Blink vs Emergent (2026)

Every app Blink generates comes wired to a Turso database, Blink Auth sign in, file storage and a live web address with SSL. Beyond that core, 63 connectors across email, calendars, team chat, CRM, projects, payments, analytics and social media are connected once by a workspace admin, most with a single sign in, and each project then picks which accounts it may use.

When a builder describes an app, the agent checks which connectors are live and writes the code that reads and sends data through them, and every call is routed through Blink's connector service so no token sits in the app's code. A few tools, among them Meta Ads, ElevenLabs, Perplexity and Firecrawl, take a pasted key instead. Claw agents reach Notion, Google Drive, HubSpot and Slack and send email, and outside assistants can list, authorize and disconnect connectors through Blink's MCP server.

Sourceblink.new/integrationsread 2026-10-09

The build agent wires outside services into each app in two ways. Integrations with a playbook follow a tested setup recipe for services such as Stripe, PayPal, Shopify, Razorpay, Paystack, SendGrid, Resend, Twilio, ElevenLabs and the OpenAI, Anthropic and Google models, and a few bill a small credit fee per run.

Anything else with a public API is coded in against the customer's own key. Teams can also register their own MCP servers under Manage Agents, giving the agent tools to read schemas, run queries, post to Slack or open Jira tickets while it builds. Servers are entered as JSON in the mcpServers format, and each tool list is cached for five minutes.

Sourcehelp.emergent.sh/connect-your-toolsread 2026-10-09

The build agent runs one multi step loop per request. It plans the change, writes the frontend and backend, sets up the database and sign in, lints and builds the app, and saves a version when the task is done.

An outside assistant can start the same loop through the blink_build_prompt tool, follow it with status and wait calls, and answer the agent when it stops on a question, since a new prompt is refused until that question is answered or the build is stopped.

Claw agents each run alone on their own machine and can be paused, resized or deleted without touching the others, so several can work in parallel, though they do not hand work to one another or share a plan.

Sourceblink.new/docs/cloud/tools/claude-connectorread 2026-10-09

Every conversation runs as a long lived Temporal workflow in which the agent reasons, calls a tool, reads the result and repeats, up to 10,000 tool calls per message and four hours of wall clock time. A main agent plans the work and spawns subagents for testing, design, code review and integration setup, which run in parallel workflows and report back.

Three agents set the workflow when a job starts: E-1 builds step by step, E-2 checks that the pieces it builds fit together, and E-3 runs E-1 as a subagent under its own loop for large, open ended goals. Runs end on named stop reasons such as end_turn, handoff, max_iterations and insufficient_credits, and a handoff passes the task to another agent.

Sourcehelp.emergent.sh/how-the-agent-runs-workflow-stop-reasonsread 2026-10-09

Blink Claw agents run around the clock on a private cloud machine and act on their own schedule, with examples such as a 6:45 AM summary, a 3am inbox check and a Friday report, and on events, such as a follow up sent within ten minutes of a signup.

They take messages from Slack, Telegram and Discord, with WhatsApp also listed in the automation docs, can be chatted with inside Blink, and message the user first when something needs attention. How a schedule is set up is unclear.

The app builder itself starts work when someone sends a prompt, in the Blink workspace or from Claude, Claude Code or ChatGPT over MCP.

Sourceblink.new/clawread 2026-10-09

Work reaches the builder when a person sends a message. A builder picks an app type (Full Stack App, Mobile App, Landing Page or Brainstorm), chooses an agent and types a prompt in the web workspace at app.emergent.sh or its mobile web app, and a new job can also start from an imported GitHub repository. In Plan mode the build waits until the plan is approved. On the self serve plans nothing starts a build on a schedule, a webhook or an event, so the agent works only while someone is driving it. The apps it makes can run scheduled jobs of their own once published.

Sourcehelp.emergent.sh/start-with-your-idearead 2026-10-09

Each build works from the project itself: its source tree, database schema, backend functions and saved versions. Blink's MCP server exposes the same material to outside assistants for reading, including file contents, the schema and a diff between any two versions, while secret files such as .env are never returned. Claw agents draw on the workspaces they are connected to, such as Notion and Google Drive. The Knowledge Base in Blink's docs is a retrieval feature for the apps Blink builds, so the build agent keeps no searchable index of a company's wider documents.

