DeepSeek-native coding. Grok-class execution.
A full-screen terminal coding agent with safe edits, cache-aware sessions, and parallel execution built around DeepSeek models.
Quick start · Why DeepSeek Build · How it works · DeepSeek Harness · Benchmarks · Documentation · Contributing
Install from npm, add your DeepSeek API key, and open the TUI.
npm 12.0.0 and newer block dependency install scripts unless you opt in. Without the flag below, npm reports success and does not install the agent. npm 11 and older install with or without the flag.
npm install -g --allow-scripts=@innocarpe/deepseek-build @innocarpe/deepseek-build
deepseek-build setup
deepseek-buildAllow the package once for later global installs, then the plain command works:
npm config set allow-scripts=@innocarpe/deepseek-build --location=user
npm install -g @innocarpe/deepseek-buildThe registry install requires Node.js 18 or newer and uses a prebuilt binary when a matching release asset is available. It does not require Rust on that path; see the npm installation guide for platform and source-fallback details.
deepseek-build is the primary command. dsb is the fully supported short
alias with the same behavior and full Semantic Version:
deepseek-build --version
dsb --versionKeep the install current with the same command on either name:
deepseek-build updateIt asks the published feeds for the latest version and replaces an npm-global install in place. When the install is already current it says so and changes nothing, and it never replaces an install with an older published version.
If the installer reports that the product bin directory is not on PATH, add
it before launching:
export PATH="$HOME/.deepseek-build/bin:$PATH"| Capability | What it means |
|---|---|
| DeepSeek-native | DeepSeek API defaults, Flash/Pro routing, reasoning effort, and a DeepSeek-branded TUI. |
| Safe edits | Version-bound snippet editing and fail-closed workspace permissions instead of silent whole-file replacement. |
| Long-session economics | Session changes append after the cached history instead of rewriting it. A cache miss names the assembled document that moved, the session logs its cumulative cache total, and a request that diverges from the session log fails the turn. |
| Wall-clock throughput | Parallel tools, background shell jobs, subagents, and opt-in worktrees run beneath the safety and cache layers. |
| Durable sessions | Resume the most recent full-screen session or address a saved session directly. |
The result is a coding agent that keeps the speed of a Grok-derived execution engine while making DeepSeek-specific cost, edit, and permission rules part of the product path.
# Open the full-screen TUI
deepseek-build
# Resume the most recent TUI session
deepseek-build --resume
# Continue the most recent TUI session (Claude Code-style -c)
deepseek-build -c
# Run one non-interactive turn
deepseek-build run "Explain the architecture of this repository."
# Use the trusted local coding profile
deepseek-build --dogfood--dogfood allows writes inside the current workspace and enables shell
execution under policy. Writes and deletes outside the workspace remain denied.
For the short command, replace deepseek-build with dsb in any example.
At or below 60 columns, a submitted prompt folds to at most two rows and drops
the decorative ❯ and blank pad rows; wider windows keep the three-line
budget. On that phone-width pane the balance, cache hit, model and permission
share one row under the input box — cost and cache on the left, model and
permission on the right. Wider panes keep the model on the box border and the
balance on its own row. The full-screen view has a configurable status line at the bottom. On
the official DeepSeek API, DeepSeek V4.1 Flash (deepseek-flash) accepts
attached images directly, and text-only models keep images off the wire with
the on-disk fallback.
The default Flash wire ID is deepseek-flash, which DeepSeek currently serves
as DeepSeek V4.1 Flash 1017. OpenRouter setup uses
deepseek/deepseek-v4.1-flash-1017. Existing product homes retain their local
[model.deepseek-v4-flash] slot; DSB updates its old wire ID on launch while
preserving custom model choices (ADR 0015).
