Trenchwork Coder

A Claude Code–class terminal coding agent you fully control — for a fraction of the price. Live at ero.solar.

Trenchwork Coder is an Ink-rendered CLI coding agent (the same terminal-UI engine Claude Code uses) that reads code, edits files, runs commands, and searches the web — rendered with the exact ⏺ / ⎿ transcript shape, permission modes, and colored diffs you know from Claude Code. The difference: it runs on DeepSeek v4 Pro at max thought with your own API keys, so a month of heavy daily use costs less than a single Claude Max 20× subscription, and it does the work you ask without an approval-nag or hosted-refusal layer in the way.

npm install -g @trenchwork/coder
trenchwork          # or: deepseek

Why Trenchwork

A fraction of the cost

DeepSeek v4 Pro lists at $0.435 / 1M input tokens and $0.87 / 1M output tokens (cache-hit input is ~$0.0036/1M — a ~99% discount on the stable prompt prefixes a coding agent reuses constantly). Tavily web search has a free 1,000-credit/month tier that covers normal use.

In the maintainer's own testing, a full month of heavy daily use — the equivalent of maxing a Claude Max 20× plan every week — came to under $50 across DeepSeek + Tavily combined. At list rates that pencils out to roughly $10–30/month for 30–50M tokens of real coding-agent work. Claude Max 20× is $200/month flat.

Numbers are list pricing as of June 2026 (DeepSeek api-docs pricing; Tavily pricing; Anthropic claude.com/pricing/max). The "<$50/mo" figure is the maintainer's measured spend, not a list-rate guarantee — your mileage depends on token volume and cache-hit rate.

Built for long horizons (ultracode, always on)

Ultracode is baked in — there is no toggle. Every session runs DeepSeek v4 Pro on its maximum thinking budget, with a system directive tuned for substantial, multi-step work:

• Phased execution — research → verify load-bearing facts → design → build the whole thing → verify the result against the real artifact.
• Living TODO plan — for large refactors/migrations, it enumerates every unit of work up front, marks progress as it goes, and re-derives what's left so it never loses the thread of a long task.
• Repo-wide multi-file changes — it finds every affected site first, edits each, then re-searches to confirm nothing was missed.
• Doesn't stop early — the task is done only when every TODO item is complete and the build + tests pass.
• Adversarial self-check — an always-on critic reviews the finished answer and high-impact tool calls (toggle with /adversarial).

Trivial turns stay direct — ultracode does not orchestrate a greeting.

Looks and works like Claude Code

• ⏺ action bullets and ⎿ result summaries — every assistant turn and tool call.
• Colored diffs after every write/edit — additions in green, removals in red, context dim, with line numbers (formatDiffClaudeStyle).
• Shift+Tab permission modes — cycle default → accept edits → plan. Plan mode is genuinely read-only: write/edit/bash/commit are blocked at the tool runtime so the agent investigates and proposes a plan before changing anything.
• Rounded input box, dim hint line, animated working spinner (✻ Synthesizing… · esc to interrupt), and a minimal status line — no emoji chrome. Follow-ups typed while the agent works are accepted live into a transient queue and processed at the next turn boundary (Claude Code parity; no waiting for the full original task).

The website's example sessions are rendered from the exact strings the binary emits (src/shell/toolPresentation.ts + ChatStatic), so what you see at ero.solar is what the installed CLI prints.

Install & run

npm install -g @trenchwork/coder

Exposes three synonyms on PATH — trenchwork, trenchwork-coder, deepseek — pick whichever you like. The bin is interactive-only: running it launches the Ink shell. There are no argv flags, no print/headless mode, and a non-TTY invocation (trenchwork < /dev/null, piping, CI without a PTY) fails fast on purpose.

trenchwork           # launch the Ink shell

Keys

Bring your own DeepSeek key (no login, no account):

• In-shell: /key sk-...
• Or env: export DEEPSEEK_API_KEY=sk-...

Optional: set a Tavily key for web search (TAVILY_API_KEY) — the free tier covers light use; without it the agent falls back to direct page fetches.

Slash commands

The surface is intentionally lean — the model is locked to deepseek-v4-pro on max thought, so there's nothing to switch:

Keyboard shortcuts

Tools the agent can call

Read · Write · Edit · MultiEdit · Bash · Grep · Glob · Web search + fetch · TodoWrite · Memory (persistent across sessions) · Notebook edit · Skills · Git · human-in-the-loop prompts. Pre/post-tool hooks load from .trenchwork/settings.json.

Authorization scope

Trenchwork Coder runs without the approval and refusal layers hosted assistants add — it executes the commands you give it, including destructive shell, sudo, and security-research tooling that mainstream agents decline. Use it only on systems you own or are explicitly authorized to test. Its security does not rest on hiding the source: the design is open to inspection (the Project Glasswing–inspired principle), and what's protected is protected by small, rotatable secrets in the OS keychain / secret store — not obscurity.

Architecture

src/
  core/          agent loop, tool runtime, permission modes, adversarial verifier,
                 context manager, secret store, hooks, preferences, sessions
  config.ts      profile + system prompt assembly (ultracode directive baked in)
  providers/     DeepSeek (OpenAI-compatible) provider, thinking forced to max budget
  tools/         Read / Write / Edit / MultiEdit / Bash / Grep / Glob / Web / Todo /
                 Memory / Notebook / Skill / HITL / Git  (+ diffUtils for colored diffs)
  shell/         toolPresentation — the ⏺ / ⎿ transcript formatters
  ui/ink/        Ink renderer — App, Prompt, ChatStatic, StatusLine, Menu, controller
  headless/      interactive shell + CLI bootstrap
GLASSWING.md     transparency-over-secrecy security principle
CLAUDE.md        contributor conventions (test discipline, research-before-custom-code)

Build & test

npm install
npx tsc                                          # build
npm test                                         # full jest suite (build + tests)

Verification runs in two places, per the Glasswing "checkable by anyone" pillar: a fast local pre-push hook and a public GitHub Actions workflow (.github/workflows/ci.yml) on every push and PR. Install the hook once per checkout:

git config core.hooksPath scripts/git-hooks

Every fix ships with a test that fails before and passes after — UI changes are end-to-end tested against the real binary under a PTY where one can fork, and the rendered output is asserted on real stdout.

Contributing

Read CLAUDE.md first — it documents the testing discipline and the "research before custom code" rules this repo enforces.

Contact

Bo Shang — building Ero.Solar.

• Email: bo@ero.solar
• GitHub: @Aroxora
• X: @ghidradragon
• YouTube: @trenchworkai
• LinkedIn: bo-shang
• Instagram: @ghidra.dragon

License

Proprietary — see LICENSE. © 2025 Trenchwork AI. All rights reserved.

trenchwork-coder is an independent project and is not affiliated with or endorsed by Anthropic. Project Glasswing is an Anthropic initiative; this repo is only inspired by it. "Claude Code" is referenced for comparison only.