Roomote

Tasks clone repositories and open pull requests on GitHub, GitLab, Bitbucket, Gitea or Azure DevOps, and a catalog of integrations lets agents act during a task: operate Cloudflare resources, query Snowflake, manage Resend email infrastructure, inspect Vercel, Railway and Neon deployments, read Grafana alerting, PostHog analytics and Sentry issues, and turn Linear issues into work.

Custom MCP servers cover tools outside the catalog, and HTTP integrations let an agent call an approved API while Roomote attaches the credential on the server, so the key never enters the sandbox. Integration keys extend the same broker to one service at a time.

Sourcedocs.roomote.dev/integrations/http-integrations.mdread 2026-10-18

A request opens a session that can answer from connected tools or delegate independently controllable tasks, and teammates can run unlimited tasks in parallel, make coherent changes across repositories and keep one goal going across many turns in goal mode.

Work is split across model roles: an orchestration model coordinates delegated launches, an explore model searches the codebase, an advisor model plans longer coding tasks, and helper, vision and judgment models handle routing, images and quick decisions.

A Review Code pipeline runs as its own task inside a session, and automation runs can answer directly or delegate a normal task when repository work is needed.

Sourcedocs.roomote.dev/how-roomote-works.mdread 2026-10-18

Shared memory draws on what the deployment can already see: pull requests, GitHub issues, public Slack channels and Discord threads the bot has joined, Notion pages explicitly shared with it, including database properties, Granola meeting notes and the Rippling employee directory.

When a remembered fact informs an answer, Roomote names it and cites the source in plain terms. Integrations bring Jira, Linear, Notion, Pylon, Asana and monday.com context into tasks, Exa adds web search and page fetching, and deployment wide agent guidance and skills give standing instructions such as review checklists or release processes.

Sourcedocs.roomote.dev/memory.mdread 2026-10-18

Changes come back as pull requests that go through the team's normal review, each with the transcript, logs, diff and a live preview so a reviewer can trust, continue or reject the work, and repository changes return through those review paths instead of bypassing them.

A session moves to Needs input when it is waiting on a person, integration keys need the session owner's individual approval, and Roomote leaves assignments and deadlines to people unless asked to coordinate.

There is no step to approve a plan before code is written and no per command approval inside a task, so a person's chance to stop a change comes at the pull request once the run is over.

Sourcedocs.roomote.dev/tasks.mdread 2026-10-18

Deployments use admin and member roles, with the last admin protected from removal, and people sign in through invite links, email and password, Slack workspace sign in or a Microsoft Entra tenant. Only admins can create or change environment configuration.

Each Roomote Cloud customer gets an isolated deployment with its own application services, database, cache and object storage, inference keys stay inside it and deleting a workspace destroys it with its backups, while self hosted deployments keep everything on the customer's own infrastructure.

Credentials stay server side through the integration broker, admins control anonymous telemetry, and a DPA covers Cloud, but no SOC 2 or ISO 27001 report is offered.

Sourcedocs.roomote.dev/users.mdread 2026-10-18

Each task view holds the transcript, logs, terminal output, diffs, previews, retries and artifacts, and the sessions board sorts shared work into Active, Needs input, Blocked, Ready and Done with owner, participants, delegated task count, pull request context and cost on every card.

Cost analytics reports inference spend in dollars and tokens by type, source, user, environment, provider and model, separating manual and automated tasks, session orchestration and memory synthesis. Automation runs stay in the web app as sessions with links to their tasks, pull requests, artifacts, costs and final report.

Sourcedocs.roomote.dev/tasks.mdread 2026-10-18

Every completed task leaves a memory written by the agent about what it decided, why and what is still open, along with the request that started it, and when a pull request it opened later merges or closes unmerged, that memory is updated with the outcome so recall can tell shipped work from abandoned work.

Each memory records who started the task and links to that person's page, so Roomote can answer what someone has been working on. Memory is one corpus per deployment, sessions persist as a shared workspace, and deleting a task or session removes the memories created for it and keeps them from being recreated.

