Ellipsis

Pick Ellipsis if your workflow is already Claude Code or Codex and you want those agents fired from a GitHub event or a Slack message without renting a VM yourself. The daily trait is the trigger surface: an issue comment, a Slack thread, an API call, all landing in the same sandbox. The review is one workload among many, and it feels like it.

Trusting your own agents in someone else's sandbox is a reasonable trade if the ops work was the thing slowing you down. Pick it for that. Pick CodeRabbit if you want a reviewer that is finished, rather than a platform for building your own.

The risk is the pivot. Three years as a code review bot, then a repositioning to Ellipsis Agent Cloud in July 2026. The review still exists, but the roadmap belongs to the platform, and a platform one quarter old has not yet met its first hard migration.

For a buyer this means the reviewer you evaluate today may be a maintenance-mode workload by the next renewal. Two quarters of changelog would show which way it went. What it does right: the reviewer never approves, requests changes, pushes commits or merges, so it cannot satisfy a required-review rule by accident.

The review pipeline is documented. 1. Each stage agent runs in its own sandbox with the repository checked out at the reviewed commit, so findings are judged against surrounding code rather than the diff alone. 2. A user-declared gatekeeper filters findings before they post. 3. Comments anchor to the commit, and a line already covered is never re-commented.

What is missing is measurement: no precision or recall figure accompanies the design, so the gatekeeper's effect on noise is asserted, not shown. A published false-positive rate would settle it. The observation: a filter whose output is never counted is a preference, not a control.

Model tokens pass through at cost with a 10% platform fee, which is a reseller margin, not a software margin, and it shrinks every time a lab cuts prices. The money is elsewhere: Standard, Advanced and Premier support at $5,000, $7,500 and $15,000 a month, a services business wearing a platform.

Moat: none against labs hosting their own agents, since the workloads are Claude Code and Codex, which their makers can run in their own clouds tomorrow. Likely outcome is acquisition by a cloud that wants the sandbox orchestration. Position: pass, revisit if the support line becomes the headline.

The demo is an agent answering Slack with a pull request. Procurement: no seat price, so sixty engineers is $0 in seats plus compute at $0.142 per vCPU-hour and $0.024 per GB-hour, a number nobody can forecast before the agents run, because run time depends on the task. Account-wide spend limits exist and would be mandatory on day one.

SSO, audit logs and a retention statement are absent from the pricing page, and a platform that runs agents against our repositories with write access needs all three in the contract, not the FAQ. CI fit is native, which is the one thing that makes this worth revisiting. Not yet.

Closed platform, no local models, and the agents inside are Claude Code or Codex, so the loop is someone else's twice over. What I can bend is real, though: agents are YAML files in my repo, code_review.yaml in a .ellipsis directory sets org-wide review rules, and tokens can bill through my own AWS account, so the meter is mine even if the machine is not.

It is an MCP client, and brew install ellipsis-dev/cli/agent gives me a CLI to trigger runs from a script. Configuration as code, execution as rental. Nothing to fork, but everything I wrote stays in git when I leave, which is the honest minimum.