Ellipsis is an execution platform for running coding agents such as Claude and Codex safely in the cloud. Agent configuration lives as YAML inside your repository, and agents launch in isolated sandboxes in response to GitHub events, manual commands, or API calls. You can hand off work like reviewing pull requests, creating fix commits, or implementing changes from an issue, while every execution step, the cost incurred, and the permissions granted are kept on record. The service is based in the US and holds SOC 2 Type I certification.
It is easiest to understand as filling the gap for teams who think “running agents locally makes me nervous, but I still want them in our development workflow.”
Key Features
- Agents defined as YAML in the repository: Which model to use, which repositories it may touch, and what triggers it are all managed as a configuration file. Because the configuration is under version control, you can trace afterwards which version ran and when
- Isolated sandboxes with scoped credentials: Each session gets its own container environment, and credentials are minted before the sandbox is created, narrowed to the target repositories and permission scopes. The design limits lateral movement and credential exposure
- Choice of multiple models and harnesses: Supports Claude Opus 5 and Sonnet 5 along with more than ten models including the GPT family. You can also choose the agent harness itself, such as Claude Code or Codex
- Three launch paths: events, manual, and API: In addition to automatic runs triggered by GitHub events or Linear issue creation, it supports manual runs from the CLI and programmatic invocation through an SDK. Python and TypeScript SDKs are provided
- Complete recording and replay of runs: Thinking tokens, tool calls, and diffs are all recorded and searchable. Spend per step, the exact instructions the session launched with and who issued them (a person or an API key), the configuration version it ran on, and a downloadable log all remain, and the history persists after the sandbox is deleted
- Budget caps and cost visibility: Spending limits can be set per session, per developer, or per time period. This acts as a brake when you widen the scope of what you delegate to agents
Pricing
There are no per-seat fees. Billing is usage-based, charged against the tokens and compute resources you actually consume.
| Item | Details |
|---|---|
| Starting credit | $100 |
| Tokens | At cost (no markup on managed SaaS) |
| Platform fee | 10% of token cost |
| CPU | $0.142 / vCPU-hour |
| Memory | $0.024 / GB-hour |
| Per-user cost | $0 (no seat fees) |
| Idle charges | None |
| Support packages (optional) | $5,000/month (Standard) / $7,500 (Advanced) / $15,000 (Premier) |
Besides the managed SaaS option, you can run it in your own AWS account. In that case tokens, CPU, and memory are billed directly by AWS, and Ellipsis charges only the platform fee. There is no feature difference between the two, and no feature gating by plan. The official FAQ covers open-source arrangements and spending limits, so check directly if either applies to you.
Pricing is as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- The platform takes on the agent execution environment, permissions, and budgets, so you do not have to build your own sandboxing and auditing machinery
- No seat fees, so you pay only for what you use, and the $100 starting credit lets you try it out
- Runs are recorded down to the instructions, tool calls, and cost, which makes accountability within a team easier
- Being able to choose the model and the harness reduces lock-in to a single vendor
- Running in your own AWS account is supported, which helps organizations with constraints on where data may live
⚠️ Cons
- Usage-based billing makes monthly costs hard to predict as agent runs increase, so designing budget caps is a prerequisite
- Defining agents in YAML carries a learning cost during setup
- Integration is centered on GitHub; support for other hosting services should be confirmed with the vendor
- Paid support starts at $5,000/month, which is not a level an individual developer would casually sign up for (all features remain available on a self-serve basis without support)
- Documentation and UI are primarily in English
Comparison with Similar Services
| Item | Ellipsis | Devin | OpenAI Codex (cloud) | CodeRabbit |
|---|---|---|---|---|
| Provider | Ellipsis | Cognition | OpenAI | CodeRabbit |
| Positioning | Agent execution platform | Autonomous AI software engineer | Cloud-run coding agent | Focused on automated PR review |
| Execution environment | Isolated sandboxes (SaaS / your AWS) | Cloud workspace | Cloud sandbox | SaaS |
| Model choice | Multiple models and harnesses | In-house agent | OpenAI models | Multiple models |
| Billing model | Usage-based (no seat fees) | Subscription + usage | Included in ChatGPT / API plans | Per-seat subscription |
| Main use | Building automation workflows | Delegating whole tasks | Delegating tasks | Raising review quality |
Who Is It For
- Development teams that want to run coding agents in a cloud environment with permissions and auditing in place, rather than locally
- Repository maintainers who want to automate pull request reviews and routine fixes on an event-driven basis
- Engineering leads who want agent execution costs made visible and capped
- Developers who want to operate several agents side by side without being locked into one model vendor
- Organizations with constraints on data location that need agents to run inside their own cloud
Summary
Ellipsis is a platform whose center of gravity is not “how cleverly can agents run” but “how safely and how traceably can they run.” The combination of isolated sandboxes, scoped credentials, budget caps, and complete execution logs may be overkill for an agent you use alone, but it earns its place when a team has to operate agents responsibly. With no seat fees and a starting credit available, a good first step is to build one small automation and see for yourself how the costs and the granularity of the records work out.