Kibble Blog

Kibble vs LiteLLM, Helicone and Portkey for coding-agent cost

By Kibble · 31 August 2026 · 7 min read

If you searched one of these names next to "Claude Code cost", here is the short answer. LiteLLM, Helicone and Portkey are gateways: they sit between the agent and the vendor, forward every request, and meter what passes through. Kibble is not in the request path at all: an open-source collector reads the usage logs the agents already write on each machine, sends counts, and merges in the vendor estimate from your own Claude Console. A gateway is a control point. Kibble is a scoreboard. Which one you need depends on whether the question is "route and cap this traffic" or "what did coding agents cost per engineer last week".

What is each product?

LiteLLM is an open-source, MIT-licensed AI gateway you self-host: one OpenAI-compatible endpoint in front of a hundred-plus providers, with usage and spend tracked per key, user, team and organization, and hard budgets that block requests when a limit is reached (litellm.ai, read 31 August 2026). The open-source gateway is free; Enterprise is custom pricing. Claude Code reaches it through environment variables, and LiteLLM's own tutorial shows per-developer cost tracking with virtual keys.

Helicone is an open-source LLM observability platform with gateway features: request monitoring, caching, rate limits, fallbacks. Pricing on 31 August 2026: free up to 10,000 requests a month, Pro at $79 a month plus usage, Team at $799 a month, on-prem on Enterprise (helicone.ai/pricing). Claude Code connects by one variable, ANTHROPIC_BASE_URL=https://anthropic.helicone.ai/<key>; the integration page notes that path is maintained but no longer actively developed, and points at their newer AI Gateway.

Portkey is an AI gateway with observability, routing, guardrails and prompt management. The open-source gateway is free to self-host; the hosted Developer plan is free with 10,000 recorded logs a month, Production is $49 a month for 100,000 logs, then $9 per additional 100,000 (portkey.ai/pricing, read 31 August 2026). Claude Code is a first-class client: npx portkey sets ANTHROPIC_BASE_URL to https://api.portkey.ai and routes to Anthropic, Bedrock or Vertex (their docs).

Kibble is a team dashboard fed by a collector on each machine. The collector reads the per-session logs Claude Code and Codex already write, and sends per-day counts: token totals, model names, opaque session ids, repository names. Never prompts, file contents or paths, and the collector is public, so the claim can be read line by line. Vendor-reported estimates comes from the organization's own Claude Console through an Admin API key and is merged per day beside the estimates. Free for one person; $100 a month covers up to 30 people, then $5 per person.

How do they compare for coding-agent cost tracking?

Every fact in this table was read from the linked public pages on 31 August 2026. Prices and limits move; the links are the source of truth.

When is a gateway the right choice?

When you need control, not just visibility. Three cases are squarely gateway problems. Hard caps: LiteLLM blocks a request the moment a key's budget is spent; Kibble will alert, and will never stand between an engineer and the model. Routing: sending the same agent to Anthropic, Bedrock or Vertex by config, failing over between them, or splitting traffic by cost is exactly what Portkey and LiteLLM are for. Compliance retention: if your regime requires every prompt kept, only something in the request path can keep them, and the engineers should know it does. The price of all three is the same: every agent and CI job must be reconfigured to point at the proxy, the proxy sees content, and its uptime becomes the team's uptime.

One quiet limit to check before choosing a gateway for cost visibility: it meters the traffic that is routed through it, which in practice means traffic on API keys you control. A seat on a subscription plan is billed flat, and a laptop where the environment variable was never exported talks straight to the vendor; both simply do not appear. The transcripts on those machines still record every token, which is the half Kibble reads.

When is Kibble the right choice?

When the question is team usage and estimated cost. Kibble shows spend per engineer, team, model and repo, this week against last, with the biggest mover named, and puts the vendor estimate from your own Console alongside local estimates so the two are compared and never summed (why that rule matters). Nothing changes in how any agent talks to its vendor, there is no proxy to keep up, and rollout is one command per machine rather than one config per agent per CI job. Engineers can read what leaves their laptop, and each of them sees their own machines while owners see totals: visibility, never surveillance.

Can you run Kibble beside a gateway?

Yes, and teams that need both should. They do not touch: the gateway holds the keys, routes and caps in the request path; the collector reads the logs the agent writes locally either way and fills the dashboard. If the gateway is doing its job you get routing and hard budgets; if a machine or a runner slips out of the gateway's config, the vendor usage merge on the dashboard is where the gap shows up as an unattributed remainder instead of silently missing.

Questions

Is Kibble an LLM gateway like LiteLLM or Portkey?

No. A gateway sits between the agent and the vendor and forwards every request. Kibble is never in the request path: an open-source collector reads the usage logs Claude Code and Codex already write on each machine and sends per-day counts to one dashboard. Nothing changes in how an agent talks to its vendor.

Do LiteLLM, Helicone and Portkey see your prompts?

Every request passes through them, so they can log prompts, files and diffs. Each one offers settings to limit what is recorded, which means the guarantee is the operator's configuration. Kibble's guarantee is architectural: the collector sends counts only, the ingest schema rejects a request carrying an unexpected field, and the code is public.

Which is cheapest for a team of ten engineers?

As of 31 August 2026: LiteLLM is free to self-host, so its price is your ops time. Portkey's Production plan is $49 a month for 100,000 recorded logs, then $9 per additional 100,000. Helicone's Pro plan is $79 a month plus usage-based pricing beyond the free tier. Kibble is $100 a month flat, covering up to 30 people (price updated 1 September 2026). A busy Claude Code seat makes thousands of requests a day, so per-request log limits are the number to check first.

Does any of them show what Anthropic actually billed?

A gateway shows its own estimate for traffic it handles. Kibble combines local estimates with your organization's Claude Console analytics, which also reports estimated cost. Vendor estimates replace matching local daily usage without double-counting. Neither is a substitute for the vendor invoice.