Ellipsis

Documented GitHub, Slack, Linear and Sentry integrations let agents open pull requests, post reviews and answer in the thread or ticket that started them, and each agent's YAML attaches MCP servers and secrets for further tools, so the customer extends the tool surface.

Sourceellipsis.dev/docs (integrations, agents schema)read 2026-09-29

Agents are YAML files in the repository declaring trigger, harness, model, prompt, environment, permissions and budget, live on merge and previewed on branches. Agents can start further sessions of their own, parallel sessions each run in their own sandbox, and a person can steer a session mid task, all configured by the buyer.

Sourceellipsis.dev/docs/agents and ellipsis.dev/pricingread 2026-09-29

Context is assembled per session: each session clones the repositories its environment names, so the Claude Code or Codex harness reads the working tree directly, alongside a shared skill library and connected Slack, Linear and Sentry context. No persistent index, graph or embeddings layer over the customer's code is documented.

Sourceellipsis.dev/docs (environments, sessions)read 2026-09-29

Ellipsis constrains what an agent can do rather than asking a person to approve it. Permissions declared per agent limit each session's GitHub credential to named repositories and to read only or specific write scopes, and Ellipsis API scopes, and hard budgets stop model requests per session, agent, developer and account; work reaches the codebase as pull requests the customer merges on its own code host. No step where a person approves an agent action before it commits is documented.

Sourceellipsis.dev/docs/permissions and docs/budgetsread 2026-09-29

The home page states SOC 2 Type I certified with Type II in progress. Access controls cover owner and administrator roles for account settings, scoped API keys, per session credentials minted at the permissions each agent declares and revoked at teardown, write only secrets, and customer set session data retention from zero hours to one year with deletion within four hours of expiry. No SSO, SAML or SCIM is documented, so the identity side is thin.

Sourceellipsis.dev/docs/security, docs/api/authentication and the home pageread 2026-09-29

Every session is recorded: the dashboard and CLI show messages and tool calls as they happen, session events, turns and records are readable over the API, transcripts are downloadable as unmodified Claude Code and Codex output, sessions are searchable and linked to the pull requests they produced, and cost and token usage are tracked per session, developer and agent.

Sourceellipsis.dev/docs/sessions and platform/observabilityread 2026-09-29

A session stays open for follow ups, and a mention starts a durable conversation per pull request, issue or Slack thread that later replies continue, with session records kept for a customer set retention of up to a year. That state stays within one thread, and no memory layer the agent reads across independent sessions, with its own scope and lifetime, is documented.

Sourceellipsis.dev/docs/sessions and docs/triggersread 2026-09-29

Ellipsis publishes two offerings: Managed SaaS, where sandboxes run in Ellipsis's cloud, and Your AWS Account, which deploys the control plane and agent sandboxes into the customer's AWS VPC with no credentials, code or LLM calls leaving it. Session data retention is customer set.

Sourceellipsis.dev/pricingread 2026-09-29

The pricing page lists agent templates deployable from a gallery in one command, and the docs publish a code review agent as example YAML. Which templates the gallery holds, and what each does, is not published.

Sourceellipsis.dev/pricing and docs/agents/code-reviewread 2026-09-29

Each agent can carry a cron schedule or an event trigger on GitHub pull requests, pushes and issues, Slack, Linear or Sentry events, and each run starts a session with no one invoking it; @ellipsis mentions and the dashboard, CLI and API start sessions on demand.

Sourceellipsis.dev/docs/triggersread 2026-09-29

The customer picks the harness and model for every session and agent, across makers: Claude models in Claude Code, GPT models in Codex, and GLM-5.2 and MiniMax M3, and a Claude or ChatGPT subscription can pay for sessions a user starts. The current docs reject settings that point Ellipsis at a customer's own Bedrock account or HTTPS proxy.

Sourceellipsis.dev/docs/modelsread 2026-09-29

A REST API at api.ellipsis.dev/v1 with bearer API keys and a published OpenAPI specification starts, streams, messages and stops sessions and manages agents, environments, secrets and webhooks, with Python and TypeScript SDKs and the ellipsis CLI over the same surface.

Sourceellipsis.dev/docs/api and docs/api/authenticationread 2026-09-29

Ellipsis tests the customer's code, not the agent. A code review agent reviews each push to a pull request for bugs, agents run the project's tests in their sandboxes, and structured output and session records support debugging after the fact. No harness, scored test set or optimization loop for evaluating agent behavior is documented.

Sourceellipsis.dev/docs/agents/code-review and docs/sessionsread 2026-09-29

Ellipsis does not operate browsers or desktops. Sessions run the Claude Code or Codex harness in an isolated cloud Linux sandbox with terminal and file access, which is code execution, and the current documentation names no browser, desktop or computer control tool; preview screenshot tooling is no longer described.

Sourceellipsis.dev/docs (environments, sessions)read 2026-09-29