Changelog

What’s new in the pre.dev API and SDKs.

User-facing changes to the pre.dev API, SDKs, and dashboard. Subscribe via the GitHub releases page to get notified on every SDK bump.

2026-10-12

Chat-only projects, in one Projects list

• Every folder you run predev in is a project. A folder without code starts chat only: the agent and all of its tools, with no keys, database, or file sync until you build something there. It becomes a full project, with Plan, Code, and Preview, when the agent writes code, a plan or build starts, or you type /project. See Chat-only and full projects.
• On the web, chat-only projects are in your Projects list with a Chat only tag and where they run, and the separate Sessions list is gone. A chat-only project you opened but never wrote in stays out of the list.
• A chat-only project’s page has a Chat only tag, shows Plan, Code, and Preview switched off, and has a Make it a full project button; they switch on in the open page when it becomes one.
• In the CLI, the top bar shows chat only until the project is full, the first line says what that means, and predev --help describes both.

2026-10-11

Sessions that move to the cloud can build

• When a session’s message runs in the cloud (Run in the cloud instead on the web), its cloud workspace starts with it. Once the agent writes code or starts a build there, the session becomes a project and Code and Preview appear.

2026-10-10

CLI: sessions and projects

• A folder without code opens a session. predev in a folder with no code, or in your home folder, Desktop, Downloads, or Documents, starts a session: a chat tied to that folder, with all of the agent’s tools, /chrome and ! commands included. A session writes no keys to .env, starts no local database, syncs no files, sets up no git, makes no build suggestions, and does not appear in Projects. See the pre.dev CLI.
• A folder that holds code (a git repo, an app manifest, or source files), a folder already linked to a project, and predev open <project> open a project as before.
• A session becomes a project when building starts: the agent writes code into the folder, a plan or build starts, or you run /publish or /project. The CLI says so in one line and starts what was off.
• A project made from a session, or with /project in an empty folder, gets only the keys you set for the project in its .env.
• predev open <session> opens a session in the folder it last ran in, or in ~/predev-sessions/<name> when that folder is gone.

Sessions on the web

• A Sessions page, next to Projects, lists your CLI sessions with where each one is (working, online, offline, or in the cloud), messages waiting for it, its last message, and a search.
• New session starts pre.dev on your computer, or shows the command to run. Each session can be opened on your computer, renamed, made a project, or deleted.
• A session’s page is the agent chat alone: no Plan, Code, Preview, or Publish until it becomes a project. Messages sent while your computer is offline wait for up to 7 days and run when the CLI opens again, or choose Run in the cloud instead.
• An answer to the agent’s question given on the web goes to the computer that asked it and closes the question there.

Browser Agents Local in the CLI starts faster

• @predotdev/mcp 2.2.2 adds npx -y @predotdev/mcp install-app, which installs or updates the self-updating copy with no sign-in and no change to your agents’ settings. The CLI runs it in the background, so /chrome starts without an npm lookup. See Browser Agents Local in the CLI.
• In the CLI, the agent is no longer offered screenshot tools from local MCP servers, since it receives their text only.
• !cmd keeps square brackets exactly as typed.

2026-10-09

pre.dev MCP 2.1.6 to 2.2.1

• Setup works before you sign in. npx -y @predotdev/mcp setup installs the server, adds it to your agents, and connects to Chrome first, then asks you to sign in. Press Enter to skip; chrome_act, plain-words waits, and the cloud tools work after npx -y @predotdev/mcp login. See Browser Agents Local.
• Pages are loaded when tools return. chrome_open and chrome_navigate wait for the new page, clicks and key presses give it up to 2 seconds to update, and chrome_wait with only a tab waits until the page stops changing. Open and navigate end with Page loaded, or say the page is still loading after 20 seconds.
• Leaner reads. chrome_read leaves out site menus, footers, off-screen panels, and the options of long dropdowns (full=true keeps them). selector reads only the matching elements and links adds their URLs. Link URLs lose tracking parameters, and sponsored redirects resolve to where they lead.
• A chrome_read selector that matches nothing, or a chrome_wait selector that times out, lists the page’s repeated items with selectors that work.
• Results name the Chrome profile, and a call without tab uses the tab this agent last worked in.
• The installed copy updates itself. It checks npm at most every 6 hours and verifies the download, and the running server uses the new version from its next call. Set PREDEV_MCP_AUTO_UPDATE=off to turn this off.

CLI

• Text pasted on the first screen (What do you want to build?) no longer disappears.

