heroku-managed-inference
Use Heroku Managed Inference and Agents with the current Heroku AI workflow. Use when the agent needs to install or inspect the Heroku AI CLI plugin, provision Heroku inference access on the current s
SKILL.mdskills/heroku-managed-inference/SKILL.md
GitHubHeroku Managed Inference
Work with Heroku Managed Inference and Agents using the current add-on and CLI plugin model.
Start with plugin and add-on inspection
Resolve this installed skill's directory first. Do not assume the current workspace contains the skill source repository.
Run the bundled inspection helper from this skill's scripts/ directory:
python3 /path/to/heroku-managed-inference/scripts/plugin_status.py
Use the output to determine:
- whether the Heroku CLI is installed
- whether the Heroku AI plugin is already installed
- whether AI subcommands are available in the current shell
Preferred workflow
Follow the current documented flow:
- install the AI plugin if needed with
heroku plugins:install @heroku/plugin-ai - create or identify the target app
- provision standard-plan access with
heroku addons:create heroku-inference:standard -a <app> - inspect available models with
heroku ai:models:list - inspect attached resources with
heroku ai:models:info -a <app> - make bounded test calls with
heroku ai:models:call <MODEL_RESOURCE> -a <app> --prompt '...'
Prefer the standard-plan add-on workflow over deprecated model-by-model creation and destroy commands. The current standard plan gives one add-on and one inference key access to the supported model catalog, so switching models should normally be a code change, not a reprovisioning step.
Resource handling
- Use aliases intentionally when attaching multiple model resources to one app.
- Keep region alignment in mind when provisioning resources.
- Treat chat, embedding, rerank, and image-capable resources as separate operational concerns even when they live on one app.
- Review the latest catalog before recommending a model. As of February 19, 2026, Heroku announced Claude 4.6 models, DeepSeek v3.2, Kimi K2.5, MiniMax M2.1, ZAI GLM 4.7, ZAI GLM 4.7 Flash, and Cohere Embed v4 under the standard-plan workflow.
- Prefer steering users away from deprecated Claude 3.5, Claude 3.7, and Claude 4 variants when Claude 4.5 or 4.6 satisfies the workload.
Guardrails
Ask for confirmation before:
- installing the Heroku AI plugin into a shared shell environment
- provisioning or destroying inference resources
- attaching a model resource from one app to another
- running test prompts that may expose sensitive data
Prefer harmless sample prompts when validating access.
References
- Read
references/inference-workflows.mdfor current commands, deprecations, and source links.