Skip to main content
Coding agents are chat completions in a tool loop, so pointing one at Pearl is the same everywhere: a base URL, an API key, and a model ID. Every model in the catalog supports tools; start with deepseek-ai/DeepSeek-V4-Pro-0813 for hard, long-context work and zai-org/GLM-5.3-Flash when latency and cost matter more. See Models.

The universal route: environment variables

Tools built on the official OpenAI SDKs read the standard variables, often that’s the entire setup:

opencode

Add a custom provider to ~/.config/opencode/opencode.json (this exact setup runs Pearl’s own engineering work), then pick a Pearl model from the model picker:
~/.config/opencode/opencode.json

Aider

Aider routes OpenAI-compatible providers through the openai/ model prefix:

Cline, Continue, and other IDE agents

IDE assistants with a provider picker all follow the same shape: choose OpenAI Compatible as the provider type, set the base URL to https://inference.pearlresearch.ai/v1, paste a Pearl API key, and enter a catalog model ID verbatim (they’re namespaced, zai-org/GLM-5.3, not glm).

Verify before blaming the tool

When an agent misbehaves, first prove the credentials and model work with one request, it separates configuration problems from tool problems in ten seconds:

What makes agents run well on Pearl

  • Agent loops are the textbook prompt-caching case, long system prompts and an append-only transcript replay on every iteration, so the cached-input discount compounds across the whole run.
  • Watch finish_reason. Agents that truncate mid-edit usually hit max_tokens during the thinking phase, raise the cap or lower reasoning.effort.
  • Give the agent its own API key (“agents-dev”), so you can revoke a runaway loop without touching production, and attribute its spend in Analytics.
Agents retry aggressively by design, make sure your retry policy respects the error semantics: back off on 429/5xx, but stop on 402, no amount of retrying refills a balance.