Sourceblink.new/docs/cloud/tools/claude-connectorread 2026-10-09

Each build draws on the job's own material: the whole conversation, the workspace code and files, the database structure, the key names in .env, and files attached to chat such as screenshots, Figma exports, PDFs, CSVs and specs. MCP servers add schemas, documents and live rows from outside systems. Pro raises the context window to one million tokens, long chats are compacted automatically, and a fork starts from an editable summary of the history. Context is gathered for each run from these sources, and the builder keeps no searchable index of a company's wider documents.

Sourcehelp.emergent.sh/a-tour-of-the-workspaceread 2026-10-09

Claw agents keep 1 GB of storage and memory that survives restarts, so preferences, goals and context carry from one conversation to the next, and pausing an agent stops its machine while keeping its data. How that memory is stored, reviewed or cleared is unclear. In the app builder, continuity comes from the project rather than from the agent: every finished task saves a version with a screenshot and a generated description, and the agent rereads the project's code and schema on each request instead of carrying memory of its own across projects.

Sourceblink.new/clawread 2026-10-09

Each job keeps its chat, code and build state across sessions, so a builder can close the workspace and come back, and the agent reads the whole conversation on every turn. Workflow state persists in Temporal, so an interrupted run resumes where it left off, and a job paused at zero credits picks up again once the balance is topped up. Less carries between jobs.

A fork condenses the history into a summary the builder can edit, early details can fade after repeated compaction, and subagents start fresh on every turn. What the agent knows about a builder or team in a new job is whatever sits in the code, the forked summary or the files attached to chat.

Sourcehelp.emergent.sh/checkpoints-undo-anythingread 2026-10-09

Plan mode lets a builder think a change through with the agent, which only analyzes and replies and does not run tools, write files or deploy, and moving from Plan to Agent asks for confirmation before implementation starts. Agent mode then makes the changes.

Production stays apart from the work: each finished task updates a private preview and saves a version, the live app changes only when someone clicks Publish, and any earlier version can be restored and republished. Outside assistants change code only through a full Blink agent turn, since the MCP server has no file write tool, and a viewer's connection is read only. Claw agents act on their own, with the secrets vault setting what each one can reach.

Sourceblink.new/docs/reference/modesread 2026-10-09

Plan mode, switched on in the composer, returns a plan card with two buttons, Request changes and Looks good, let's build it, and nothing is built until the builder approves.

Build mode, the default, starts as soon as a prompt is sent, Plan mode is unavailable in Brainstorm and E-3 sessions, and in autonomous runs a feature called Auto-HITL answers the agent's own questions with sensible defaults so the run never pauses.

Spending and release controls sit beside the plan step: a maximum credit cap per prompt that asks before going over, a budget per session, cancel from the chat menu, and a live app that changes only when someone publishes it again.

Sourcehelp.emergent.sh/talk-it-throughread 2026-10-09

SOC 2 is in progress and not yet complete. The DPA sets out TLS 1.2 or higher with HSTS in transit and AES 256 or equivalent encryption at rest, and, for the company's own production systems, role based access, single sign on, mandatory multi factor authentication, central application and access logs and audit logging on privileged operations, with breach notice within 72 hours.

Inside a project, one Owner controls visibility, Admins and deletion, Admins edit with the AI, invite people and publish, and Users can only view. Both roles can be invited on every plan, Free included, with no cap on numbers, and Business adds access levels set per project. None of the plans lists single sign on for logging in to Blink.

Sourceblink.new/dparead 2026-10-09

Projects have four roles. Owners control billing, plan and deletion, Admins manage settings, members, the database and domains, Members chat with the agent and publish, and Viewers, on Enterprise only, see the app and its history.

Business adds single sign on through SAML or OpenID Connect and role based access control, Enterprise adds user groups and workspace wide two factor enforcement, and anyone can review and revoke active sessions. Emergent holds ISO 27001:2022 and SOC 2 Type I, with GDPR alongside them in its trust center, and its DPA sets out technical measures, breach notice within 72 hours and audit reports on written request. Secrets are encrypted at rest and data in transit, and customer personal data is kept out of model training under the DPA.

