Quick Answer
Running DeepSeek V4 Pro inside DeepSeek Harness for free is entirely possible. TeamoRouter's deepseek-v4-pro-free model ID provides 200 requests per account per day, reset daily, over an OpenAI-compatible endpoint (https://api.teamorouter.com/v1).
Three lines to configure:
export DEEPSEEK_API_KEY="sk-teamo-your-key"
export DEEPSEEK_BASE_URL="https://api.teamorouter.com/v1"
export DEEPSEEK_DEFAULT_MODEL="deepseek-v4-pro-free"
Launch with npx @deepseek-ai/dsh web and dsh runs on free V4 Pro. Below: whether 200 requests is enough for agent work, and how to make it last.
The Key Insight: Quotas Count Requests, Not Tokens
This is the easiest thing to misunderstand and the most consequential. The free quota counts requests, not tokens.
- One "request" = one complete model call (one turn). The model can read 100K+ tokens of context, reason, and produce a long diff — and that is one unit against the 200.
- Inside an agent loop, each tool round-trip is its own request: the model says "read this file" → the tool returns content → the model says the next step. That round-trip is 1 request. A gnarly multi-step refactor can burn 10–15.
So 200 requests is not a pitiful "200 tokens" — it is 200 rounds of model decision-making. That's a comfortable day for a solo developer, but letting the agent thrash in circles will drain it fast.
What 200 Requests Buys You in dsh
From a week of normal use, 200 requests roughly covers:
| Task | Approx. requests |
|---|---|
| Read a file, explain a function, write a commit message | 1–3 each |
| Single-file edit (read + change + verify) | 3–5 each |
| Multi-file refactor (cross-file + run tests) | 10–15 each |
| Long agent loop (build feature + self-test + fix) | 20–40 each |
Bottom line: single-file daily edits are fully covered — a dozen or more per day; heavy agent loops cover 3–5 hard tasks a day. Past that, paid pricing is cheap enough not to hurt.
Make It Last: Route by Difficulty, Not Habit
dsh can hold both the Pro and Flash free tiers, each 200/day. The sensible split:
deepseek-v4-pro-free(200/day) — multi-file reasoning, long agent loops, anything where a wrong answer costs you a redo.deepseek-v4-flash-free(200/day) — the volume work: explain-a-function, commit messages, single-file edits, quick lookups.
The classic mistake: burning pro-free on trivia before lunch, then finding it drained when you need it for real problems. Route by difficulty, not by muscle memory — default light work to Flash, and reserve Pro for the genuinely hard turns.
When the Quota Runs Out: A One-Word Switch
Once a free tier hits zero, switch the model ID to the paid version — paid models have no account-level rate limit:
export DEEPSEEK_DEFAULT_MODEL="deepseek-v4-pro" # drop the -free
Paid pricing: deepseek-v4-pro is $1.74/$3.48 (input/output per 1M tokens), deepseek-v4-flash $0.14/$0.28 — flat, no peak windows. Even on paid, a heavy day typically stays under a few dollars.
FAQ
Does the quota reset daily or accumulate?
It resets daily. It's per-account, not per-session; multiple sessions share the same 200-request pool.
Is deepseek-v4-pro-free weaker than deepseek-v4-pro?
Same capability, different rate limits. The -free variant caps at 200/day; the paid one has no account-level limit. deepseek-v4-pro currently resolves to DeepSeek-V4-Pro-0813.
Does it work for teams?
The free quota is per account with no team-level pooling. Each member needs their own account and their own 200. For team-level throughput, go straight to the paid model.
How do I check my remaining quota?
See the TeamoRouter rate-limits docs. At the limit, requests return a rate-limit error — switch to the paid model.
How do I use both free tiers in dsh?
Switch deepseek-v4-pro-free and deepseek-v4-flash-free in the Web UI's model picker, or set DEEPSEEK_DEFAULT_MODEL and switch per session. Light work on Flash, hard work on Pro is the most quota-efficient combo.
Register at TeamoRouter, point dsh at it, and take the 200 free V4 Pro requests daily before paying flat rates past that.