2026-10-08

pre.dev MCP 2.1.4 and 2.1.5

• Agents finish a browser task with the closest answer the pages show, and say what differs, instead of stopping to ask. They still ask you for sign-ins, passwords, and payments.
• chrome_snapshot lists the URL each link goes to.
• Updates no longer make Chrome ask Allow remote debugging? again, and tool fixes reach a running server without reconnecting.
• chrome_act takes target, or description, instruction, element, or query; chrome_wait takes condition or until. Both say what they need when given nothing.

/chrome in the CLI

• The agent uses your signed-in Chrome for real websites and its own headless browser for the app it is building.
• Plain-words actions work whether you signed in to the CLI in the browser or with PREDEV_API_KEY.
• /chrome is listed in predev --help.

2026-10-07

Your own Chrome in the pre.dev CLI

• /chrome in the pre.dev CLI lets the agent open tabs, read pages, click, and type in the Chrome you already use, in every profile and signed in as you. It uses your CLI login and needs Node.js 22 or later. /chrome off turns it off. See Browser Agents Local in the CLI.
• predev mcp setup adds the pre.dev MCP server to your other coding agents, such as Claude Code, Codex, and Cursor.
• npx -y @predotdev/mcp setup now adds the server to the pre.dev CLI too, and uses your CLI login when you are signed in there.
• When a local MCP server returns an image, such as a screenshot, the CLI saves it as a file and gives the agent its path.

The pre.dev MCP in every agent

• npx -y @predotdev/mcp setup now also sets itself up in Hermes, Pi, OpenClaw, GitHub Copilot CLI, Antigravity, Cline, Kiro, Qwen Code, Factory Droid, Augment, Amp, Kilo Code, Zed, Kimi Code, Junie, Warp, Rovo Dev, OpenHands, LM Studio and VS Code Insiders, on top of the agents it already covered. See manual setup for the rest.
• Agents without MCP can use every tool from a shell: npx -y @predotdev/mcp tools lists them and npx -y @predotdev/mcp call <tool> '<json>' runs one.

Switch from Claude in Chrome

• Switch from Claude in Chrome: move your coding agent to Browser Agents Local in two steps, with a tool-by-tool map from Claude in Chrome’s tools to the chrome_* tools. Your Chrome profiles, tabs, and logins carry over.

pre.dev MCP 2.1.2 and 2.1.3

• chrome_open and chrome_navigate read localhost:3000 and other host-and-port addresses as hosts, and open local and private-network addresses over http.
• Plain-words waits see the page’s fields, as chrome_act does.
• Plain-words actions mask password, card, security code, one-time code, and secret-named fields, and drop token-like values from URLs, before anything is sent to pre.dev. See Safety.
• The server’s files under ~/.predev/mcp are readable only by you, and chrome_upload and file:// opens refuse SSH keys, cloud credentials, .env files, the saved pre.dev key, and Chrome’s own data.

AI Gateway: auto-recharge before a 402

• When a /v1 call finds the workspace short of credits and auto-recharge is on, pre.dev refills first and runs the call instead of answering 402. See Errors.
• The topup_url in an insufficient_credits error, https://pre.dev/billing, now opens the plans for a free account and credits for a subscriber.
• insufficient_credits carries trial_blocked when the workspace’s free credits are on hold, for example until its email is verified. Plain-words actions then say to verify the email.

Plan mode on every plan, and one stop at 0 credits

• /plan works on free plans too, in the web workspace and the CLI. At 0 credits a plan is refused before any work starts, with Subscribe or Buy credits next to it.
• The composer’s plan chip shows only while plan mode is on, to turn it off; turn plan mode on with /plan.
• Out of credits, the chat shows one card with one line and one button: Subscribe or Buy credits while the balance is empty, then Resume once credits are back. A refused message, a paused preview, and a stopped publish name the same fix.
• On a project another account pays for (its owner, or a team), the chat checks and shows that account’s balance.

