densilo

For the Claude Code GitHub Action

Cut your Claude PR review cost in half,
and get twice as many reviews finished.

Add Densilo to the Claude Code GitHub Action you already run for PR reviews: one step before it. Your workflow and your Anthropic account stay exactly as they are.

- uses: densilo/setup@v1
  with:
    densilo-key: ${{ secrets.DENSILO_KEY }}

Works with anthropics/claude-code-action, any number of seats.

How to add it

Four steps, about five minutes.

Two on densilo.com, two on GitHub. Works with anthropics/claude-code-action on Linux and macOS runners.

On densilo.com

  1. Sign in with Google at the dashboard, and create an organization from the account menu (top left). Trying it alone? Skip this and use Key for GitHub Actions on your Profile.
  2. Create a key for the repository: on the Organization page, add a service account (for example github: acme/api) and copy its key.

On GitHub

  1. Add the key as a secret: repository → Settings → Secrets and variables → Actions → New repository secret, named DENSILO_KEY.
  2. Add three lines before your Claude step, in the workflow you already have:
    steps:
      - uses: actions/checkout@v4
+     - uses: densilo/setup@v1
+       with:
+         densilo-key: ${{ secrets.DENSILO_KEY }}
      - uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          # ...your existing inputs, unchanged

That's it. Your next pull request's job summary shows what Claude used and what Densilo spared it; the organization's dashboard adds it up per repository. If Densilo can't start, the review runs exactly as before; it never fails your job.

Where the cost goes

Why Claude PR reviews cost what they do.

A review is many turns, and every turn re-sends everything so far: the diff, each file Claude opened, each command's output. Most of what a long review pays for is reading the same tool output again.

Turns, not diff size

A small diff that sends Claude exploring the repository can cost more than a large one it understands at a glance.

Failed runs still cost

A review that hits its turn limit can end as a failed job with no comment posted, after paying for every turn.

What Densilo changes

Old tool output Claude has already seen is sent in short form, with the full text kept so Claude can fetch it back exactly. Reviews finish in fewer tokens.

Measured

Real pull requests, with and without Densilo.

The same model, the standard review prompt, a 15-turn limit, the same tool permissions in both runs.

2× more reviews finish

Without Densilo, more than half the reviews ran out of turns before writing anything.

2.5× more bugs caught

Planted bugs found in the reviews, with Densilo against without.

Half the cost

per finished review. Your own numbers are in every job summary.

Questions

What teams ask.

How much does a Claude PR review cost?

Mostly it depends on how many turns the review takes. Every turn re-sends the diff, every file Claude has read and every command's output so far, so long reviews cost far more than their final context size. With Densilo, each job summary shows the tokens your review used and what Densilo spared it. In our tests, Densilo halved the cost per finished review.

Why do Claude reviews fail with max turns?

claude-code-action stops when a review reaches its turn limit, and the job can show as failed with no review posted. You still pay for every turn. Without Densilo, more than half the reviews in our tests ran out of turns before writing anything; with it, twice as many finished.

What does Densilo see?

Token counts only. Densilo runs on your GitHub runner, next to Claude Code; your code and your Anthropic key go from the runner to Anthropic as before.

What if Densilo is down?

The step warns and your review runs exactly as it would without Densilo. It never fails your job unless you ask it to.

Who sees the usage?

Everyone in your organization, on the dashboard: usage and savings for the whole team, broken down by service account.

Does it work with Codex?

Densilo for Codex runs on developers' machines today; the GitHub Action is for Claude reviews first.

See your team's own numbers.

Create an organization, add a service account, and the next pull request shows what Claude used and what Densilo spared it.