Interactive setup asks how to reach the DeepSeek models — the DeepSeek API or
OpenRouter — then stores that provider's API key in
~/.deepseek-build/credentials.json with mode 0600:
deepseek-build setup
deepseek-build setup --provider openrouter --api-key "$KEY" # non-interactive
deepseek-build auth status
deepseek-build auth logoutOpenRouter keeps prompt caches separate for each upstream provider. DSB sends
{"order":["relace","deepinfra"],"allow_fallbacks":false} by default, which
tries Relace first and limits routing to those two providers. Override it per
model with OpenRouter's provider object; a configured object replaces the
default entirely. Set it to false to disable routing:
[model.deepseek-v4-flash]
openrouter_provider = { order = ["relace", "deepinfra"], allow_fallbacks = false }
# openrouter_provider = { only = ["deepinfra/fp8"] }
# openrouter_provider = falseFor CI or another non-interactive environment, set DEEPSEEK_API_KEY; the
environment variable takes precedence over a DeepSeek credentials file.
Line mode (run / chat) uses the DeepSeek API only. Product
configuration, credentials, sessions, and user skills live under
~/.deepseek-build/ by default.
Source installation is intended for contributors and unsupported release
platforms. It requires Rust 1.94 or newer plus protoc or DotSlash; the first
agent build can take several minutes.
git clone https://github.com/innocarpe/deepseek-build.git
cd deepseek-build
./scripts/install.sh
deepseek-build --version
dsb --versionnpm install in this checkout does not install deepseek-build or dsb.
It does not download a prebuilt and it does not compile. Use
./scripts/install.sh here, or
npm install -g --allow-scripts=@innocarpe/deepseek-build @innocarpe/deepseek-build
for the registry package. npm 12 does not install the agent without that flag.
See the installation guide for Cargo and custom prefix options.
deepseek-build | dsb
│
▼
product launcher ── auth · config · model routing
│
▼
deepseek-build-agent ── full-screen TUI · tools · sessions
│
▼
DeepSeek API
Three layers have explicit ownership. Higher-throughput machinery cannot bypass the edit, permission, or cache contracts beneath it.
| Layer | Source | Owns |
|---|---|---|
| L1 | Deep Code CLI | Snippet-safe edits, skills as context, and side-effect permissions. |
| L2 | Reasonix | Stable-prefix economics, Flash/Pro behavior, and tool-call repair. |
| L3 | Grok Build | The base runtime, TUI, parallel tools, subagents, background work, and worktrees. |
Reading DeepSeek's official harness added no layer to this map: the append rule and the request-versus-log check came from that evidence, described in the next section.
The normative conflict rules live in the harness philosophy, with the complete system map in SYSTEM_ARCHITECTURE.md.
dsh is DeepSeek's official open-source harness. It is not this product's architecture, and it is not a fourth layer: DeepSeek Build stays a Rust full-screen TUI, with edits owned by Deep Code, the cache contract owned by Reasonix, and execution owned by Grok Build. dsh is evidence for how one family behaves — a cached prefix survives a session change, and a request that does not match the session log is not sent.
| What changes in a session | What the product does |
|---|---|
| A later change of the stable system body | Earlier system messages stay byte-for-byte; the new body is appended after them. The model treats the most recent system message as the current body. |
| tools, skills, environment, or project instructions inside that body | The same append. The wire tools array is the full list on every request. There is no tool add/remove history message; Chat Completions has no field for one. |
| A cache miss | The turn names which assembled document moved, as prefix_change=. When the epoch moved and every document hash matches, it says unattributed instead of inventing a cause. |
| The session's cache | Once any response carried cache fields, every turn logs one cache_session= line. Hits and misses are token sums; a response without cache fields is unreported and is not counted as a miss. deepseek-build run and the REPL store the totals with the session, so a resume continues them. The full-screen status chip stays the last turn's ratio, the full-screen counter is not written to the session file, and a new process starts it at zero. |
| A request that diverges from the log | The turn fails before sampling. An item that cannot be serialized fails closed too. |
| deny | A later allow does not replace a deny. |
Deliberately not taken: a plugin host, agent teams, sandbox escalation,
request-series bookkeeping (initial / resume / change), and the
Anthropic Messages transport — this product speaks DeepSeek Chat Completions
(ADR 0005).
