Your team is burning tokens faster than coffee.
Coding agents are the heaviest per-seat LLM workload most companies will ever run. modelux gives engineering leadership the budget controls, policy surface, and audit trail this actually deserves — without getting in the way of the developers using the tools.
Point Claude Code at modelux once.
Claude Code and other coding agents read ANTHROPIC_BASE_URL.
Set it once per engineer; from then on every agent request
flows through modelux — tagged, logged, budgeted, and
auditable.
- ▸ Works with Claude Code, Cursor, Aider, Cline, Continue
- ▸ Per-engineer
mlx_sk_keys - ▸ Tag each request with repo, branch, and session
- ▸ MCP tools expose analytics and config to the agent itself
# Point Claude Code at modelux — uses your team's keys. export ANTHROPIC_BASE_URL="https://api.modelux.ai/v1" export ANTHROPIC_API_KEY="mlx_sk_..." # Optional — tag every request so analytics can break down by repo. export MODELUX_TAG_repo="acme-web" export MODELUX_TAG_author="alice"
Coding agents are different. They need different controls.
One dev burns through the team's Anthropic quota
A single refactoring session runs up 5M tokens. Everyone else gets rate-limited the rest of the day.
You can't tell what agents spent on what
The provider bill has one number. You have no way to say which task, which PR, which engineer.
Switching models means rewriting configs
Claude ships a better model, but updating config across ten engineers' setups is a chore nobody wants.
No audit trail for agent-driven changes
An agent rewrote your auth code. Nobody can answer "what prompts led to that commit" without digging.
Governance that respects the developers using the tools.
Per-engineer API keys, one central budget
Each developer gets a modelux key. The team shares a monthly budget cap. One engineer spiking doesn't block the others — auto-downgrade kicks in only for that key.
Tag requests with git metadata
Pass the repo, branch, and task ID in request tags. Analytics breaks spend down by repo, feature, or engineer — whatever tag key you care about.
Change default model in the dashboard
Update @default-coding to point at Claude Sonnet 4.6 once. Everyone's agents pick it up on the next request — no config push, no restart.
Full request logs with decision traces
Every agent prompt and response is captured, searchable by tags. Paired with audit logs of config changes, you can reconstruct any session.
Try a new default model without emailing the team
Before swapping @default-coding from Sonnet to Haiku, replay last week's agent traffic against the candidate. Routing-only mode is free and shows the projected spend; with-responses mode calls the candidate and scores it against Sonnet's outputs with embedding similarity. If the numbers hold, promote in one click — and every agent picks it up on the next request.
engineer requests input_tok output_tok cost ───────────── ───────── ─────────── ─────────── ─────── alice 4,821 12,480,312 1,104,227 $42.18 bob 3,204 8,711,558 891,402 $29.64 carol 2,977 7,409,921 823,110 $26.81 dan 1,248 3,102,847 294,501 $10.44 ───────────── ───────── ─────────── ─────────── ─────── total 12,250 31,704,638 3,113,240 $109.07 budget: $500/month used: 21.8% projected eom: $167
Cost by engineer, by repo, by agent session.
Tag requests with whatever dimension you care about — author, repo, branch, session, task — and analytics breaks down spend along that dimension in real time. Projected end-of-month spend updates live so you know when to tighten the belt.
- ▸ Top engineers by spend, volume, latency
- ▸ Budget alerts scoped to one engineer or tag
- ▸ Auto-downgrade near cap to avoid surprise bills
13× cheaper than GPT-5.
Token-hungry agents make ensemble routing pay for itself in a week — same accuracy, a fraction of the bill.
240ms TTFT.
Faster first tokens for the IDE feedback loop your devs feel — and a warm connection pool that beats direct calls at the tail.
Replay last week's agent traffic.
Test a cheaper default model against real agent sessions before rolling it out — measured savings with embedding-based quality scoring.
Get visibility into agent spend by Monday.
Free tier works for one developer. Team tier covers unlimited engineers with shared budgets and role-based access.