title: "One ¥99 Coding Plan, Three Protocols: Route Each Coding Client Correctly"
published: true
tags: ai, devtools, tutorial, programming
cover_image: https://cdn.acedata.cloud/uploads/blog-covers/f8235a19fe270b29ac57b54f12c50e34ad9feb0166ec43d0bf8f9844af9550b0.webp
---Disclosure: This article was prepared with AI assistance for Ace Data Cloud and discusses its Coding Plan. The implementation guidance is the point of the article; links to Ace Data Cloud are commercial product links.
The Ace Data Cloud Coding Plan starts at ¥99 for 30 days and provides one quota for supported Claude, GPT/Codex, GLM, Kimi, and DeepSeek model families. The practical benefit is not “one endpoint fits everything”: it is keeping clients such as Claude Code, Codex, OpenCode, Kimi Code, or OpenClaw while deliberately routing each one through the protocol it actually speaks.
The current Coding Plus tier contains 400 base Credits plus 100 promotional Credits (a 25% bonus) through November 30, 2026, 23:59:59 Beijing time. Credits are metered usage, not unlimited requests, and the plan is an API package rather than an official Claude or ChatGPT consumer membership.
Start with the client protocol, not the model name
A common configuration mistake is to choose a model first and then paste the same base URL into every client. A model may exist in the catalog while the client still fails because its request shape, streaming events, or tool-call format expects a different protocol.
Use this decision table first:
| Client or workflow | Protocol contract | Current API path | Why it matters |
|---|---|---|---|
| Claude Code-style agent | Anthropic Messages | /v1/messages | Native message blocks, tool use, and Anthropic-style streaming |
| Codex agent workflow | OpenAI Responses | /openai/responses | Response items and agent-oriented tool interactions |
| OpenCode/custom-provider clients | OpenAI-compatible Chat Completions | Family-specific .../chat/completions path | Broad provider compatibility, but capabilities vary by model and client |
Coding Plan currently exposes Chat Completions paths for several families, including /openai/chat/completions, /glm/chat/completions, /kimi/chat/completions, and /deepseek/chat/completions. That does not mean every client supports every model. Treat “model available” and “client/protocol compatible” as two separate checks.
Keep routing policy in one small file
If a team uses several coding clients, document the intended route rather than relying on tribal knowledge. The following YAML is an illustrative policy file, not a configuration format that every client imports directly:
coding_routes: repo_refactor: client: claude-code protocol: anthropic-messages path: /v1/messages model_env: ACE_REFACTOR_MODEL autonomous_patch: client: codex-cli protocol: openai-responses path: /openai/responses model_env: ACE_AGENT_MODEL quick_review: client: opencode protocol: openai-chat-completions path: /deepseek/chat/completions model_env: ACE_REVIEW_MODEL credentials: api_key_env: ACE_CODING_API_KEY base_url_env: ACE_CODING_BASE_URL This separation creates three useful boundaries:
- Credentials stay in environment variables rather than source control.
- Protocol is selected by client behavior, not by a convenient-looking model label.
- Model choice can change without rewriting the routing architecture, as long as the replacement is supported on that route.
The exact environment-variable names and base-URL syntax differ among clients, so translate this policy using the current Coding Plan integration guide.
Choose models by task after the route is valid
Once the protocol is correct, assign models according to the work rather than forcing one model into every stage.
For example:
- Cross-file refactoring: prioritize repository comprehension, reliable edits, and tool-use behavior.
- Autonomous patch loops: prioritize compatibility with the agent's response and tool-call lifecycle.
- Fast review or explanation: use a supported lower-cost route when the task does not need a long autonomous run.
- Second-opinion debugging: switch model families while preserving the same issue statement and acceptance tests.
This is where one multi-model quota can simplify budgeting. You are managing one 30-day Coding allowance while retaining task-level model choice, rather than assuming that each model family requires a separate coding workflow. The live catalog and credential scope remain the source of truth.
Validate one route at a time
Do not begin with a large repository and three clients. Use a small, reversible validation sequence for each route:
- Confirm that the client reads the expected base URL and placeholder-backed API key.
- List or select a model that the current Coding credential can access.
- Ask for a short response without tools.
- Ask the client to inspect one harmless file.
- Request a tiny patch and review the diff before accepting it.
- Only then enable longer context, tool calls, or parallel agents.
These are validation steps, not claims that the sample above was executed in your environment.
Failure modes that look like model problems
A 404 or HTML response
This usually indicates a wrong base URL or path. Check whether the client appends /v1/messages, /responses, or /chat/completions itself before adding the path manually.
Authentication succeeds but the model is rejected
The key can be valid while the requested model is outside its current catalog or scope. Re-read the model list exposed to the Coding credential; do not infer availability from a marketing family name.
Text works but tools or streaming fail
That is a strong protocol signal. A “compatible” text response does not prove that tool calls, structured output, or streaming event types match what the agent expects. Return to the protocol matrix rather than repeatedly changing model names.
Usage disappears faster than expected
Credits cannot be converted into a fixed number of prompts. Context size, generated output, cache behavior, tool loops, and parallel agents all affect consumption. Measure with a real task and keep acceptance criteria fixed when comparing routes.
The design trade-off
A shared plan reduces credential and budget fragmentation, but protocol differences do not disappear. The maintainable approach is explicit routing: client → protocol → supported model → task. It is slightly more configuration up front, yet it makes failures easier to classify and model changes less disruptive.
You can inspect the current tiers and model scope on the Coding Plan page. As of November 8, 2026, the verified entry tier is ¥99 for 30 days; promotional Credits and available models should always be rechecked before purchase.
Sources
This is an independently structured implementation article, adapted from and cross-checked against:
- Ace Data Cloud, Is the ¥99 Coding Plan Right for You? (October 2, 2026)
- Ace Data Cloud, One Coding Plan for Claude Code, Codex, Kimi Code and OpenClaw (October 10, 2026)
- Coding Plan Integration Center
Package values and API paths were rechecked against Ace Data Cloud's public Coding Plan service data on November 8, 2026.
