DSCode Coding Agent

A local-first coding agent that routes repository tasks across providers, with DeepSeek as the default.

Stars
★ 365
Last updated
23d ago
License
MIT
Primary language
TypeScript

At a glance

How it runs
CLIDesktop appLibrary / SDK
Works with
Universal · cross-platformOpenAI API · Claude APICodex (Partial support)
Cost
Free software; you pay for model usage
Setup effort
Low · running in minutes
You'll need
Node.js 22.19+GitripgrepShell / CLINetwork accessLocal filesystemMCP Server
Typical use
Developers who want DeepSeek V4 Flash for routine coding tasks and the option to switch providers when needed.
Not a fit if
  • Teams that require a published VS Code Marketplace extension
  • Linux or Windows users unable to configure a Docker sandbox
Source review
66/100 · Some gaps

What does this agent do, and when should you use it?

DSCode is a local-first coding-agent runtime that defaults to DeepSeek V4 Flash and also supports Codex, OpenAI, Anthropic, OpenRouter, Z.AI, Kimi, MiniMax, xAI, and OpenCode Zen Go. Use it through the terminal TUI, one-shot commands, JSONL, or RPC, with optional Electron desktop, VS Code, and @thinkany/dscode-core clients. It reads project files and instructions, runs sandboxed commands, and edits code with patches that create durable checkpoints on success. Session transcripts are stored locally as JSONL; SQLite holds thread metadata and related state. It requires Node.js 22.19+ and Git, while isolation on Linux and Windows depends on a Docker sandbox configured by the user.

Start DSCode with dscode -C to select a project directory, then choose and authenticate a provider with /login. The agent reads workspace code and AGENTS.md or CLAUDE.md instructions, asks the selected model to analyze the task, and uses sandboxed commands, background-process interaction, and apply_patch to make changes; the safe harness also adds file reading, search, and language diagnostics. Users can inspect diffs, restore checkpoints, run tests, or delegate to explorer, implementer, reviewer, and tester roles, with implementers working in isolated Git worktrees. Results are available in the interactive terminal, one-shot -p output, JSONL automation, or RPC events for clients; session transcripts are written locally in JSONL.

  1. Developers who want DeepSeek V4 Flash for routine coding tasks and the option to switch providers when needed.
  2. Individuals or teams that want an agent to inspect a local repository, run restricted commands, and change files through patches.
  3. Teams that want parallel exploration, implementation, review, and testing, with implementation isolated in Git worktrees.
  4. Engineering teams integrating a coding agent into scripts or CI through JSONL, stdin/stdout RPC, or @thinkany/dscode-core.
  5. Developers who prefer an Electron desktop client or local VS Code integration for the same agent runtime.

How do you install or deploy this agent?

Requires Node.js 22.19+ and Git. Install globally with npm and ensure ~/.local/bin is on PATH. You can also install the latest source build or use a desktop installer from the README. macOS desktop builds are signed and notarized by Apple; Windows and Linux packages are currently unsigned.

How do you use this agent?

On first launch, enter /login in the TUI and choose a provider. DeepSeek requires an API key; other providers use their documented account or API-key authentication. You can also authenticate with dscode login before starting a project. The default DeepSeek model is deepseek-v4-flash; override it with --provider and --model. Start a project with:

What are this agent's strengths and limitations?

Pros
  • Defaults to DeepSeek V4 Flash while supporting multiple providers, so users can choose models for different tasks.
  • Local JSONL transcripts, patch checkpoints, and workspace sandboxing provide inspectable history and a recovery path.
  • Supports up to four parallel roles, with implementers working in isolated Git worktrees.
  • Offers CLI, desktop, Core RPC package, and local VS Code integration entry points.
Limitations
  • Cloud model use requires the relevant account or API key and may incur provider charges; OpenAI API-key usage is billed separately from ChatGPT plan access.
  • DeepSeek V4 Flash is text-only; image input requires a vision-capable model.
  • Command isolation on Linux and Windows depends on a user-configured Docker sandbox; macOS uses Seatbelt.
  • The VS Code extension is not yet published to the Marketplace, and the project describes its ecosystem as less mature than Claude Code and Codex.

How does this agent compare with similar options?