2026-10-06

The pre.dev MCP server: one command, every pre.dev tool, and your own Chrome

• One command sets it up. npx -y @predotdev/mcp setup signs you in to pre.dev in the browser, or creates a free account, and adds the pre.dev MCP server to every coding agent it finds on your Mac: Claude Code, Codex, Cursor, Windsurf, VS Code, Gemini CLI, OpenCode, and Claude Desktop. It registers the server as predev and connects to Chrome. Running it again updates to the newest version. See Connect the MCP server.
• Every pre.dev tool in one server. The local server passes through the hosted tools on your key: browser_agent, browser_agent_get, browser_agent_list, fast_spec, deep_spec, get_spec, list_specs, and get_plan. Setup replaces a pre.dev server you added before, such as the hosted one, because the local server carries all of its tools. The hosted server at https://api.pre.dev/mcp keeps working for clients without Node.js.
• Browser Agents Local: your own Chrome. 17 chrome_* tools let your coding agent drive the Google Chrome you already have open: every profile, your logins, and your open tabs. All your agents share one connection, so Chrome asks Allow remote debugging? once each time it starts. See Cloud or Local to choose.
• Plain-words actions: chrome_act clicks or types into an element you describe, and chrome_wait can wait for a plain-words condition. They use your workspace credits and appear in GET /v1/usage as browser-agents-local. Every other Chrome tool runs on your computer and never calls pre.dev.
• Free and open source under MIT: @predotdev/mcp on npm, source at predotdev/mcp. check confirms the setup, login signs in again, and uninstall removes it from every agent.

Plans list additions under their feature

• In GET /get-plan, the MCP get_plan tool, and the plan in spec status and get_spec, work added to a feature that was already built is listed right after that feature, with partOf naming it, and counts as its own item until it is built. On a feature, partOf is null.
• The Markdown summary says (added to <feature>) and counts features and additions separately.

Plan mode is back on the web

• On paid plans, the Plan chip and /plan turn on plan mode in the web workspace again, and /auto turns it off.
• Plan mode now stays on until you leave it. Each message changes the plan instead of the code and ends with a Plan updated card: the stories added, changed, or removed, the spec version, and the credits used. Build it leaves plan mode and starts building. /plan in the CLI works the same way.
• A first message that asks for a plan and no build yet, such as “Plan only, don’t build yet: a habit tracker”, plans the app without building it and opens the Plan view on a Plan ready card.
• The Plan view’s tabs (Overview, Document, Roadmap, Architecture, Spec, App map, and Changes) read the same plan, and each says when it last changed.

2026-10-05

Project plans through the API and MCP

• GET /get-plan returns a project’s verified plan by specId or projectId: what the app is for, each feature’s status from pre.dev’s checks, progress, Ready to launch with a plain fix for each open item, and where screens, API routes and stored data live in the code. Add format=markdown to get the Markdown summary alone.
• The MCP server has a get_plan tool with the same plan, for a coding agent to read before it works on a project. See get_plan.
• Spec status and get_spec carry the plan as plan once a spec is completed.
• Reading a plan uses no credits.

Progress counts features, and “works” comes from checks

• Progress now counts the features of your plan (for example 3 of 4 built), and the App map shows each feature’s status, from Planned, not built to Works, checked … ago. See How progress is counted.
• A card after each run shows what changed, whether the app runs and builds, and which features its checks confirmed.
• Build next picks the next feature from the plan, and work a sprint builds stays Built, not checked yet until a check confirms it.

2026-09-27

CLI: sign in with an API key, and run one task with predev run

• Sign in without a browser. Set PREDEV_API_KEY to your workspace API key and the CLI signs in with it when the computer has no saved login, as in CI, coding agents and fresh machines. A saved browser login still wins. See Sign in without a browser.
• predev run "<prompt>" sends one request to the agent in the current directory, prints the result, and exits: 0 finished, 1 failed or stopped, 2 not signed in or the key was rejected. With no prompt argument it reads the prompt from stdin. See Run one task without the app.
• predev logout forgets the saved login on this computer, so the CLI signs in with PREDEV_API_KEY when it is set.

Documentation for agents

Documentation update. No product change.

• Guide for agents covers the interfaces, the API key, plan access, five complete recipes, polling and error rules, and limits. It comes first in llms-full.txt.
• llms.txt is written by hand: agent entry points first, then every page by tab with a unique title.
• skill.md is reviewed by hand. Install it with npx skills add https://docs.pre.dev.
• The AI Gateway has an OpenAPI file, ai-gateway.openapi.json, covering all 25 /v1 operations with error examples. Every operation has a reference page, and every REST response in openapi.json now has an example.
• The docs search server for agents is at https://docs.pre.dev/mcp.

AI Gateway: a model list without a key, credit-only prices, and workspace-scoped files

• GET /v1/models/public lists every model an app can call, with its lab, type and modalities, and needs no API key. Credit prices stay on GET /v1/models.
• The catalog listings carry prices in credits only, in each row’s predev object. When the catalog cannot be read, they answer 502 with code catalog_unavailable.
• GET /v1/usage reports requests, credits and tokens only, and workspace is an opaque ws_… id.
• GET /v1/files lists only files this workspace uploaded, 20 per page by default and at most 50, paged with after.
• Video jobs are polled and downloaded by the id their submit response returns.
• A request that refers to another workspace’s file, response, or video job returns 404 before anything runs.

