OpenClaw: Tested Installation and Fishapi Cloud Plugin Configuration

Obtain Fishapi Cloud API Key

  1. Open Console → Application List and log in.
  2. Select the application based on the billing method: enter Coding Application when using the Coding Plan; select a general/service application with corresponding service permissions for pay-as-you-go usage. Coding packages and general balances are managed separately.
  3. Copy the application's API Key / API Token, do not copy the Application ID or platform- management Token. The API Key field below should only contain the key itself, without the Bearer prefix.
  4. Confirm that the application is valid and has quota. Active and free quotas are displayed in real-time on the console.

Use the same Key to query the model directory:

export ACEDATACLOUD_API_KEY="YOUR_ACEDATACLOUD_API_KEY"
curl --fail-with-body https://api.fishapi.cloud/v1/models \
  -H "Authorization: Bearer $ACEDATACLOUD_API_KEY"

In PowerShell, the environment variable is written as $env:ACEDATACLOUD_API_KEY = "YOUR_ACEDATACLOUD_API_KEY"; when using the built-in curl on Windows, write curl.exe. /v1/models is the entry point for model discovery; the presence of a model in the directory does not mean the current Key has corresponding API permissions. You also need to verify the selected protocol, application scope, and client capabilities.

The MODEL_ID below must be replaced with the specific model ID. Do not submit the real Key to Git, paste it into logs, or take screenshots.

Version and Runtime Environment

The tested combination is OpenClaw 2026.9.2 + @acedatacloud/openclaw-provider 2026.5.34 + Node.js 26 (macOS isolated environment). The Node requirement for the OpenClaw main program is higher than that of the plugin itself; the old tutorial Node 20+ is no longer applicable.

node --version
npm install -g openclaw@2026.9.2
openclaw --version

This is the version for this validation, not a requirement to be permanently fixed; recheck after upgrading.

Install Plugin

openclaw plugins install npm:@acedatacloud/openclaw-provider@2026.5.34 --pin

OpenClaw 2026.9.2 will prompt to review the npm source, and you need to agree to authorize the declared capabilities. After confirming the package name, publisher, and plugin source code, the non-interactive installation for this round is:

openclaw plugins install npm:@acedatacloud/openclaw-provider@2026.5.34 \
  --pin --force --accept-capabilities

--force here means confirming the npm source, not skipping platform authentication; --accept-capabilities accepts the plugin capability declaration. If the Gateway is already running, restart the Gateway as prompted after installation. There is no need to install chat channels or system daemon services just for testing models.

Configure Key and Check

Set in the same terminal:

export ACEDATA_API_KEY="$ACEDATACLOUD_API_KEY"
openclaw plugins info acedatacloud --runtime
openclaw models list --all --provider acedatacloud

The plugin supports both ACEDATA_API_KEY and ACEDATACLOUD_API_KEY. This round of runtime shows text-inference and web-search capabilities; do not take the image/video/music text in the plugin description as registered capabilities.

For persistent configuration, you can use openclaw onboard --auth-choice acedatacloud-api-key and input the Key as prompted. --acedata-api-key is an onboarding parameter, not a global parameter that can be used with all subcommands.

Successfully Run: One Model Call

openclaw infer model run --model acedatacloud/gpt-4.1-mini \
  --prompt "Reply only ADC_OPENCLAW_OK" --json

Actual result on 2026-09-07:

{
  "ok": true,
  "capability": "model.run",
  "transport": "local",
  "provider": "acedatacloud",
  "model": "acedatacloud/gpt-4.1-mini",
  "attempts": [],
  "outputs": [{ "text": "ADC_OPENCLAW_OK", "mediaUrl": null }]
}

The running log confirms that the request was sent to https://api.fishapi.cloud/v1/chat/completions and returned HTTP 200. infer model run is a single model request and does not include the Agent tool loop; the next step is to verify:

openclaw agent --local --session-id adc-first-task \
  --model acedatacloud/gpt-4.1-mini --json \
  -m "Reply only ADC_AGENT_OK"

In this round, agent --local was also used to perform a real file task: creating a file in the isolated directory with the content ADC actual file contents 7093, requiring reading that file without writing its content into the prompt. The running result returned that content verbatim, with logs showing read called 1 time, tool failures 0 times, and no backup Provider used. This result verifies the local Agent file reading capability of this version; it does not represent that the chat channel, all tools, and the backend Gateway have been validated.

To use this model by default, run openclaw models set acedatacloud/gpt-4.1-mini. For web searches, first check openclaw infer web providers --json, configure/select acedatacloud, and then call; it requires search service permissions and cannot infer authorization based solely on successful model calls.

Practical Application and Acceptance

First, create a test directory, write hello from Ace Data Cloud in fixture.txt, and then let the client execute:

Read fixture.txt, quote its content word for word, and explain the purpose of the file. Do not modify any files.

Check whether the file was actually read and whether the correct content was quoted; then go to Platform Usage History to verify requests by time. For chat modes that do not support file tools, first attach the file, then ask the same question. Do not infer that editing, images, tools, and auto-completion are all available based on a single successful chat.

Phenomenon Troubleshooting
401 Is the API Key complete, is the Bearer prefix duplicated, is the process able to read the variable?
403 / Quota Error Application to which the Key belongs, validity period of the package, API permissions and balance
404 Is the Base URL duplicated with /v1 or has an incorrect full interface path been included; is the model ID correct?
Chat successful but Agent failed Check the current model's tool calling capabilities and client mode, first complete the above read-only test
Changes not effective Restart the client; configurations and variables for Remote/WSL/containers usually need to be placed on the actual running end

Verification Basis

Configured on 2026-09-07 against the following official materials; except for explicitly listed tests, do not mark all system and client versions as successfully run.

Return to Coding Plan Configuration Center.