LumeAPIShort path: Cheap LLM API · Cheap OpenAI API · OpenAI-compatible API · Multi-model API · AI API pricing · Models
Last verified: July 26, 2026
Not automatically. Claude Code Hooks react to lifecycle events and can run commands, HTTP handlers, MCP tools, prompts or experimental agents; Codex Skills package instructions, scripts and resources that Codex discovers for specialized work. Neither feature is a general model router. To use a cheaper LumeAPI model for a substep, an explicit hook command, script, MCP service or application orchestrator must call LumeAPI and pass a controlled result back to the main agent.
This distinction prevents a common architecture error: installing a skill or hook does not silently move the main Claude Code or Codex conversation onto a different provider.
Quick answer
| Question | Claude Code Hooks | Codex Skills |
|---|---|---|
| Primary purpose | React to tool/session events and allow, block or enrich behavior | Package reusable instructions, scripts and resources |
| Automatically selects a cheaper model? | No | No |
| Can invoke external code? | Yes, depending on hook type | A skill can instruct Codex to run bundled scripts/tools |
| Safe LumeAPI pattern | Explicit sidecar call with bounded input/output | Explicit script or application tool called under skill instructions |
| Changes the main agent provider? | Not by itself | Not by itself |
In short
Not automatically. Claude Code Hooks react to lifecycle events and can run commands, HTTP handlers, MCP tools, prompts or experimental agents; Codex Skills package instructions, scripts and resources that Codex discovers for specialized work.
What Claude Code Hooks actually do
Anthropic documents hook types including command, HTTP, MCP-tool, prompt and experimental agent hooks. Hook exit status or structured JSON can allow, block or influence an action. MCP tool hooks operate after Claude Code connects to the server. Prompt hooks use a Claude model for a focused decision; they are not a generic gateway switch.
A cost-aware command hook could call a small internal service:
Claude Code event
↓
command/http hook
↓
redact and bound the payload
↓
LumeAPI mini classification call
↓
structured allow/block/tag result
↓
Claude Code continues on its configured Claude routeThe main Claude Code model remains governed by Claude Code and its supported Anthropic-format gateway. See the Claude Code compatibility boundary.
What Codex Skills actually do
OpenAI’s public openai/skills catalog describes a Skill as a folder of instructions, scripts and resources that an agent can discover and use for a specialized task. The catalog had 24,184 GitHub stars when checked July 26, 2026.
A skill can standardize when to run a routing script, what fields it accepts and how to verify output. The skill itself is not an API proxy and does not change model_provider. The main Codex provider remains configured through Codex settings. For LumeAPI, that provider path remains version-sensitive because current catalog routes use Chat Completions while Codex is moving toward Responses.
Read the Codex custom-provider report before mixing skills with a gateway.
Explicit LumeAPI routing pattern
Use a small sidecar with an allowlist:
{
"task": "classify_test_failure",
"model": "gpt-5.4-mini",
"max_input_chars": 12000,
"response_schema": "failure_classification_v1"
}The sidecar—not an untrusted repository prompt—chooses the upstream model. It sends a bounded OpenAI-compatible request to https://api.lumeapi.site/v1, validates structured output and returns only the fields the main agent needs.
Use gpt-5.4-mini for classification or log compression, gpt-5.6-terra for routine delegated reasoning and gpt-5.6-sol for high-risk execution only when the main workflow can safely consume the result.
Calculated cost saving
Assume a monolithic session sends 300k input and 50k output tokens to GPT-5.6 Sol:
- Input: 300k × $1.50/M = $0.45
- Output: 50k × $9.00/M = $0.45
- Total: $0.90
Now route 120k input/15k output of classification and summarization to mini, leaving 180k/35k on Sol:
| Stage | Model | Tokens | Cost |
|---|---|---|---|
| Classify and compress | GPT-5.4 mini | 120k / 15k | $0.0473 |
| Main reasoning and edits | GPT-5.6 Sol | 180k / 35k | $0.5850 |
| Total | Hybrid | 300k / 50k | $0.6323 |
The calculated saving is $0.2677, or 29.7%. This is not a measured hook/skill benchmark. If compression removes a critical error and causes a retry, the hybrid may cost more.
Security and quality gates
- Never send arbitrary repository secrets to the sidecar.
- Allowlist tasks and model ids server-side.
- Bound input characters, output tokens and timeout.
- Use a strict response schema.
- Treat hook/skill output as untrusted until validated.
- Log caller, task, model, request id, cost and result.
- Fail closed for policy decisions.
- Keep the main agent’s provider boundary explicit.
What most guides get wrong
They describe every extension point as “routing.” A hook can call something, but calling is not the same as changing the host agent’s model. A skill can teach a workflow, but instructions are not network policy. More extensions can also increase tokens because schemas, outputs and retries enter context.
Use the broader quality/cost routing guide and Claude Code versus Codex CLI for the surrounding decision.
Production scenario
A platform team can implement one audited agent-subtask-router service. Claude Code command hooks and Codex skill scripts may call it for approved tasks such as test-log classification. Each caller gets a separate scoped key and policy. The service records cost and returns a compact schema; it never receives unrestricted repository context by default.
This centralizes the product benefit—multiple LLM routes and usage logs through one LumeAPI surface—without pretending that Claude Code and Codex share the same native protocol.
A realistic production scenario
A six-person platform team standardized on one premium model for Can Claude Code Hooks or Codex Skills Route Agent Steps Through LumeAPI? after a strong demo score. Early invoices looked fine; by week three, failing test loops had doubled input tokens without more accepted output.
They kept the orchestrator but tightened routing: a lower-cost catalog model handled exploration, a mid-tier model handled implementation, and premium routes required a logged failure condition. They truncated test logs and stopped attaching full diffs on every retry.
The breakthrough was reconciling LumeAPI usage records with task ids. Token list price mattered less than eliminating duplicate context across retries.
Expert take
LumeAPI is an independent third-party OpenAI-compatible gateway—not OpenAI, Anthropic, or Google. Published catalog rates can be materially below official reference pricing, but compatibility does not guarantee identical behavior for every native feature, streaming edge case, tool schema, Batch workflow, or prompt-cache semantics.
Follow this advice when you control prompts, retries, and routing and need auditable USD billing across GPT, Claude, and Gemini on one key. Do not cut over blindly if you depend on provider-native capabilities your stack has not tested end to end.
Run a bounded smoke test on your exact SDK, model id, streaming mode, and tool path; reconcile usage logs with orchestrator task ids; and keep a fallback route until error rates and latency are acceptable.
FAQ
Can a Claude Code hook call LumeAPI?
A command or HTTP hook can call an external service, subject to your security policy. That does not change Claude Code’s main model provider.
Can a Codex Skill choose a different model?
A skill can instruct an explicit script or tool to make a delegated call, but it is not itself a provider router.
Which substeps are safest for mini models?
Bounded classification, log compression and deterministic extraction are better candidates than final code edits or security decisions.
Does hybrid routing always save money?
No. It saves only when cheaper substeps preserve enough information to avoid premium-model rework.
Sources and evidence boundary
The architecture and savings table are an explicit design scenario, not a native Hooks/Skills feature claim.