Ellipsis vs the Claude Code GitHub Action

Feature by feature

Capability	Ellipsis	Claude Code GitHub Action
Infrastructure to run it	Managed cloud sandboxes, nothing to host	You provision and maintain the CI runners and workflows
Startup time	Warm sandboxes in seconds, dependencies baked into a cached image	Every run boots a runner, checks out, and reinstalls dependencies
Triggers	Pull requests, pushes, cron, and @ellipsis in Slack and Linear	GitHub events inside your CI only
Scheduled agents	Native cron schedules	Scheduled workflows you build and maintain
Team-wide observability	Every session, cost, and diff in one dashboard	Scattered across workflow logs, one run at a time
Spend controls	Per session, day, week, and month caps, enforced live	You track API spend yourself
Permissions	Per-agent GitHub token scopes, enforced at the mint	Whatever the workflow and token happen to allow
Audit trail	Every action and every config version recorded	CI run logs, retained per your CI
Code review on every PR	Built in, from any agent or engineer	A separate thing you build and maintain
Agent definition	One versioned YAML per agent, deployed with git push	Workflow YAML focused on CI steps, not the agent
Secrets	Central, write-only sandbox variable store	GitHub Secrets, managed per repo
Isolation and retention	Ephemeral sandbox per session, zero source code retention	Depends on how your runner is set up

Sandbox ready time

Measured on the Ellipsis monorepo: a full poetry install, baked into the image

First run at a commitbuilds and caches the image

56.8s

Every run afterspawns from the cached image

2.9s

boot 16.4sclone 8.4sinstall dependencies 22.9ssnapshot image 9.1s

A CI runner repeats the checkout and dependency install on every run. Ellipsis pays it once per commit, then starts each sandbox from the cached image with the packages already on disk.

Speed

Sandboxes start in seconds, dependencies included

A CI workflow pays the same tax on every run: boot a runner, check out the repo, install dependencies. Ellipsis bakes the install into a cached sandbox image, so the first session at a commit pays it once and every session after starts with your packages already on disk. Measured on our own monorepo: 56.8 seconds once, then 2.9 seconds.

Dependency installs run at image build, not in every session
A new commit reinstalls only what changed
Two YAML fields: image.dockerfile_append and image.setup

Sessionslive

Flagged 1 risk on splitshift-api #1841

Dropped column without backfill

$0.21

12.0K tok

just now1m30s

Posted daily standup to splitshift-hq #engineering

4 merged, 2 awaiting review

$0.05

3.3K tok

just now22s

Opened splitshift-web #314

Implement CSV export

$0.42

31.2K tok

22s ago4m02s

Approved splitshift-api #1839

Add rate limiting to /v1

$0.18

9.1K tok

1m ago48s

Reviewed splitshift-web #1842

Fix connection leak on error path

$0.31

18.4K tok

3m ago1m12s

Observability

Every session in one place

A CI job leaves its output in scattered workflow logs, one run at a time. Ellipsis records every session in one dashboard: how it was triggered, what it changed, what it cost, and the full step-by-step trace, across every repo and every agent your team runs.

Tool callsmerge-referee

Bash

pytest -k auth · passed

Allowed

Edit

tests/test_auth.py

Allowed

gh pr merge

GH_TOKEN is read-only

Blocked

Grep

"flaky" · 3 matches

Allowed

Read

tests/test_auth.py

Allowed

Permissions

Decide what agents can do before they run

The GitHub Action can do whatever its workflow and token allow. With Ellipsis each agent's GitHub token is narrowed to the permissions and repos its YAML grants, enforced by GitHub at the mint, and spend is capped per session, per day, and per month while the agent runs.

Steps · 152Show thinking

152Commit pushed28 tok

151Bashgit commit -m 'Release connection on error path'20 tok

150thinkingWrite a clear commit message

149assistantGreen. Committing the fix.14 tok

148Bashpytest -k pool · 142 passed24 tok

147Editapi/pool.ts · +12 −426 tok

146thinkingPatch the finally block and add a test

Tracing

See every step, from thinking to diff

See the reasoning behind every change, not just the change. Each session records the agent's thinking, the tools it called, the tests it ran, and the diff it shipped, so a failed or surprising session is something you can actually inspect.

Claude Code agents your whole team can run, govern, and observe

Feature by feature

Sandboxes start in seconds, dependencies included

Every session in one place

Decide what agents can do before they run

See every step, from thinking to diff

Keep the agent, add the cloud layer