How to use 2,000+ AI models in Claude Code or Codex—with price checks before every call

[denial123789]

SandBase CLI gives Claude Code, Codex, Cursor, and 22 other AI client targets a consistent MCP workflow for choosing and calling models or APIs:

discover → inspect schema and pricing → run → poll if asynchronous

This is useful when an agent needs a capability that is not already available through a dedicated tool: another LLM, image or video generation, embeddings, web extraction, social data, or a structured API.

1. Connect the client

Use the pinned v0.1.17 GitHub Release while npm trusted publishing is being enabled:

npx -y https://github.com/sandbaseai/cli/releases/download/v0.1.17/sandbaseai-cli-0.1.17.tgz connect

Approve browser authorization and select the installed client. The CLI installs a local stdio bridge; it does not start a background daemon.

2. Ask the agent to discover candidates

Example prompt:

Use sandbase_discover to find image-generation endpoints suitable for a square product illustration. Do not run anything yet.

The agent can narrow the catalog with short queries, endpoint type, vendor, and result limit. SandBase currently exposes more than 2,000 AI models and APIs through this interface.

3. Inspect schema and price before execution

Inspect the best two candidates with sandbase_inspect. Compare their required inputs and current per-call pricing, then recommend one. Do not execute yet.

sandbase_inspect returns the endpoint's input schema, pricing metadata, and request template. This avoids guessing parameter names or committing to a paid call before the user can compare options.

4. Run only after choosing

Run the selected endpoint with the approved prompt and settings. Report the exact endpoint name and whether the result is synchronous or asynchronous.

The agent calls sandbase_run with the exact endpoint name and schema-valid arguments.

For asynchronous work such as video generation, the response includes a run ID. Poll it with sandbase_run_get rather than starting duplicate jobs:

Poll that run every 10 seconds until it completes or fails. Do not create a second run.

5. Review usage and balance

Use sandbase_runs to summarize my five most recent calls and their recorded costs, then use sandbase_account to show the remaining balance.

The six MCP tools are:

• sandbase_discover — search models and APIs
• sandbase_inspect — inspect schema, pricing, and request template
• sandbase_run — execute a model or API
• sandbase_run_get — retrieve an asynchronous result
• sandbase_runs — review recent calls and costs
• sandbase_account — check account balance

Safety notes

• Prefer an existing dedicated integration when the user already has one.
• Inspect current pricing before every billable call.
• Do not submit secrets or regulated data without reviewing SandBase and upstream-provider policies.
• Keep a human in the loop for endpoints that can write to external systems.
• For provenance-sensitive installation, verify the SHA-256 published with the v0.1.17 release before executing the archive.

Resources:

• Source and installation
• v0.1.17 release
• Official MCP Registry record
• 25-client setup guide

If you try this workflow, share the client and capability you tested—without credentials or private model inputs. If it saves integration time, starring SandBase CLI helps other agent developers discover it.

Disclosure: This is an official SandBase technical walkthrough, drafted with AI assistance and checked against the v0.1.17 source, managed Skill documentation, and release metadata.

[denial123789]

Practical follow-up: you can preview the supported client targets before changing any configuration:

npx -y https://github.com/sandbaseai/cli/releases/download/v0.1.17/sandbaseai-cli-0.1.17.tgz catalog --json

Then use connect only after reviewing the target. The CLI keeps the bridge local and the repository documents checksum verification for release artifacts. See the latest release and checksum guidance.