Sourcedocs.roomote.dev/memory.mdread 2026-10-18

The same product runs as Roomote Cloud, one isolated single tenant deployment per customer, or self hosted from a one command Linux installer, on Railway, Render or other hosting accounts, or on a home network, with a macOS or Windows host running it inside an Ubuntu VM. Task sandboxes run as local Docker containers or on Modal, E2B, Daytona, Azure Container Apps, Blaxel or Box, and local models through Ollama or vLLM keep inference in house. Self hosted code, prompts and conversations stay on the customer's infrastructure, and the source is available under the Fair Core License.

Sourcedocs.roomote.dev/self-hosting.mdread 2026-10-18

Built in automations come ready to switch on: Review Code, Resolve PR Conflicts, Suggest Ideas, Summarize Merged PRs, Manager Stats, Inference Provider Usage Alerts, Triage Sentry Issues, Dependabot and CodeQL alert triage, a Security Auditor, a Code Quality Auditor, CI Failure Triage and a Merge announcer. A cookbook of recipes turns them into standing systems, among them CI failure auto fix, scheduled housekeeping, vendor outage triage, support channel triage, product update drafts and narrated feature demo videos, and packaged skills such as feature demo load by name in a task.

Sourcedocs.roomote.dev/automations.mdread 2026-10-18

Work starts from a mention in Slack, Microsoft Teams, Telegram or Discord, an email to the deployment's AgentMail inbox, a voice call, a pull request comment, a Linear issue, the web dashboard, the API or an MCP client.

Automations add schedules, provider events, a Run now button and private webhooks, so Roomote can review every pull request, resolve merge conflicts, triage CI failures, Sentry issues, Dependabot and CodeQL alerts, or summarize merged work without a prompt.

Results return to the thread, pull request, channel, direct message or email the work came from, and a person who started a session is notified when it needs input.

Sourcedocs.roomote.dev/automations.mdread 2026-10-18

More than 30 inference sources plug in, among them Anthropic, OpenAI, Google Gemini, DeepSeek, xAI, Moonshot, MiniMax and Z.AI direct, Amazon Bedrock and Azure, gateways such as OpenRouter, LiteLLM, Requesty and Vercel AI Gateway, open model hosts such as Together AI and Baseten, local Ollama and vLLM, and subscriptions including ChatGPT, GitHub Copilot, Kimi for Coding and SuperGrok.

Admins map models to roles, for example a strong model for planning and review and a fast one for routing, with reasoning levels per role, and people can pick a model for a session or task. Code review uses a different model from the one that wrote the change.

Sourcedocs.roomote.dev/models.mdread 2026-10-18

Roomote MCP lets any OAuth capable MCP client start a session, attach up to 20 files, search sessions and tasks, read summaries and send follow ups, launch a task in a named environment and poll for new activity with a cursor, so a local coding tool can hand work to Roomote and pick up the result. Tasks also start from the API, automation webhooks start runs over HTTP with plain text or JSON input, custom MCP servers and HTTP integrations add tools, and skills add reusable playbooks. Because the code is source available, teams can read and modify the product itself.

Sourcedocs.roomote.dev/integrations/roomote-mcp.mdread 2026-10-18

Before it calls work done, Roomote runs the real app in its sandbox, reviews its own change with a different model and attaches proof such as screenshots, recordings or a live preview URL that reviewers can click through on desktop or mobile layouts. The Review Code pipeline can run on any pull request on demand, CI Failure Triage fixes broken builds and checks they pass, and the task view keeps the logs and diffs needed to debug a run. Model choices per role are backed by published benchmarks of how candidate models handle Roomote's own work, from codebase exploration to reading screenshots.

Sourcedocs.roomote.dev/tasks.mdread 2026-10-18

Inside its task sandbox, Roomote drives a browser over the team's running product to record feature demo videos, planning story beats, clicking and scrolling with a synthetic cursor and captions at 30 or 60 frames a second, and it captures screenshots or recordings as visual proof for UI changes. This runs on an agent browser runtime in the sandbox and works on the app under development. There is no general browser agent for operating outside websites or portals, and no desktop control.

Sourcedocs.roomote.dev/cookbook/feature-demo-videos.mdread 2026-10-18