Sourcehelp.emergent.sh/team-roles-permissions-collaborationread 2026-10-09

Every finished task saves a version with a screenshot, a generated description of the change, a timestamp and, for projects connected to GitHub, a commit SHA, so a builder can see what the agent changed and when a problem first appeared, then restore a working state. Outside assistants can read the version list, the source tree and a diff between versions over MCP.

For apps with a Blink Backend, the Backend tab shows every request live with its route, status code, response time and any error, and function logs can be pulled over MCP. Whether the build agent's individual tool calls are kept as a trace, or a workspace keeps an audit log of agent activity, is unclear.

Sourceblink.new/docs/launch/publishread 2026-10-09

While an app builds, the workspace shows the agent's activity live: each tool call with the arguments passed and the result returned, subagent logs nested under the main agent that expand to show their own calls, the step count for the turn and the credits being spent.

A checkpoint is saved after every agent turn, publish history keeps up to three earlier production images tagged with a timestamp, and Account Settings breaks credit use down by agent runs, publishes and Universal Key calls. Enterprise adds audit logs and credit usage reports with an analytics dashboard. How long the Enterprise audit log keeps entries is unclear.

Sourcehelp.emergent.sh/custom-agentsread 2026-10-09

Core platform data, app databases on Turso, sign in and the build sandboxes sit in the United States on a hosted service. API servers run in the US, the Netherlands and Singapore with requests handled by the nearest one, and Claw agent machines run on Fly.io in US cities, Amsterdam, Frankfurt or London, with each agent's data stored where it runs. Customers do not pick a region.

The way off the platform is the code: web apps are standard React and TypeScript on Vite and mobile apps are React Native with Expo, a code download costs 5 credits on Starter and Pro and nothing on Max, and projects sync both ways with GitHub, with the agent's own Git operations coming on Pro, so the code can be taken elsewhere.

Sourceblink.new/dparead 2026-10-09

Emergent runs as a hosted service, and customer data is processed and stored in the United States, the EU and India under its DPA. Builders choose the hosting region of a published app themselves, US or EU, in the Resources tab of Manage Publishing, with the EU priced higher. App databases run on MongoDB Atlas in the United States unless moved.

Enterprise deploys apps into the customer's own cloud through a VPC setup and supports a self hosted database, and any published app can move to the customer's own MongoDB by a manual runbook. The build platform and its subprocessors stay where they are whatever region an app uses, and EU only hosting for all data goes through a request to the privacy team.

Sourcehelp.emergent.sh/where-your-data-is-storedread 2026-10-09

Blink Claw ships four named agents, each with its own role, instructions and recommended model. Gerald works as a chief financial officer on financial models, books and taxes, and Jamie as an executive assistant on morning briefings, tasks and email drafts.

Alex researches prospects and handles outreach and follow ups as a sales development rep, and Maya triages an inbox and drafts replies.

For apps, step by step tutorials walk through building a CRM, an analytics dashboard, a project manager, an online store, a course platform and several AI tools, and any public project can be remixed as a starting point.

Sourceblink.new/clawread 2026-10-09

New jobs start from four app types, Full Stack App, Mobile App, Landing Page and Brainstorm, each on its own stack template (React with FastAPI and MongoDB, Next.js, Python only, or Expo for mobile), and from a choice of three build agents, E-1, E-2 and E-3, which are levels of one builder.

Integrations with a playbook package a tested setup recipe for each supported service, and the build page offers example prompts for websites, mobile apps, dashboards and SaaS tools. On Pro, builders make their own custom agents in a four step wizard covering the system prompt, tools and subagents. There is no gallery of finished apps or ready made agents to clone.

Sourcehelp.emergent.sh/how-apps-work-here-mental-modelread 2026-10-09

Auto, the recommended setting, routes simple interface changes to fast models, complex features to advanced ones, debugging to reasoning models and design work to models that read images.

A builder can switch Auto off and pick the build agent's model directly from Google (Gemini 3 Pro and Flash), OpenAI (GPT-5.2 and GPT-5.1), Anthropic (Claude Opus, Sonnet and Haiku 4.5), xAI (Grok Code Fast) and Z.ai (GLM 4.6), with advanced models included from the Pro plan and costing more credits per request.

