CCDash: product and architecture
The short verdict from our CCDash review: if you are on a Claude Pro or Max subscription and you have no idea how close you are to your next 5-hour throttle, CCDash solves that specific problem better than most things in its class — and it does it without ever asking for an API key. This is a self-hosted usage and cost observability panel for Claude, not a general developer utility, and that narrow scope is the whole reason it works. It is MIT licensed, written in Python, sitting at 71 GitHub stars, with v0.9.2 shipped on 13 April 2026. It is also young, pre-1.0, and has already had one significant feature amputation. Treat it accordingly.
Overview
CCDash reads the JSONL session transcripts that Claude Code writes to your local disk and turns them into a unified panel covering Claude Code, claude.ai, and direct API usage. The no-API-key design is the architectural decision that defines the product: because it parses files already on the machine, there is no credential to provision, no token to rotate, and no outbound call to Anthropic required to populate the dashboard. For teams with security review processes, that removes an entire approval conversation.
The project launched on 26 March 2026 as a session monitor with task scheduling attached. The current release drops scheduling entirely and repositions around usage and cost. We consider that a good call — scheduling was a different product competing for the same maintainer attention — but it is worth knowing that features have been removed from this codebase before, and at v0.9.2 the public API surface should not be treated as stable.
Adoption signals are modest and should be read as such. 71 stars and a last push of 13 April 2026 indicate a live, actively maintained single-maintainer project with an early community, not evidence of enterprise deployment. We found no published enterprise case studies, no independent benchmark data, and no detailed user review scores in the sources we examined. That absence matters if you are trying to justify this as sanctioned tooling rather than something an individual engineer runs locally.
The core value proposition is legible: Claude's subscription quotas are opaque from inside the terminal, and CCDash makes them visible before you hit them rather than after. Everything else — per-project cost attribution, model mix analysis, multi-machine aggregation — builds on that foundation.
Key Features and Architecture
The stack is deliberately plain: a Python backend with a simple JavaScript frontend, no heavyweight framework, MIT licensed. That means the deployment story is a Python process on a box you control, which is realistic for an engineer to stand up in an afternoon and equally realistic for a platform team to containerize.
Local JSONL ingestion, no API key. CCDash's collector parses Claude Code's on-disk session files directly. The practical consequence is that your prompt and session metadata never leave your infrastructure to produce the dashboard, and the tool works even for accounts that have no API credentials provisioned at all — which is exactly the situation most Pro and Max subscribers are in.
Dual-window quota tracking. The dashboard tracks both the 5-hour and the 7-day subscription quota windows as distinct gauges. These are different failure modes: the 5-hour window burns during a single intense refactoring session, while the 7-day window is what silently strands you on a Friday. Tracking both, with burn-rate projection and rate-limit prediction, is the feature we would keep if we could keep only one.
Cost attribution against published pricing. Sessions are costed using Anthropic's published rates and broken down per model and per project. This produces a counterfactual — what your usage would have cost on pay-per-use API pricing — which is the single most useful number for anyone deciding whether a subscription seat is earning its keep, or whether a given project should migrate to the API.
Model, cache, and context analytics. CCDash reports which models dominate your token spend across Opus, Sonnet, and Haiku, average response times, cache hit rates, context window utilization, and time-of-day work patterns. Cache hit rate is the underrated metric here: on Claude workloads, prompt caching is often the difference between a sustainable and an unsustainable burn rate, and very few tools surface it at all.
Multi-machine aggregation via a lightweight agent. A small agent runs on each additional machine and reports into a single panel, with live status badges showing whether each local or remote session is currently working, idle, or done. This is what makes CCDash usable for a small team rather than one laptop, and the status badges double as a cheap way to see whether a long-running agent session has actually finished.
Webhook budgets and a CLI. Token budgets fire Slack or Discord webhook alerts when spending crosses a defined threshold — push, not pull, so nobody has to keep the tab open. A companion CLI covers terminal-only quota checks without loading the web UI.
Ideal Use Cases
The solo or two-person Max-subscription shop. A single engineer running Claude Code hard against a monorepo hits the 5-hour window several times a week and rarely knows why. CCDash's burn-rate projection and per-model breakdown answer the actual question — usually "Opus is doing work Sonnet could do" — and the CLI check fits into a shell prompt. Setup is a Python process against local files, so the time-to-value here is under an hour.
