A dependency-free LLM gateway and self-hosted dashboard. Point any project at it - any language, one command - and get cost, traces, prompt versions, budgets and anomaly alerts across your whole AI portfolio. No SDK to adopt, no database to run.
Every provider SDK already supports a custom base URL, so onboarding a project is one line - or zero, via an environment variable. Nothing else about your code changes, and your API keys are forwarded to the provider untouched.
from openai import OpenAI # the only change: point base_url at the gateway client = OpenAI(base_url="http://localhost:4321/p/invoice-bot/openai/v1") # your OPENAI_API_KEY is passed straight through, untouched
Your app calls the LLM as before, but through the gateway. It forwards the request untouched, streams the response straight back, and reads token usage on the side to compute cost. Added latency is well under a millisecond - a thin, auditable proxy, nothing more.
Also: install a thin helper (pip install aicc-sdk or npm i @ai-command-center/sdk) that sets the base URL for you; group calls without touching the path via an x-aicc-project header; report batch or unsupported-provider usage to /api/track; and swap /p/<project> for a /k/<gateway-key> when auth is on. Full integration guide →
Every capability runs from the same one-command install, with metadata-only storage and no external services.
Per-request cost from real token counts (including cached tokens), grouped by project and model, with latency percentiles and error rates. Shown in ₹ / $ / € with live exchange rates and stored in USD.
Group the calls in one request or agent run with a trace header, then read the timeline.
Tag prompts by version and compare cost, latency and error rate as they change.
Per-project monthly budgets, error-rate and latency thresholds, and an optional webhook.
Explainable rules flag cost spikes and error bursts per project - no ML, no black box.
Effective cost per million tokens, p50/p95 latency and error rate, compared side by side.
Opt-in routes fail over and load-balance across providers - no client change, no new dependency.
Provider keys pass through and are never logged. Prompt and response bodies are never stored.
Platforms like Langfuse, Helicone and LangSmith go deeper on tracing and evaluation; LiteLLM and Portkey are richer gateways. They are also a database, a queue and an analytics cluster to operate. This is one command and a file of JSONL - the answer to “what is each project spending, and is anything off?” without standing up infrastructure.
Clear about what exists today and what is deliberately not built yet.
No signup, no database, no vendor. One command and a base URL.