DeepSeek Harness vs Claude Code vs Codex: Open-Source Agent Frameworks Compared

AI Tools Insight • 2026-10-27 • AI Coding DeepSeek Harness Claude Code Codex AI Agents Open Source Comparison

DeepSeek Harness passed 170,000 GitHub stars within 11 days of release. Claude Code took more than a year to reach 140,000. Codex took 15 months to reach 103,000. The star count alone does not make a tool better. It does tell you where developer attention is moving.

This comparison covers three coding agent frameworks: DeepSeek Harness (dsh), Claude Code, and OpenAI Codex. We compare architecture, benchmarks, pricing, and extension models as of August 24, 2026.

Sources: DeepSeek Harness repo, OpenAI Codex stats, Claude Code pricing, DeepSeek API pricing, DeepSeek Claude Code integration, Codex custom providers

Checked against vendor pages on 2026-10-27.

Codex entry at $20/mo through ChatGPT Plus is correct, and Claude Pro at $20/mo is correct, with $17/mo on annual billing.

DeepSeek renamed the model and changed the rates. The model name deepseek-v4-flash is retired. Use deepseek-flash, which serves DeepSeek-V4.1-Flash-1017 at the Flash price. Official rates per 1M tokens: output costs $0.60 off-peak and $1.20 at peak for deepseek-flash, and $1.98 off-peak and $3.96 at peak for deepseek-v4-pro. A cache-miss input token costs $0.15 or $0.30 on Flash and $0.66 or $1.32 on Pro. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday.

What Is an Agent Harness?

An agent harness is the layer between a language model and the real world. It gives the model file access, shell commands, tool calls, session logs, and approval rules. Without a harness, a model can only answer text. With a harness, it can edit code, run tests, and finish tasks.

Claude Code and Codex package this layer as commercial products. DeepSeek Harness makes the entire layer open source and replaceable. That difference shapes everything else in this comparison.

Quick Verdict

Use Case Pick Why
Full control, any model, inspectable code DeepSeek Harness MIT license, every component is a plugin
Best out-of-box coding agent Claude Code Mature product, broad IDE support, high user satisfaction
OpenAI ecosystem and cloud agents Codex Tied to ChatGPT, strong Terminal-Bench scores, cloud tasks

DeepSeek Harness: Everything Is a Plugin

DeepSeek published dsh on August 13, 2026 under the MIT license. The repo shows 170.8k stars, 18.4k forks, and 12,940 commits as of this writing. It is written in TypeScript and built on Cordis, a plugin kernel described in the paper A Programming Paradigm for Spatiotemporal Composability.

The core idea is simple: every capability is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the web UI are all swappable from configuration. There is no privileged core to patch.

Three concepts carry most of the design:

Profiles and bundles. A running instance is a plugin tree built at boot. The shipped profiles are web and headless. Each profile stacks bundles, and any config row can be replaced with a patch file. Teams override behavior without forking the code.

The session log. The append-only event log is the single source of truth for what the model sees. The documentation calls this invariant "model-visible means logged". Forking, resuming, and telemetry all come from this stream.

Capability seams. A seam is a swappable capability with three roles: service definition, provider, and consumer. Pointing the filesystem and subprocess providers at a remote sandbox moves Bash, PTY, and LSP tools with them.

The model adapter is just another plugin. You can run dsh against DeepSeek, Anthropic, OpenAI, or any OpenAI-compatible endpoint, including local models via Ollama. Claude Code and Codex are optimized for their vendors' models.

Claude Code: The Mature Commercial Agent

Claude Code is Anthropic's terminal-native coding agent. It reached roughly 140k GitHub stars by July 2026 and 20.2 million VS Code extension installs. It runs in your shell, reads your codebase, and handles git workflows in plain English.

Its extension system covers skills, hooks, MCP servers, and plugins. Anthropic is building a platform, not a chat wrapper. A change on June 15, 2026 split programmatic usage into a separate credit pool. Pro plans get a $20 monthly credit for the Agent SDK, claude -p, and GitHub Actions.

Claude Code is proprietary. The default setup works only with Anthropic models and an Anthropic account. You can point it at DeepSeek through the Anthropic-compatible API. DeepSeek documents this setup for Claude Code with environment variables. Local models are not supported.

OpenAI Codex: The Cloud-Connected Agent

Codex CLI launched in April 2025 under Apache 2.0 and was rewritten in Rust in June 2025. It reached 103,504 stars and over 5 million weekly users by July 2026. npm downloads hit 46.9 million over the last 30 days.

Codex spans more than the terminal: CLI, IDE extensions for VS Code and Cursor, a desktop app, a cloud agent at chatgpt.com/codex, and iOS. Cloud integrations include automatic code review and Slack.