Already present, so not ported again: spill of an oversized tool result (head, tail, and a path to the rest), compaction aligned to the warm prefix, and a snippet staleness guard stricter than dsh's path-scoped token. These are not new features taken from dsh.
Read more: dsh research note · what this work changed · design sources.
On 2026-10-15, dsb itself ran two public agent benchmarks — a Linux x64
source build driving DeepSeek V4.1 Flash (deepseek-v4-flash) with
thinking on, reasoning effort max, and one attempt per task (-k 1). Both
went through their official runners: Harbor for Terminal-Bench 4.0 and Pier
for DeepSWE v1.1.
| Model · harness | Terminal-Bench 4.0 | DeepSWE v1.1 |
|---|---|---|
| dsb + DeepSeek V4.1 Flash (max) — measured here | 33.3% (2 of 6 tasks) | 63.6% (7 of 11 tasks) |
| DeepSeek V4.1 Flash (max) — reported | 26.8% (announced 31.2%) | 74.2% |
| Claude Haiku 5.5 (max) — reported | 32.8% (announced 39.2%) | not published (83.7% on SWE-bench Multilingual) |
| GPT-6 Luna (max) — reported | 12.6% (announced 16.4%) | 66.6% |
Reported figures are public numbers from the sources linked below; announced
marks the model vendor's own figure next to the evaluation number. For scale,
the same Terminal-Bench 4.0 board carries Claude Sonnet 5.5 at 63.6% and
Claude Opus 5.5 at 59.6%.
Harness is part of the score. On Terminal-Bench 4.0, Artificial Analysis ran both lightweight models with the same mini-swe-agent harness (DeepSeek V4.1 Flash 26.8%, GPT-6 Luna 12.6%), and Vals AI measured 19.7% and 13.6% in a shared environment. On DeepSWE v1.1, DeepSeek V4.1 Flash is reported at 65.6% under Codex and 69.8% under Claude Code. List prices per 1M tokens (input/output): Claude Haiku 5.5 $0.10/$0.50 (5× on prompts above 100k tokens), GPT-6 Luna $0.10/$0.50, DeepSeek V4.1 Flash $0.30/$1.20 (half price in off-peak hours).
Our row is an early directional signal, not a like-for-like comparison with those boards: it covers a task subset (6 of 66 Terminal-Bench 4.0 tasks, with GPU and network-isolated tasks excluded; 11 of 113 DeepSWE v1.1 tasks), one attempt per task, and an environment that moved mid-run from Modal to a local Docker Desktop (macOS, amd64 emulation) when Modal's spend limit was reached.
Sources: aievals.app — Terminal-Bench 4 · decode.agency — AI model comparison · DeepSeek V4.1 Flash model card
| Start here | Use it for |
|---|---|
| User guide | Installation, setup, daily use, and the complete feature index. |
| First-run setup | API keys, credential precedence, and headless setup. |
| Sessions | Full-screen resume and line-mode session storage. |
| Permissions | Interactive asks, headless denial, and workspace boundaries. |
| Subagents · background tasks · worktrees | L3 execution surfaces. |
| Known limitations | Current packaging, live-smoke, and platform boundaries. |
| Product SSOT | Which artifact wins when product documents disagree. |
cargo build -p dsb-cli
cargo test --workspace
./scripts/check-semver.sh
./scripts/test-owner-bar.shThe root Rust workspace covers the product crates. Avoid vendor-full Cargo runs for everyday checks; the owner-bar scripts use the bounded product path.
For the crate map, see crates/README.md. For the repository map and documentation ownership, start at docs/README.md.
Read CONTRIBUTING.md before making changes. All meaningful work lands through a focused PR with an atomic Conventional Commit, an existing kind label, honest test evidence, and the review narrative defined in the PR authoring guide.
DeepSeek Build is available under the Apache License 2.0. Vendored and third-party code retains its original licensing; see NOTICE.