MCP tools say what they do and cost

• Tool descriptions and /mcp/info say what each tool does, when it helps, and what it costs. get_spec and list_specs use no credits.
• get_spec returns creditsUsed: the credits recorded so far for the spec’s project.

2026-09-26

Agent Skills and sign-in for MCP servers

• Skills follow the Agent Skills format. Paste a SKILL.md or upload a skill folder as a .zip. The agent sees each skill’s name and description and loads the rest only when a task needs it.
• Remote MCP servers can sign in with OAuth. Add the server’s https URL in Integrations and approve access, instead of pasting a token.
• Local command (stdio) MCP servers run only from the CLI; the web workspace uses remote servers.

A project header that fits any window

• The header no longer overflows at narrow widths. The Sandbox chip shows its status, with CPU and memory on hover, and Publish is a rocket icon.
• Clicking the Sandbox chip explains where the project runs and how to work on it from a terminal: install the CLI, then predev open.

2026-09-25

Plan mode removed on the web

• The web workspace no longer has a Plan toggle or the /plan and /auto commands. The agent decides when to plan before it builds.

CLI: publish, usage, undo, and more

• /publish and /unpublish put the app on its live address or take it down.
• /usage lists what the project’s credits went on, turn by turn.
• /undo and /redo, with a checkpoint before every run on your machine.
• /rename, /logout, /mcp add and /skills add.
• Up and down arrows recall earlier prompts, each session tab keeps its draft, and suggested next messages appear after a turn.

2026-09-23

AI Gateway: choose a model, not a provider

• Responses carry pre.dev generation ids (pdg-…), which GET /v1/generation accepts. usage no longer includes cost fields: read the charge from x-predev-credits-charged, or from GET /v1/generation for a streamed call.
• A provider object in a request is ignored, and /v1/providers and the per-model endpoints lists are gone. Choose a model id from GET /v1/models, and send models for fallbacks.
• Errors from the model keep their status and message.

Browser Agents: leaner results

• Task results and events no longer include totalTokenUsage or llmCallsUsed. Data, credits, status and step counts are unchanged.

2026-09-22

Publish apps to a live address

• Publish in the project header builds a production version of the app at its own pre.dev address, or your own domain. Every publish is kept as a version, and you choose who can open it.
• Ask the agent to publish or unpublish; it checks with you first.
• Share leads with two links: Live, the published app, and Draft, your latest changes.
• Free plans publish for 7 days; Premium and above keep apps live and can add a custom domain. See Publish.

WebMCP

• Forms and actions the agent builds are registered as WebMCP tools that run the same handler as the button, so AI agents in the browser can use your app directly.
• Browser Agents call a site’s WebMCP tools when it registers them, and work the page as before when it does not.

2026-09-17

Integrations opens on a Built-in tab

• The workspace API key, the model catalog and payments are on Integrations → Built-in. Each project’s settings has a Test or Live switch for payments.

2026-09-16

AI Gateway: hundreds of models inside every app

• https://api.pre.dev/v1 serves hundreds of models from every major lab through one OpenAI-compatible API, plus the Messages format: chat, Responses, embeddings, rerank, images, video, speech and music, transcription, and files.
• Every project built on pre.dev gets the workspace key as PREDEV_API_KEY, and calls are charged to the workspace’s credits. The catalog at GET /v1/models lists credit prices per model. See the AI Gateway.

One API key per workspace

• Each workspace, personal or team, has one fixed key that starts with pdk_. It authenticates the REST API, the MCP server, Browser Agents and the AI Gateway. Existing keys keep working.
• Rotating the key keeps the old one valid for 15 minutes.

Payments built in

• Apps built on pre.dev take payments with no setup: checkout, subscriptions and webhooks work in test mode while you build. To take real payments, connect an account you own and switch a project from test to live. See payments.

Faster code browsing, with editing when you need it

• Code opens as a viewer. Browse generated files, search names and contents, copy code, and inspect unified or split changes without loading a full editor.
• Choose Edit for manual changes. Find and replace, undo and redo, reset, and saving remain available. Switching files keeps drafts; failed saves can be retried.
• Large files stay responsive. They open in plain text with optional syntax highlighting. Mobile uses a file drawer and wrapped unified diffs.

See The workspace for the controls and shortcuts.

2026-09-14

CLI projects stay local

• A project created in the CLI and opened on the web no longer starts a cloud sandbox or reseeds a template over your local files, even when it has a GitHub remote.

