Reasonix: Free DeepSeek Coding Agent with Cache-First Cost Control

Reasonix is a free, open-source AI coding agent that works inside your project with DeepSeek and other compatible model providers.

Give it a coding task, and it can inspect files, edit code, run commands and tests, call MCP servers, and keep working from the results.

This coding agent is built for long DeepSeek sessions. Reasonix keeps the system prompt and tool schemas byte-stable from turn to turn, then prunes or compacts older context as the session grows. DeepSeek can reuse the repeated prefix from cache, which can lower cached-input costs during long repository work.

An earlier Reasonix 0.x real-world cache case study recorded 435 million cache-hit input tokens in one day at a 99.82% cache hit rate. At the DeepSeek V4 Flash prices used for that test, the reported cost was $1.38 versus an estimated $61.06 with no input-cache hits.

Features

• Runs in the terminal, desktop app, browser, and ACP-compatible editors.
• Reads and edits files, runs commands and tests, and manages background tasks.
• Delegates bounded work to subagents with controlled tools and write scopes.
• Connects to DeepSeek, compatible cloud APIs, and local model endpoints.
• Maintains cache-aware context for long DeepSeek repository sessions.
• Controls writes with permissions, workspace boundaries, deny rules, and shell isolation.
• Extends projects through MCP, skills, hooks, commands, plugins, and extensions.
• Restores code or conversation from per-turn checkpoints outside Git history.
• Runs continuous Goal tasks with optional token, time, cost, or step limits.

Use Cases

• Inspect an unfamiliar repository and trace the files involved in a change.
• Edit several files, run tests, diagnose failures, and revise the patch.
• Automate a defined coding task with reasonix run from a script.
• Keep a DeepSeek coding session running across a large repository task.
• Review sessions, checkpoints, approvals, and tasks from the desktop app.
• Work on Linux or macOS development hosts through Reasonix Remote SSH.
• Run Reasonix in VS Code, VSCodium, Zed, Eclipse Theia, or another ACP host.
• Connect repository tasks to MCP servers, messaging bots, hooks, and reusable skills.

Reasonix CLI, Desktop App, Browser, and VS Code

Models and Provider Setup

Reasonix ships DeepSeek presets and can also connect to OpenAI-compatible or Anthropic-compatible providers. Run reasonix setup to create or edit providers, test connections, choose a default model, and store API keys in the global <Reasonix home>/.env file.

DeepSeek V4 Flash and V4 Pro support Reasonix reasoning controls. The recent version also added the experimental deepseek-v4-flash-vision-exp entry for image input through inline data, public URLs, or DeepSeek Files API IDs.

Skills, Plugins, MCP, and Project Instructions

Skills and subagent profiles

A Reasonix skill is a Markdown playbook stored in a project, user, or configured skill directory. A skill can run in the active session or start an isolated subagent with its own model, effort level, tool set, iteration limit, and write scope. Built-in roles include explore, research, review, and security review. Use /skills to inspect the loaded set.

Plugin packages and Extension Protocol v1

Plugin packages can bundle skills, commands, agents, hooks, MCP declarations, themes, and other project extensions. Extension Protocol v1 also supports versioned sidecars that contribute providers, tools, interceptors, and structured UI elements. Use /reload after changing extensions when you want the active session to rebuild its runtime.

MCP servers

Reasonix connects to local stdio servers and remote MCP servers. Imported prompts become slash commands, and MCP resources can be referenced inside a session. Project MCP configuration can live in .mcp.json or Reasonix configuration, with explicit Reasonix entries taking precedence when names overlap.

REASONIX.md, AGENTS.md, commands, and hooks

REASONIX.md and AGENTS.md carry project instructions into sessions. Project and user command folders create reusable slash commands, and lifecycle hooks can inspect or gate events such as session start, prompt submission, tool use, and completion. Run reasonix doctor capabilities when an expected skill, command, hook, package, or MCP server is missing.

