Skip to main content
The Coding Agents experience is in beta.
Claude Code speaks the Anthropic Messages API, and Verlon provides an Anthropic-compatible endpoint — so Claude Code can run through a Verlon coding gate, backed by any model in the registry (GPT, Gemini, Mistral, or Claude). You get dashboard-controlled model switching, per-session observability with turn-by-turn timelines, and cost tracking for every coding session, with no changes to Claude Code itself.

​
How it works

Claude Code reads a handful of environment variables that redirect its API traffic. You point its base URL at Verlon, authenticate with a Verlon API key, and put a gate id where Claude Code expects a model name. One gate covers the whole session:
  • The gate’s coding model answers your sessions (Claude Code’s main-model traffic).
  • Claude Code’s automatic housekeeping calls (session titles, summaries — its “haiku tier”) carry a /background suffix on the gate id and run on the gate’s background model, which you’ll typically point at something cheap.
  • Subagent traffic carries a /subagents suffix so it’s attributed separately; it runs on the coding model unless you route it elsewhere in the dashboard.
Claude Code decides which role each call belongs to on its own — you only decide what each role means, and you can change that from the dashboard at any time, mid-session.

​
Setup

From the project you want to connect:
It authenticates (browser device flow, or reuses VERLON_API_KEY), creates or picks your coding gate, mints a dedicated API key, writes the settings into .claude/settings.local.json (git-ignored, so the key stays out of your repo), and enables usage telemetry. Restart Claude Code — settings are read at startup — and your next session appears on the dashboard within seconds. Verify any time with verlon doctor --gate <gate-id> (traffic arriving, sessions grouping, costs accruing), switch models mid-session with verlon switch <model-id> (takes effect next turn, no restart — verlon models lists the ids), and undo everything with verlon disconnect. Prefer staying inside Claude Code? Install the Verlon plugin:
You get /verlon:connect, /verlon:switch, /verlon:models, and /verlon:doctor — thin wrappers over the same CLI — plus the bundled Verlon MCP server, so “switch my model to gemini-3-pro” works conversationally too.

​
Manual setup (what connect does under the hood)

The dashboard also shows this block ready to copy: Coding agents → New Claude Code gate.

​
Prerequisites

  • Verlon AI account (sign up) and API key from your Dashboard
  • A coding gate (Dashboard → Coding agents → New Claude Code gate): pick a coding model and a cheap background model; the gate id is on the page header

​
Terminal

The variables live only in that shell — other terminals keep talking to Anthropic directly.

​
VS Code extension (per project)

The extension doesn’t inherit shell exports. Put the same values in the project’s .claude/settings.local.json (Claude Code keeps this file out of git, so the API key stays out of your repo):
Open the project and start a session as normal.

​
What each variable does

​
What you get

  • Sessions, not just logs — every Claude Code conversation appears under the gate’s Sessions tab, grouped into turns you can open: each model call with its role, model, duration, and cost.
  • Tool visibility — with telemetry on (the connect default), the tools Claude Code runs locally (Bash, file edits) appear in the turn timeline between the model calls, with timing and status.
  • Live model switching — verlon switch <model>, /verlon:switch, or the gate’s Settings page; the running session switches on its next request, no restart.
  • Cost clarity — spend per session and per turn, model mix, cache savings, and daily totals on the gate’s Overview.
  • Failover — optionally add a fallback model; sessions ride out a provider outage automatically. Note that failover sends session content to the fallback provider — it’s opt-in for that reason.

​
Privacy & security

  • Where your code goes: routed sessions transit Verlon to reach the model provider. Payload storage follows your account’s data controls — zero-data-retention is the default posture; see your dashboard’s privacy settings for exactly what is stored and for how long.
  • Where your key lives: connect writes it to .claude/settings.local.json, which Claude Code keeps out of git. The key is minted specifically for this machine — revoke it any time under Settings → API Keys without touching your other keys.
  • What telemetry sends: metadata only — token counts, cost, model, tool names and timings, session ids. Your prompts, code, and tool outputs are not included (Claude Code redacts them by default; we don’t ask you to change that). Remove the OTEL_* lines from the settings file — or run verlon connect claude-code --no-telemetry — to turn it off; the connector works either way.
  • Leaving: verlon disconnect removes exactly what connect wrote and leaves your other Claude Code settings untouched. Your gate history stays on the dashboard; the key stays valid until you revoke it.

​
Gotchas

  • Lots of MCP tools? Handled. Sessions with many MCP servers can exceed OpenAI’s legacy 128-tool limit — Verlon routes OpenAI-backed gates through the newer Responses API automatically, so large tool sets just work (500+ tools verified).
  • Use aliases, not full model ids, in subagent definitions. A subagent whose frontmatter pins model: haiku resolves through the env mapping and stays on the gate; a literal id like model: claude-haiku-4-5-20251001 bypasses it and fails with a clear error.
  • Model self-reports are unreliable. Don’t ask the assistant which model it is — the gate’s request log records model_used for every call.
  • Experimental client features. If you see 400s naming unfamiliar request fields, set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 to stop Claude Code sending pre-release capabilities.
To return to normal Claude Code, run verlon disconnect in the project (or close the terminal if you used shell exports) — no other state is kept.