Claw agents choose separately from 13 models, Claude Sonnet 4.6 by default alongside GPT-5.4, Gemini 3.1 Pro, DeepSeek, Kimi and GLM, and can switch at any time. The AI Gateway's models serve the apps Blink builds rather than the builder.

Sourceblink.new/docs/reference/modelsread 2026-10-09

Two choices shape a build. The agent dropdown sets the workflow, and a separate model selector picks the LLM that powers it, offered again when a fork opens, so a builder can move to another model when a provider slows down. Pro adds Ultra thinking and Maxx mode, a deeper reasoning toggle that costs significantly more credits.

Model providers in the data path include Anthropic, OpenAI, Google Gemini, xAI and Groq. The Universal LLM Key, which reaches 44 or more models from seven providers on Emergent credits or the customer's own provider keys, serves the apps being built rather than the builder.

Sourcehelp.emergent.sh/understanding-models-e1-e2-e3-maxxread 2026-10-09

A remote MCP server at blink.new/mcp, signed in with OAuth and bound to one workspace, carries scopes for reading the workspace and projects, writing projects and running builds.

Its 37 tools cover projects, the database, backend logs, builds, source and versions, publishing, and skills and connectors, so an assistant can start a build, answer the agent's questions, roll back to an earlier version and publish, and setup is laid out for Claude, Claude Code and ChatGPT.

Billing, workspace deletion and ownership transfer stay out of reach, and adding secrets or switching models has to happen in the browser. Separately, the AI Gateway API, reached with workspace API keys, gives code chat completions, media generation, web search and file upload.

Sourceblink.new/docs/cloud/tools/claude-connectorread 2026-10-09

Emergent also works as an MCP server for clients such as Claude Desktop or Cursor, with example tools that create an app from a prompt, publish a workspace and fetch build or runtime logs, and personal access tokens with scoped permissions reach an Emergent API. The server's address and full tool list are not given alongside those examples, so how much of the platform outside code can drive is unclear. On paid plans the Universal LLM Key also works as a LiteLLM compatible proxy for calling models from scripts and CI, restricted by IP, and code leaves through Save to GitHub.

Sourcehelp.emergent.sh/mcps-connectorsread 2026-10-09

Each finished task runs a lint check and a build before the preview updates and a version is saved, and preview errors are surfaced to the agent and, over MCP, to outside assistants. When something breaks, the agent investigates in chat before changing code, and an earlier version can be restored. In mobile apps, features such as the camera, health data, push notifications and haptics only behave correctly on real hardware, so Blink points builders to testing on a phone through Expo Go. Whether the agent writes and runs its own tests for the app is unclear.

Sourceblink.new/docs/launch/publishread 2026-10-09

Testing sits in the agent's own toolset. A testing subagent, called through a run_tests tool or when a builder asks for it, writes unit, integration and end to end tests for the app, runs them and reports coverage, and when a test fails the agent fixes the code and runs the tests again.

E-2 adds a workflow that proves the services it builds actually work together, a Codex review subagent lints and scans pull requests for security and style, and testing runs include visual regression scans.

Builders can ask for specific checks, such as the checkout flow end to end, results show in chat with what failed and why, and mobile apps get TestFlight and APK test builds on paid plans.

Sourcehelp.emergent.sh/debugging-testing-with-the-agentread 2026-10-09

Claw agents come with a web browser beside Python, Node.js and bash, and browsing is part of the jobs they take on, such as prospect research, morning briefings and inbox work. Whether they can click through pages, fill in forms or work inside sites that need a login is unclear. The app build agent works through code, its preview and the build tools rather than operating a browser or a desktop.

Sourceblink.new/clawread 2026-10-09

Agents can carry out tasks in a hosted web browser run by Browserbase, seeing and typing on pages, including ones the user is signed into, and content viewed or entered there is part of the data the platform processes. Firecrawl fetches web pages as a separate agent tool, and web search costs a small credit fee per query. Inside the app builder, the browser work Emergent describes is the preview server screenshotting the app under construction for the agent to check, so whether the builder drives the hosted browser on outside sites is unclear.

Sourceapp.emergent.sh/privacy-policyread 2026-10-09