The CLI is open source. The cloud agents and desktop app sit behind ChatGPT subscriptions, and billing moved to token-based credits on April 2, 2026. Codex also supports custom model providers through config.toml, so you can point it at DeepSeek or other OpenAI-compatible endpoints.

Architecture Comparison

Dimension DeepSeek Harness Claude Code OpenAI Codex
License MIT Proprietary Apache 2.0 (CLI)
Language TypeScript TypeScript Rust
Repo read, edit, test Yes Yes Yes
Shell and dev tools Yes Yes Yes
Planning and subagents Yes Yes Yes
Sandbox and approvals Configurable via plugins Mature built-in Granular controls
Model choice Any (DeepSeek, Anthropic, OpenAI, local) Claude by default, DeepSeek via env config OpenAI by default, DeepSeek via config.toml
Primary interfaces Web UI, headless, Python SDK Terminal, VS Code, JetBrains, desktop, Slack CLI, IDE, desktop, web, iOS
Hosted background agents Not offered Yes Yes
GitHub PR workflow Not documented Yes Yes
Maturity Developer preview Established Established
VS Code extension installs N/A (web UI) 20.2M 11.5M

Sources: VentureBeat launch analysis via flowtivity, gradually.ai Codex stats

Benchmarks: What the Numbers Measure

Benchmark results are the hardest part of this comparison to read. Each vendor reports its own numbers with its own methodology.

Benchmark DeepSeek Harness + V4-Pro-0813 Claude Code (Opus 4.8) Codex (GPT-5.6 Sol)
Terminal-Bench 2.1 87.9 78.9 88.8
Toolathlon-Verified 74.1 Not published Not published
DSBench-FullStack 71.1 Not published Not published
SWE-bench Pro Not published Not published 64.6

Two caveats matter.

DeepSeek's numbers are self-reported. The company tested V4-Pro-0813 inside its own Harness in minimal mode. Some headline results measure the model inside DeepSeek's execution environment, not the model alone. The same table shows Fable 5 ahead of DeepSeek on Toolathlon-Verified (77.9) and DSBench-FullStack (77.2).

Claude Code's score comes from a third-party scan. The 78.9 figure for Opus 4.8 on Terminal-Bench 2.1 comes from morphllm's 2026 CLI comparison, not from Anthropic. Anthropic does not publish a Terminal-Bench score for Claude Code.

Treat all four rows as directional, not as a race result.

Pricing: Free Software, Paid Models

DeepSeek Harness is free. The MIT license permits commercial use, modification, and redistribution. You pay only for model API usage. Because the model adapter is a plugin, you can point it at Ollama and pay nothing beyond your own hardware.

Claude Code starts at $20 per month for Pro. Max plans cost $100 or $200 per month. Team Premium is $100 per seat with a 5-seat minimum. Pay-per-token API billing is also available. The June 15, 2026 programmatic credit change means heavy SDK use draws from a separate pool.

Codex starts at $20 per month through ChatGPT Plus. Billing is token-based since April 2, 2026. The cloud agent and desktop app require a paid plan.

Cost Item DeepSeek Harness Claude Code Codex
Software license $0 Proprietary CLI free, cloud paid
Entry subscription None required $20/mo Pro ($17/mo annual) $20/mo ChatGPT Plus
Model choice for lowest cost DeepSeek V4-Flash Claude (DeepSeek via env config) OpenAI (DeepSeek via config.toml)
Local model option Yes (Ollama) No No

DeepSeek API Pricing Changed on August 16, 2026

DeepSeek replaced flat API pricing with peak and off-peak rates on August 16, 2026. Peak windows are 01:00-04:00 UTC and 06:00-10:00 UTC, Monday through Friday.

Model Cache-miss input (peak) Output (peak) Output (off-peak)
deepseek-flash output $0.60 / 1M off-peak $1.20 / 1M peak Input misses cost $0.15 off-peak or $0.30 at peak
deepseek-v4-pro output $1.98 / 1M off-peak $3.96 / 1M peak Input misses cost $0.66 off-peak or $1.32 at peak

V4-Pro output at peak is $3.96 per million tokens, up from $0.87 flat. That is a 355 percent increase. Off-peak rates are half of peak, and cache hits cost a fraction of cache-miss input.

The pricing shift does not change the tool comparison. dsh software stays free. It changes the operating cost of running DeepSeek models inside any agent. Early testers report a 99 percent cache hit rate inside dsh, which cuts effective input costs by a large margin on repeated tasks.

Source: DeepSeek API pricing, techtimes pricing report

Extension Model: The Real Difference

All three tools support plugins. The difference is scope.