The project compares DSCode with Claude Code and Codex, describing those products as having broader and more mature ecosystems. DSCode positions itself as a smaller, DeepSeek-first, locally controlled runtime under the MIT license.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
DSCode Coding Agent This agent 66 · Some gaps CLIFree + model costs ★ 365 23d ago TypeScript OpenAI API · Claude API
zerostack 57 · Major gaps CLIFree + model costs ★ 1.7k 1d ago Rust OpenAI API · Claude API
Synara 52 · Major gaps Desktop appFree + model costs ★ 2.1k today TypeScript Codex · Claude Code
OpenClaude — Universal Terminal AI Coding Agent 42 · Major gaps CLIFree + model costs ★ 34k today TypeScript Claude Code · OpenAI API · Claude API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Some gaps
66/ 100 5-point scale 3.3 / 5
Trust 17/29
Reliability 8/14
Adaptability 14/18
Convention 14/18
Effectiveness 9/13
Verifiability 4/8
Why each dimension lost points
Trust17 / 29 · 2.9/5

The README describes workspace-write and network-blocked defaults, approval for network or host access, project trust and full-access modes. It also says API keys are removed from child environments, credentials use the OS keyring by default, and the file fallback is owner-only. Permission modes and external effects have practical descriptions, and successful patches are said to create checkpoints. Deductions: the supplied material lacks implementation code to verify these boundaries; the Windows document is a contract for an experimental backend that is disabled by default, not evidence of an implemented safeguard; no dependency vulnerability review is supplied, and publisher identity is unverified, so attribution is only partial.

Reliability8 / 14 · 2.9/5

The README, package manifest, and CI workflow are broadly consistent about Node, pnpm, build, and check commands. CI specifies typechecking, tests, build, and package smoke checks; the Windows sandbox contract requires clear failure and no execution when setup or enforcement checks fail. Deductions: dependencies use version ranges, and the supplied material gives no lockfile or detailed availability policy; test and CI configuration show a validation process exists but do not show that this revision passed.

Adaptability14 / 18 · 3.9/5

The materials describe terminal, desktop, IDE, CI/RPC scenarios, multiple providers, model capability differences, and limits such as DeepSeek's text-only mode, platform sandbox dependencies, and the unpublished IDE extension. Permission and network modes are concrete, and the Windows native sandbox scope is bounded. Deductions: adaptability claims come mainly from the README without implementation evidence for each deployment; commands and settings provide trigger examples, but these materials do not specify activation conditions for every extension feature.

Convention14 / 18 · 3.9/5

The README is organized around installation, configuration, safety, commands, and limitations, with concrete command examples; the license is explicitly MIT. It lists platform signing status, Docker image requirements for Linux and Windows isolation, and the project's early-stage status. The package manifest gives a version, and CI describes the main-branch release flow. Deductions: the supplied files do not include a full changelog, version history is limited, and maintenance ownership or support commitments are unclear; publisher identity is unknown.

Effectiveness9 / 13 · 3.5/5

The README describes local sessions, parallel agents, provider routing, usage and cost reporting, checkpoints, and automation entry points, suggesting practical coverage of common coding tasks and usable outputs. It also states the low-cost positioning and product boundaries. Deductions: this static material contains no real task results or execution evidence to establish output correctness, comparative value, or cost advantage; product value is primarily self-described.

Verifiability4 / 8 · 2.5/5

The README gives concrete modes, commands, paths, and configuration examples for many feature and safety claims. The Windows contract supplies checkable requirements and acceptance cases, while the CI file identifies planned checks. Deductions: these files do not cross-verify the core runtime implementation, and the Windows acceptance list is a specification rather than a record of passing tests; no build or test result for this revision is supplied, so traceability and fact-versus-claim separation receive limited scores.

Risks and how to mitigate them
  • Sandboxing, credential isolation, and checkpoint behavior are project claims; the supplied materials contain no implementation code or revision-specific validation results to confirm runtime behavior.
  • The native Windows sandbox is explicitly experimental and disabled by default; Linux and Windows isolation depends on a Docker image configured by the user.
Evidence confidence: Low Reviewed Oct 22, 2026 Reviewed revision e2a131c7f87a
See the full review method →

FAQ

Does using DSCode cost money?
The software is MIT-licensed. Cloud model access requires the relevant provider account or API key and may incur model charges; OpenAI API-key usage is billed separately by the API platform.
Can commands access the network or host files by default?
The default command boundary is workspace-write with network blocked. The TUI can request one-time or session approval for network or host access; full and -y relax those restrictions.
Can it work with screenshots or other images?
Yes, paste an image in the TUI or pass it with @file, but use a vision-capable model. DeepSeek V4 Flash is text-only.
How are commands isolated on Linux and Windows?
The README specifies a user-configured Docker sandbox for Linux and Windows. If no sandbox backend is available, DSCode fails closed rather than running commands on the host.
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents