Your coding tools. Our models. One flat monthly bill.
Keep Cline, Roo Code, Continue, Claude Code or Aider exactly as they are. Point them at one endpoint and code all month on GLM, DeepSeek and Kimi — with a hard spending ceiling and no per-token surprises.
$export OPENAI_API_BASE="https://api.coding.ecohash.com/v1"$export OPENAI_API_KEY="eco_****************"$aider --model openai/GLM-5.2›connected · GLM-5.2 · openai-compatible›prompt cache · 53.5% of input tokens hit✓plans from $9/month · 1 credit = $0.01, always$Change a base URL. Keep everything else.
We speak both the OpenAI and Anthropic wire formats, so the tools you already use connect without a plugin, a proxy, or a migration.
Create a key
Sign in to the console and create an API key. It starts eco_ and is shown once.
Point your tool at us
Paste our base URL and your key into the tool you already use. Nothing to install.
Carry on coding
Your tool works exactly as before — the bill just stopped being per token.
# Provider: OpenAI CompatibleBase URL https://api.coding.ecohash.com/v1API key eco_****************Model GLM-5.2
Cline · Roo Code · Kilo Code · Continue · Aider
$export ANTHROPIC_BASE_URL=\ "https://api.coding.ecohash.com/anthropic"$export ANTHROPIC_AUTH_TOKEN="eco_..."$claude
Anthropic-shaped endpoint — no /v1 in the URL
$curl https://api.coding.ecohash.com\ /v1/chat/completions \ -H "Authorization: Bearer $ECO_KEY" \ -d '{"model": "GLM-5.2", ...}'
A JSON reply means you're connected
Per-tool walkthroughs, limits and error codes live in the setup guide
Models we name, not tiers we hide
You call a model by name and we serve that model — no silent rerouting, no “smart” downgrades. One key and one budget across the whole lineup.
GLM-5.2
Zhipu AI
Frontier coding flagship — the workhorse for agentic edits and long sessions.
DeepSeek-V4-Flash
DeepSeek
Strong reasoning at aggressive throughput — great for refactors and reviews.
Kimi-K2.6
Moonshot AI
Long-context specialist that keeps big repos and long threads in its head.
Embeddings & rerank
Same key, same endpoint
Index a private repo without your source ever leaving this endpoint.
- qwen3-embedding 0.6b / 4b / 8b
- jina-embeddings v3 / v4
- bge-reranker-v2-m3
The exact model list on your plan is data, not marketing — see it any time on your plan page in the console. If a named model’s upstream blips, a fallback serves the request and the response tells you which model actually answered.
A credit is a cent. Forever.
If you can’t answer “why did my credits go down so fast?” without filing a ticket, the pricing has failed. Ours is built so you never have to ask.
A credit is a cent. The exchange rate never moves, so a balance always means something.
That's $25 of face value on a $19 plan — every tier's allowance is deliberately bigger than its price.
Measured in production on our flagship. Cached input costs a fraction of uncached — and coding tools resend the same context every turn.
- Hard cap — exhaustion is a stop, never an overdraft
- No silent downgrades to a cheaper model
- Top-ups carry over between months, valid 12 months
- Your plan page shows exactly where every credit went
Fast from Singapore, fair for teams
The plan is more than a price — it’s infrastructure placed where you work and rules designed so one person’s burst never becomes everyone’s outage.
Served from Singapore
An agentic session makes many sequential calls — latency is paid on every one.
Serving from ap-southeast-1 removes the 350–400 ms trans-Pacific round trip from each request, and lets us offer the local payment methods the region actually uses.
Fair for teams
Seats pool their credits; pace limits stay per member.
One teammate's giant refactor can't freeze the rest of the team. Usage is attributed per member, so you can see who spent what this period.
Server-side tools built in
Web search, web fetch and vision work from the gateway — nothing to install.
Search costs 1 credit per successful call (failures are free), fetch respects robots.txt and caches results, and vision bills as ordinary tokens on the serving model.
Private repos stay private
Embeddings and reranking run on the same key and endpoint as your chat models.
Point your tool's codebase index here and your source never fans out to a third-party embedding service.
Flat monthly, hard ceiling
Starter
$9/month
1,000 credits/month
- Up to 1 concurrent request
- Every model on the lineup
Pro
popular$19/month
2,500 credits/month
- Up to 2 concurrent requests
- Every model on the lineup
Max
$49/month
6,000 credits/month
- Up to 4 concurrent requests
- Up to 5 seats, pooled credits
Top-ups from $10 carry over for 12 months. Per-model rates and a live burn-down live in your console — the public page only ever promises what the plan includes.
Questions, answered straight
Do I have to change my workflow or install anything?
No. You keep the tool you already use — Cline, Roo Code, Kilo Code, Continue, Claude Code, Aider or anything OpenAI-compatible — and change two settings: the base URL and the API key. That's the whole migration.
What happens when my credits run out?
Requests stop with a clear error (HTTP 402) — never an overdraft, never a silent switch to a cheaper model. Buy a top-up or wait for the monthly renewal. Auto top-up exists but is off by default and requires an explicit dollar cap.
Do top-up credits expire?
Top-ups convert at face value ($10 → 1,000 credits), stay valid for 12 months, and carry over between months. They're only spent after your monthly allowance is used up — and unlike the allowance, purchased credits can be spent on any model.
Why did a request get refused when I still had credits?
Besides the monthly allowance there are two independent limits: a spending pace (a ~5-hour and a 7-day window that refill continuously — just wait) and a per-model-group share cap (switch to a model in another group). Your plan page shows all three separately so you can tell which one you hit.
Which model answers when I ask for sonnet or opus in Claude Code?
Your administrator maps each alias to a model on the plan. The response always reports which model actually served the request — including when a fallback steps in during an upstream outage.
How do teams work?
Plans are per seat. Seats pool their credits into one budget, but pace limits and concurrency are per member, so one person's marathon session can't starve the rest. Usage is attributed per member in the console.