DeepSeek Harness vs OpenCode: A Developer-Preview Plugin Harness vs an Open-Source Coding Agent
DeepSeek Harness
DeepSeek's open-source, plugin-based harness for running AI coding agents locally
OpenCode
The open-source coding agent for terminal, desktop, IDE and web
Verdicts par tâche
DeepSeek's README says breaking changes will come in its developer preview; the OpenCode CLI and docs pages we read carry no preview label.
OpenCode's desktop app is marked BETA in its README, and its permissions are mostly allow by default, so isolate it in a VM or container.
DeepSeek's own SAFETY.md says it must not be treated as secure or production-ready; the OpenCode pages we read carry no such disclaimer.
A narrow call. The OpenCode permissions page we read describes no sandbox, so neither is recommended on a machine with secrets without a disposable VM.
Setup is a Settings, Models card with one API-key field in a local web UI (providers guide); OpenCode's route is /connect in its terminal or desktop interfaces (providers page).
Both support DeepSeek through an API key, so this is a setup-steps call only, untested by us.
The architecture doc places model adapters, tools, persistence, sandbox and approval policy in a plugin-based base layer.
OpenCode also lists plugins, an SDK and a server on its docs home; this is a design-intent call, not a tested one.
Comparaison des fonctionnalités
| Dimension | DeepSeek Harness | OpenCode |
|---|---|---|
| IA native | IA native | IA native |
| Licence and software cost | MIT; the harness itself is free and needs a DeepSeek API key or another configured provider (README; providers guide) | MIT; free base, with optional paid Zen (pay as you go) and Go ($10/month, Go Plus $40/month) model plans (GitHub repo; opencode.ai/docs/go) |
| Release maturity | Developer preview. README: "DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES." Repo page shows 245.7k stars and 29.5k forks (GitHub repo, read 2026-10-09) | Repo page shows 212.2k stars and 28.3k forks, and marks the desktop app "(BETA)". The CLI and docs pages we read carry no preview label (GitHub repo; docs home, read 2026-10-09) |
| Interfaces | Web UI at http://127.0.0.1:3080 started with npx @deepseek-ai/dsh web; the user guide also references a Python SDK and other CLI modes (README; user guide) | Terminal interface, desktop app (marked BETA in the repo) and IDE extension per the docs home (docs home; GitHub repo) |
| Model choice | Built-in provider ids anthropic, openai, moonshotai (Kimi) and zai (GLM); a DeepSeek card in Settings, Models with one API-key field; custom providers over openai-completions, openai-responses or anthropic-messages (providers guide) | Providers page: "75+ LLM providers" via the AI SDK and Models.dev, support for local models, and a DeepSeek section: create an API key in the DeepSeek console, run /connect, search DeepSeek, enter the key, then pick a model such as DeepSeek V4 Pro with /models (opencode.ai/docs/providers/) |
| Extensibility design | "Everything is a plugin" on Cordis: model adapters, tools, persistence, sandbox and approval policy sit in the dsh-base layer, with web-app, headless, SDK and ACP bundles on top (architecture doc) | The docs home lists SDK, Server, Plugins and Ecosystem sections and configuration for Tools, Rules, Agents, Models and Permissions; the live profile adds MCP servers, Agent Skills and opencode serve from the earlier docs fetch (docs home, read 2026-10-09; profile, 2026-10-03) |
| Sandbox and permissions | SAFETY.md: "has not undergone a security audit and must not be treated as secure or production-ready"; OX Research describes bubblewrap, Landlock or Seatbelt sandboxes that confined file writes but left networking open (SAFETY.md; OX Research) | Permissions docs: most permissions default to "allow", doom_loop and external_directory default to "ask", .env reads denied; the permissions page we read does not discuss sandboxing (opencode.ai/docs/permissions) |
| Disclosed vulnerability | CVE-2026-82533, CVSS 9.4: a sandboxed process could reach the unauthenticated local agent-control API via a client-supplied loopback Host header and raise its session to danger-full-access with approvals off. OX Research: report disclosed to VulnCheck Aug 24, 2026; published Sep 8, 2026; fixed in 0.1.2-alpha.1 (OX Research; VulnCheck advisory) | GHSA-632h-h47v-g4x4: High, CVSS 7.5, no CVE, published Sep 24, 2026; /global/upgrade lacked origin validation; affects npm, pnpm and Bun installs running opencode serve; patched in 1.18.22 (GitHub advisory) |
Both free to run. OpenCode adds optional paid model plans; DeepSeek Harness sells none on the pages we read.
Stars and forks do not separate them (DeepSeek Harness shows more of both). The separating fact is the README: only DeepSeek Harness states that breaking changes will come.
On the pages we read, OpenCode lists more interfaces; DeepSeek Harness is browser-first.
Both support DeepSeek models through an API key. OpenCode lists the wider provider range on the pages we read. We tested neither.
By design intent, DeepSeek Harness puts more of its core (including sandbox and approval policy) behind plugins; OpenCode exposes extension points around the core. Untested by us, and OpenCode also has plugins.
Neither is a safe default on a machine with secrets. On the pages we read, only DeepSeek Harness puts a not-production-ready warning in its own notice.
Both have a fixed version. The DeepSeek flaw let an agent escape its own sandbox on default configurations; the OpenCode flaw needed a victim running opencode serve plus a malicious web page.
Toujours indécis entre les deux ?
Choisir est la partie facile. Le travail, c’est de le faire tourner dans votre entreprise, sur vos données, avec votre équipe qui s’en sert. Nous faisons les deux.
La conversation tourne sur Gnosari, l’un des outils de ce répertoire. Une vraie conversation, pas un script de vente.