changes, and persistent project instructions in a terminal-friendly or IDE-native workflow
Paid
DeepSeek
Developers who want an open-source, DeepSeek-native agent harness they can inspect and extend with plugins, and who are comfortable running preview software in an isolated environment
Free
OpenAI
Developers and teams already paying for ChatGPT Plus, Pro, Business, or Enterprise who want one coding agent across the terminal, IDE, and parallel cloud tasks
FreemiumAbout Reasonix
Reasonix began in April 2026 as a DeepSeek-only terminal agent whose whole loop was designed to keep DeepSeek's prefix cache warm: context is serialized deterministically and only ever appended, so long sessions keep replaying cached tokens. The project says this means 90%+ of each request in hours-long sessions comes from cache and input-token cost drops to about a fifth; that is the maintainer's claim, not an independent measurement. The project has since been rewritten in Go and opened up to other models. DeepSeek is the built-in preset, and any OpenAI-compatible or Anthropic Messages-compatible endpoint can be added, with ready-made presets for Kimi, Qwen, GLM/Z.ai, MiniMax, MiMo, StepFun, OpenCode Go, Vercel AI Gateway, NVIDIA NIM, Ollama Cloud, and others. One local engine sits behind four surfaces: Reasonix Studio (the recommended desktop app for macOS, Windows, and Linux), a full-screen terminal UI with `reasonix run` for scripts, a browser UI served from your machine with `reasonix web` or `reasonix serve`, and editors through a VS Code extension or the Agent Client Protocol (`reasonix acp`). Control is the other focus. You pick an approval posture (Ask, Auto, or YOLO) and can switch on Plan mode so it waits for your approval before working. Commands run inside a sandbox (Seatbelt on macOS, bubblewrap on Linux, restricted tokens on Windows) unless you choose Full access. A snapshot is taken before each edit-tool change, so you can rewind code, conversation, or both by turn without touching git. MCP servers, Markdown skills, hooks, parallel sub-agents, long-running Goals, remote SSH hosts, and code extensions built on a Go SDK round it out. Reasonix is maintained by an individual developer, Huahui Yu (GitHub: esengine), with community contributors. There is no company, paid plan, or support contract behind it, and you pay your model provider directly. Development moves fast: the 2.x line (Studio and the current CLI) ships several releases a week, while the 1.x line only receives fixes.
Updates from Reasonix
Reasonix Studio and CLI 2.58.0 added a phone-width layout for approvals, the composer, and settings, grouped memories with importance and a forgetting curve, ranked settings search by relevance, and made Notify the default update mode for new installs. Sub-agents are now one `spawn_agents` tool that runs 1 to 8 parallel tasks and stays read-only in Plan mode.
The 1.x desktop app and CLI stopped treating `python3 -v <script>` as a read-only version check, so it now needs approval like any other command that runs code. Before the fix, a repository opened from an untrusted folder could run its own scripts without a prompt in Ask or read-only mode (GHSA-qrrg-f72x-hx7r).
The workbench can preview .docx and .xlsx files read-only in an isolated frame, `reasonix serve` gained `--password-file` so the password no longer appears in the command line, and a model watching a long background job is told a new turn starts when the job ends.
Are you the founder? Claim this listing →