Sessions, Memory, Goals, and Recovery

• Project instructions: Hierarchical REASONIX.md and AGENTS.md files define durable repository rules.
• Background memory: Reasonix stores approved facts as individual records and recalls them when relevant.
• Conversation branches: /tree, /branch, and /switch preserve alternate lines of work inside saved sessions.
• Checkpoints: /rewind restores code, conversation, or both from file snapshots outside Git history.
• Goal: /goal keeps a long-running objective active until completion, a real blocker, or a configured limit stops it.
• Session inbox: Follow-up messages persist while the agent is working, and steering can modify the active turn without losing queued input.
• Task Monitor: CLI and desktop views track agent tasks and subagent progress during longer runs.
• Recovery: Session recovery copies, checkpoints, and /rewind restore earlier work after interruptions or unwanted changes.

DeepSeek Reasonix vs Claude Code

Both agents inspect repositories, edit files, run commands, plan work, call MCP servers, and resume sessions. Reasonix publishes its Go code under the MIT License, works with several model providers, and maintains cache-aware context for DeepSeek. Claude Code uses Anthropic’s models, accounts, billing, and commercial product stack.

I would run Reasonix for long DeepSeek sessions, self-managed provider access, and repositories where cached-input cost can become significant. I would pay for Claude Code when Claude’s model judgment and Anthropic’s commercial support justify the price. Reasonix asks for more attention to releases, provider configuration, extension permissions, and shell confinement.

How to Use Reasonix

1. Install the Reasonix CLI

The npm package installs the prebuilt native binary on supported platforms. Homebrew is also available on macOS, and GitHub releases contain direct archives for macOS, Linux, and Windows.

npm i -g reasonix
brew install esengine/reasonix/reasonix

2. Configure a provider and API key

Run the provider manager, select DeepSeek or another provider, test the connection, choose a default model, and save the API key. Provider secrets are stored in the global <Reasonix home>/.env file.

reasonix setup

3. Start inside a repository

Change into the project directory and start the terminal UI. Give the agent a bounded task with a concrete result, such as a failing test to fix or a function to implement.

cd /path/to/your-project
reasonix

Run /init when you want Reasonix to create project instructions. Review the generated REASONIX.md or use an existing AGENTS.md for build commands, tests, code conventions, repository boundaries, and other persistent rules.

4. Choose a permission mode

Interactive sessions cycle through Ask, Auto, and Plan with Shift+Tab. Ask opens approval prompts for normal write decisions. Auto approves ordinary fallback operations under the configured rules. Plan starts the plan-first workflow, and tool calls continue to follow permissions and sandbox policy. Ctrl+Y toggles YOLO independently. Explicit deny rules and sandbox boundaries continue to apply.

5. Run a one-shot task

Use reasonix run for a defined task that should execute and exit. The command supports model selection, resume options, permission flags, extra writable directories, and structured output for automation.

reasonix run "implement the TODOs in main.go"
reasonix run --auto "run the focused tests"
reasonix -p "summarize this repository" --output-format json

6. Open Reasonix in a browser, desktop app, or editor

For local browser use, reasonix web starts the Web UI, opens your browser, and generates token authentication by default. reasonix serve starts the lower-level server for remote hosts, supervisors, tunnels, reverse proxies, and shared sessions; enable authentication before exposing it beyond a trusted loopback connection.

reasonix web
reasonix serve --auth token

The desktop app manages visual sessions, settings, approvals, checkpoints, Task Monitor, Automation, Remote SSH, MCP status, and bot connections. The VS Code extension uses the locally installed Reasonix CLI, and other ACP hosts can start the agent with reasonix acp.

7. Configure Skills, MCP servers, and extensions

Project skills can live under .reasonix/skills/ and appear in /skills. MCP servers can be configured through Reasonix settings or project MCP configuration. Plugin packages and Extension Protocol v1 sidecars can contribute commands, hooks, providers, tools, skills, and other runtime capabilities.

