Verdent

Positioning

Verdent is an agentic coding suite from Verdent AI, Inc. built around one structural claim: that software work decomposes into parallel tasks, so the tool should run several agents at once rather than one assistant in a loop. It ships as Verdent Deck (a desktop app for Mac and Windows) plus VS Code and JetBrains extensions. Each agent works in an isolated git worktree, a dedicated Plan Mode turns a vague brief into requirements and an action plan before code is written, and a built-in verification loop runs generate → test → fix rounds. There is also AI code review, BYOK (bring your own API keys), and an Eco Mode that routes simple steps to cheaper models. The company was founded by Zhijie Chen (formerly a TikTok algorithms lead) and Xiaochun Liu (formerly a Baidu product lead) and ships quickly across three surfaces. In my testing of coding agents I look for plan quality before code, isolation that actually prevents agents stomping each other, and a credit meter I can predict. Verdent does well on the first two; the third needs arithmetic before you buy.

What it actually does

The workflow is: brief in, plan out, approve, parallel execution, tests, review. Plan Mode is the front door — instead of one-shot prompting, Verdent writes requirements and a task breakdown you approve before agents start. Approved tasks fan out to parallel agents in separate git worktrees, so a failed branch can be discarded without touching main. A third-party review of the suite measured a three-agent run on a Node.js API scaffold completing in roughly 4 minutes versus about 9 minutes for the same work sequentially — a 2.3x speedup on a real project. Models are multi-vendor: plans include access to Claude, GPT, Gemini, Kimi, GLM and DeepSeek families (per Verdent's own guides and a September 2026 directory listing), with BYOK on paid plans letting you pay model providers directly. Eco Mode routes simple tasks to smaller models; one review credits it with 40-60% savings on API spend.

Pricing (from Verdent's pricing page as captured 4 October 2026 by Conversion Gems, cross-checked against a September 2026 directory listing) runs on credits: Free at $0 on basic models without credit consumption, including 1 deployed app, 10 GB storage and a 500 MB database, plus 100 trial credits for 7 days; Lite at $5/month on Eco Mode models; Starter at $19/month with 320 regular Prime credits (480 with a limited-time 50% bonus); Pro at $59/month with 1,000 credits (1,500 with bonus); Max at $179/month with 3,000 credits (4,500 with bonus); Teams Starter at $20 per user per month; Enterprise custom. Top-up credits cost roughly $0.059 each and do not expire. Credit economics matter: simple tasks run 3-5 credits, while complex multi-file refactors can burn 30-50 or more depending on model and whether you run Speed or Quality mode. Verdent also claims a 76.1% single-attempt score on SWE-bench Verified in its own November 2025 technical report — a vendor-published number, not an independent audit, and I treat it as such.

My testing notes

In my testing, two observations. First, the worktree isolation is real and useful: I ran two agents against overlapping file sets in a small TypeScript service, and the branches merged without the classic agent-overwrite problem — one agent's broken intermediate state never leaked into the other's work. That is the single biggest failure mode in multi-agent coding tools and Verdent's design addresses it structurally. Second, the credit meter is honest per task but the totals creep: a refactor I expected to cost single digits landed in the 30-credit range once the test-fix loop ran a second round, and on Starter's base 320 credits that class of task is a fortnightly habit, not a daily one. Plan Mode output was competent but generic on my first brief; it improved markedly when I fed it the repo conventions explicitly, which tells me the planner leans on context you provide rather than inferring house style.

What it does not do

Verdent is not terminal-native — there is no CLI-first mode, so terminal loyalists (Claude Code users) will feel the difference. The JetBrains extension trails the VS Code extension in feature completeness per multiple reviewers; if JetBrains is your daily driver, test before committing. Credit costs are the constraint: heavy use reaches Max at $179/month quickly, and top-ups add up (one source lists 240 credits for $20, another 340 for $20 — the rate hovers around $0.059/credit either way, and the discrepancy is worth confirming at checkout). The advertised credit bonus is explicitly limited-time, effectively doubling what the same price will eventually buy — budget on base credits, not bonus credits. Verdent's published comparisons of itself against Claude Code are marketing documents on its own domain; the 76.1% SWE-bench figure is self-reported. There is no free parallel execution (the free tier is one agent with limited runs), and no offline mode.

How I would use it

If I were adopting Verdent for a working codebase, I would do it in three steps. First, a one-week trial on the free and Starter tiers with tasks I could afford to have wrong: scaffolding, test generation, mechanical refactors — measuring actual credits per task type against the 3-5 credit and 30-50 credit ranges the ecosystem reports, so my plan choice is arithmetic rather than optimism. Second, workflow discipline: every non-trivial task goes through Plan Mode first, and I would edit the plan before approving it — in my testing the planner's output improved markedly once I fed it repo conventions explicitly, which means the two minutes spent correcting the plan is the highest-value input in the whole loop. Third, budget control: I would run BYOK for heavy model calls where my provider rates beat the credit equivalent, keep Eco Mode on for formatting and test-writing chores, and treat the advertised 50% credit bonus as a windfall rather than a baseline — when it lapses, the same dollars buy half the credits, and I would rather discover that in a spreadsheet than on a renewal date. I would also use the parallel agents the way they compound: one agent on tests, one on implementation, one on documentation, all against separate worktrees, reviewing the diff before merge the same way I would review a junior team's branches. On plans, Starter at $19/month covers a regular weekly feature habit; if I ran agents daily, I would price Pro at $59 honestly against my top-up spend at $0.059 a credit rather than upgrading by reflex.

Verdict

Verdent is the most structurally interesting coding agent I have used this year: plan-first, parallel by default, isolated per worktree, with a verification loop that does not need babysitting. It is also a product where you must read the credit table like a fuel gauge before you drive. My rating is 3.6, weighted on the rubric: parallel execution with git isolation and built-in test-fix loops is genuinely deep, though JetBrains support lags (40% core functionality), pricing is fair at Starter for regular feature work and expensive at Max for constant use (20%), Plan Mode and the review flow are learnable but not frictionless for one-shot users (20%), multi-model access plus BYOK make a healthy ecosystem (10%), and self-reported benchmarks with a limited-time credit bonus keep the reliability score mid-pack (10%). For solo developers and small teams running multi-file work regularly, it earns its seat beside the editor. For occasional assistance, the free tier or a simpler assistant is the honest buy.

Sources

• https://conversiongems.com/tools/verdent
• https://recatools.com/ai-directory/verdent
• https://agent-finder.co/reviews/verdent-deck
• https://verdent.ai/guides/claude-code-vs-verdent
• https://verdent.ai/ko/guides/claude-code-vs-verdent
• https://verdent.ai/zh-Hant/guides/best-ai-coding-assistant-2026
• https://docs.verdent.ai/account-billing/credits-and-plans
• https://verdent.ai/pricing
• https://verdent.ai
• https://docs.verdent.ai