Claude Code extends through skills, hooks, MCP servers, and plugins. The core product is fixed. You add around it.

Codex extends through skills, MCP, and custom agents. The core is fixed. You add around it.

DeepSeek Harness replaces the core itself. The agent loop, the model adapter, the session log, the sandbox, and the UI are all plugins. Remove one, and the system returns to the exact state it had before that plugin loaded. The Cordis paper proves this property, called compositional recovery. It validates the design on Koishi, a chatbot framework with 4,000 plugins running in production for four years.

For most developers, Claude Code's fixed core is a feature, not a limitation. It works well without configuration. dsh's replaceable core matters most to agent builders. It also matters to teams that swap models often, and to anyone who audits every token that reaches a model.

A real dsh plugin, read at the source

The design claim above is testable. TapTap, a game platform, ships a plugin for dsh and publishes its source. The plugin is called @taptap/dsh-maker and it shows how the patch-row design works in practice.

It comes in two layers.

Layer Install What it writes
L1 taptap-maker install --ide dsh One mcp-client line in $DSH_HOME/cordis.patch.yml, with hot module replacement
L2 dsh plugin add @taptap/dsh-maker A bundle that carries an MCP server and four skills as one package

The L2 bundle adds one row to the same patch file:

- insert:
    - id: taptap-maker
      name: '@taptap/dsh-maker'

The plugin module then does three things at activation. It mounts a skill-filesystem instance for its own skills. It mounts @deepseek-ai/dsh-mcp-client for the tool server. It registers a shell variable that holds the absolute path of its CLI.

The implementation detail is the interesting part. The plugin resolves its MCP entry point with require.resolve at runtime, starts it with an absolute Node.js path, and binds no profile path. The dsh documentation claims that a config row can be replaced without forking the code. A third-party vendor built its integration on exactly that row, and the bundle is self-contained enough to list in a plugin market.

The MCP server behind it is not a demo either. It covers leaderboards, cloud saves, sharing and multiplayer for TapTap mini games. It authenticates with OAuth 2.0 device code flow, so a user scans a code and no key is copied. It speaks three transports: stdio, server-sent events, and HTTP. The published bundle is 567 KB in one file.

One limit: a game platform plugin says nothing about coding agents. Its value here is evidence, not utility. It shows that the plugin kernel, the patch layer and the skill and MCP seams work outside the vendor's own code.

Verified against the TapTap repository on 2026-11-06. Sources: taptap/instant-games-open-mcp, DSH plugin documentation.

When Each Tool Makes Sense

Choose DeepSeek Harness when:

Choose Claude Code when:

Choose Codex when:

How to Start with DeepSeek Harness

Install Node.js, then run one command:

npx @deepseek-ai/dsh web

The command starts a web UI at http://127.0.0.1:3080. Open Settings, then Models, and paste a DeepSeek API key. Pick a workspace, then start a session.

The download is heavy. The first npx pull takes several minutes on a fast connection. Expect a full application, not a lightweight script.

A headless profile ships for one-shot runs, and a Python SDK covers programmatic use. Community guides describe working Ollama and OpenRouter setups in under ten minutes.

Bottom Line

DeepSeek Harness is the most architecturally ambitious of the three. It is also the youngest, labeled a developer preview with breaking changes ahead. The repo warns in capitals: "THERE WILL BE COMPATIBILITY-BREAKING CHANGES."

Claude Code remains the safest default for everyday coding. Codex is the strongest choice inside the OpenAI world. dsh is the one to evaluate now if you build agents or want control over every layer.

The pattern dsh legitimizes, a fully composable agent runtime with auditable session logs and swappable everything, is the direction serious agent teams are moving. It is now available to anyone for free.

What changed in this revision

Claim in the August version What the vendor page shows now
Model name deepseek-v4-flash Retired. Use deepseek-flash, which serves V4.1-Flash-1017
V4-Flash row of $0.44, $1.32 and $0.66 Output costs $0.60 off-peak and $1.20 at peak. Those three figures are gone
V4-Pro row of $1.32, $3.96 and $1.98 Output costs $1.98 off-peak and $3.96 at peak. The $1.32 figure is the peak cache-miss input rate, not output
V4-Pro peak output $3.96, up from $0.87 Correct
Price sources were a news site and a blog Both citations now point at DeepSeek and the DeepSeek GitHub repository

The old table mixed an input rate and an output rate in one row. A reader who compared the two models by that row compared the wrong numbers.

Try DeepSeek Free

DeepSeek V4 is free to chat with and the API is among the cheapest at the frontier. Run the same model that powers DeepSeek Harness.

Start Free — from $0
Share this article
AI Tools Insight may earn a commission from some links. Editorial independence is always maintained.