8. Diagnose the loaded environment

Run the capability check when a project instruction, skill, command, hook, package, or MCP server fails to appear. Use --live only when the diagnostic process should start configured third-party MCP servers.

reasonix doctor capabilities
reasonix doctor capabilities --json

9. Update Reasonix

Check the available production release before upgrading when you want to review the target version first.

reasonix upgrade --check
reasonix upgrade

Permissions and Sandbox Boundaries

Permission rules resolve explicit deny and ask decisions before ordinary fallback behavior. Read operations usually proceed, and writes or shell commands follow the active rules and permission mode. Headless runs fail closed when an approval would require a person unless the selected mode authorizes that operation.

The file sandbox resolves symlinks and .. segments before accepting a path. File-writing tools stay inside [sandbox] workspace_root unless another destination appears in allow_write. Private paths such as ${HOME}/.ssh can be placed in forbid_read.

Shell isolation depends on the operating system. Seatbelt handles macOS, and bubblewrap handles Linux. With bash = "enforce", Reasonix refuses Bash when the required backend is unavailable. Windows has no equivalent OS-level Bash jail. Keep sensitive work in a disposable branch, container, VM, or WSL environment and retain explicit deny rules for destructive commands.

Free Access, Privacy, and Limits

Reasonix is released under the MIT License and can be downloaded, modified, and built from source. Model inference is billed by the provider you configure. DeepSeek usage depends on the selected model, session length, cache hit rate, and DeepSeek pricing.

The agent, session files, approvals, and file operations run locally. Prompts and code placed in model context go to the provider endpoint you configure. Optional CLI telemetry uses a once-per-day anonymous install ping and bounded reliability counters after the telemetry preference is saved. It excludes prompts, responses, credentials, repository paths, exact token counts, provider names, and model names. Use reasonix config telemetry off to disable reporting and delete pending counters.

Offline inference needs a local OpenAI-compatible endpoint that you operate. Tool calling, context limits, reasoning controls, image support, and caching then depend on that local model server.

Quick Reference

Core Commands

Useful Slash Commands

Configuration and Memory Paths

Pros

• MIT-licensed Go project
• DeepSeek cache-aware context handling
• CLI, desktop, browser, and editor access
• Cloud and local compatible providers
• MCP, skills, hooks, and extensions

Cons

• Frequent release and configuration changes
• No OS-level Bash jail on Windows
• Remote SSH needs Linux or macOS
• Model costs depend on provider

Alternatives and Related Resources

• DSH Desktop: Run DeepSeek Harness as an Open-source Desktop App
• DeepSeek Harness: Open-Source Plugin-Based AI Agent Harness
• 10 Best CLI AI Coding Agents: Open-Source & Commercial
• CodeWhale (DeepSeek TUI): Free, Open-source Claude Code Alternative
• The Ultimate Claude Code Resource List: Agents, Skills, Plugins & More
• Aperant (Auto Claude): Open-Source AI Coding Agent Harness
• Best DeepSeek Harness Plugins for DSH
• Best & Open-source MCP Servers

FAQs

Q: Should you use npx reasonix code?
A: DeepSeek’s integration guide documents npx reasonix code. The current Reasonix 1.x documentation centers the native CLI path: install reasonix, run reasonix setup, then start a session with reasonix. This article follows the native 1.x workflow.

Q: Can Reasonix use OpenRouter?
A: Yes. Configure OpenRouter as an OpenAI-compatible provider through reasonix setup or reasonix.toml. Tool calling, reasoning controls, image input, and context limits depend on the model you select.

Q: Can Reasonix use a local model?
A: Yes. Connect Reasonix to an OpenAI-compatible local endpoint. Tool calling, context limits, image support, reasoning controls, and cache behavior depend on that model server.

Q: Can Reasonix work with images?
A: Yes, with an image-capable model. Reasonix v1.31.3 added the experimental DeepSeek deepseek-v4-flash-vision-exp catalog entry for image input.

Last Updated: Aug 23, 2026