A 5–20 person engineering team on individual subscriptions, consolidating visibility. This is where the aggregation agent earns its place. Each developer runs the agent, one shared instance collects, and a lead gets per-project cost attribution across the group without provisioning a single API key or routing anyone's prompts through a vendor. For agencies and consultancies that bill client work, per-project cost breakdown is close to a billing input.
Deciding between subscription and API for a specific workload. A data platform team piloting agentic ETL generation can run the workload for two weeks and read the API-equivalent cost directly off the dashboard. That is a real procurement artifact, and it is cheaper to produce with CCDash than with a spreadsheet.
Don't use this if you need audited, org-wide governance with SSO, role-based access control, retention policies, or per-user chargeback enforcement — none of that is present, and a 71-star pre-1.0 project run by one maintainer is not the foundation to build a finance control on. Equally, don't use this if your team's Claude usage is predominantly through the API with billing already visible in Anthropic's console; CCDash's differentiated value is in the subscription quota windows, and if you're not subject to them you are installing infrastructure to duplicate a bill you already receive.
Pricing and Licensing
CCDash is free and open source under the MIT license, with the full source published on GitHub. There are no seats, no usage tiers, no metered events, and no vendor relationship — MIT specifically permits commercial use, modification, and redistribution without a reciprocal source-release obligation, which means a platform team can fork it, strip the frontend, wire it into an internal portal, and ship that internally with no legal review beyond attribution.
The zero licence cost does not make total cost of ownership zero, and we would push back on anyone who models it that way. The real costs are operational. Someone must host the Python process, keep it reachable, and maintain the agents on every machine feeding the aggregated view — a per-developer install that grows linearly with headcount. Someone must also own upgrades: at v0.9.2 with a release cadence tied to a small maintainer base, breaking changes between minor versions are a reasonable expectation, not a remote risk. The March-2026 removal of task scheduling is direct evidence that features can disappear between releases, so any internal workflow you build on a CCDash behaviour is a workflow you may have to rewrite.
The offsetting economics are favourable for most teams. The tool exists to stop you from over-provisioning Claude seats or over-spending on Opus tokens, so it pays for its own hosting the first time it catches a model-selection problem. Compare that against commercial usage-analytics products, where cost typically scales with seats or ingested events — precisely the dimension that grows as your Claude adoption grows. CCDash's cost curve is flat in seats and rises only with your own ops effort. For a small team that is clearly the better shape; for a 500-engineer org that needs access control and support SLAs, the flat curve stops being the deciding factor.
Strengths & Trade-offs
Pros
- No API key required, ever. Reading local JSONL files sidesteps credential provisioning entirely and keeps session data on your own infrastructure — the difference between a tool an engineer installs today and one that waits three weeks in security review.
- Both subscription quota windows are tracked, with prediction. The 5-hour and 7-day gauges plus burn-rate projection let you see a throttle coming rather than discovering it mid-task. This is the specific gap in Anthropic's own surfaces that CCDash exists to fill.
- Cache hit rate and context window utilization are first-class metrics. Most usage trackers stop at token counts. Surfacing cache effectiveness gives you a lever you can actually pull to reduce spend, not just a number to feel bad about.
- Per-project cost attribution plus Slack and Discord webhook budgets. Costs are broken out by project and thresholds push alerts to where the team already is, which makes the tool useful without anyone keeping the dashboard open.
- MIT license and a plain Python/JavaScript stack. No framework archaeology required to read or fork the code, and no copyleft obligation on internal modifications.
Cons
- Pre-1.0 with demonstrated feature churn. v0.9.2 is the current release and task scheduling — the original launch headline from March 2026 — has already been removed. Anything you automate against this tool carries rewrite risk.
- Costs are estimates against published pricing, not your invoice. CCDash computes what usage would cost at Anthropic's listed rates; it does not reconcile against actual billing. For finance reporting that gap is disqualifying.
- Multi-machine coverage requires an agent on every machine. The aggregated view is only as complete as your agent rollout, and there is no published mechanism for enforcing installation. One developer who skips it becomes a silent hole in your cost picture.
- No governance layer. We found no evidence of SSO, role-based access control, or data retention configuration. Self-hosting is a security posture, not an access control model, and anyone who can reach the panel sees everyone's usage.
- Thin independent validation. 71 stars and a single-maintainer commit history are adoption signals, not proof of production hardening, and we found no published user review scores or third-party benchmarks to corroborate reliability claims.