2026-09-07

Reject unusable browser tasks before starting work

• The dashboard now keeps Run disabled until each task has a valid website URL and instructions or an output schema, with inline guidance and consistent URL normalization for drafts and CSV rows.
• The backend checks task input and DNS before charging or queueing, so empty tasks and nonexistent domains do not consume sandbox time. Output-schema-only extraction remains supported.
• Permanent task and URL failures no longer get another sandbox attempt. Failed runner outcomes display as failed in the run details.

2026-09-06

Browser Agents API: task mode

• POST /browser-agent accepts mode per task: extract for schema-driven extraction, agent for step-by-step interaction with a completion check, or auto (the default) to choose. Requests without mode behave as before.

2026-08-20

CLI projects on the web dashboard

• Projects started in the CLI appear in the web dashboard’s Projects list.

2026-08-02

predev --version and --help

• Both print and exit instead of opening the terminal UI, so they work in scripts and CI.

2026-07-25

MCP specification tools report failures

• fast_spec, deep_spec and get_spec report a failed generation as a failure, with the reason.
• get_spec reports a specification that is still generating as in progress instead of returning an error (since 2026-07-22).

2026-07-21

CLI commands and per-project integrations

• /clear resets the feed, /help lists every command, and /integrations shows what is connected and opens setup in the browser.
• MCP servers and skills can be scoped to one project, with a list of what each build can use and where it came from.

2026-07-15

Act from your inbox, and a much snappier app

• Answer the agent’s questions straight from email. If the agent asks a clarifying question while you’re away, the email now has one-click answer buttons: tap your choice and the build picks up immediately. Links are signed and safe from email scanners.
• Credits run out mid-build? You get the full picture. The build pauses safely and you receive an email with current progress, screenshots, and the live preview link. Top up and resume where it left off. A “we’ll email you when it’s ready” cue now appears during sprints so you can close the tab with confidence.
• A broad speed pass. The dashboard, opening a project, and chat are all significantly faster, and the hottest screens paint instantly from cache.
• Renaming a project keeps old links working, and builds now survive server restarts and idle timers instead of stalling.

2026-07-13

Watch the agent think, and dial the effort

• Live thinking ticker. While the agent reasons, its thinking streams in real time; finished reasoning collapses into a “Thoughts” block you can expand later.
• Effort levels. Choose how deep each sprint goes with /effort: low (fast direct pass), medium (task-list loop), or high (full research → code → verify pipeline), with auto routing each sprint for you by default. See How the agent works.
• Fewer unnecessary questions. The agent now asks clarifying questions only when it genuinely needs an answer, and pings you when a direct sprint finishes.
• Cleaner parallel-agent view with per-branch objective, progress, and conclusion.

2026-07-09

The terminal-style workspace, on the web

• Terminal chat everywhere. The web workspace now matches the CLI: a terminal-style chat with a / command palette (/sprint, /effort, /model, /balance, /fork, and more) and live view tabs for Plan, Code, and Preview. See The Workspace.
• The pre.dev CLI brings the same agent, slash commands, and views to your terminal.
• Before/after screenshot evidence on every browser verification step: see exactly what the agent checked and what changed.
• Builds pick themselves back up. Interrupted sprints resume automatically instead of waiting for you to notice.

2026-07-08

Generate images, video, and audio inside your build

• Images, video, audio, and music in chat. Ask the agent for a logo, a hero video, a voiceover, or a soundtrack mid-build; the generated files land directly in your project.
• More reliable builds against your acceptance criteria, with a hardened verification loop.
• Faster, cheaper generated images, and GitHub sync no longer breaks from build cache.

2026-06-25

Parallel agents you can actually inspect

• The parallel-agent panel shows real work: per-agent live streaming, full tool detail, and real file diffs as multiple agents fan out across a sprint.
• Upgraded coding agent model across all build paths, plus fixes for the agent getting stuck repeating itself.
• Faster app load and a cleaner agent step view with per-step timestamps.

2026-06-14

Usage-based billing for Browser Agents

• Pay per task, not per seat. Browser agent tasks bill against your credit balance, with a per-task minimum. Failed tasks are charged the minimum and nothing more. See billing and limits.
• Tier-aware rate limits raise throughput for paid plans.
• Typed billing errors with actionUrl deep-links (see the 2026-04-30 entry) now cover the full task lifecycle.

2026-06-09

Sessions: fork, merge, and build in parallel

• Run multiple build sessions at once. Fork a session to try something risky, work several in parallel, and merge back what works. Merges no longer drop fork edits, even when git is in a bad state.
• OAuth connectors, MCP servers, and agent skills arrive in the coding agent: connect your services once and every build inherits them. See Integrations.
• Secrets and build artifacts now stay strictly out of your GitHub repos.

2026-05-26

Real-time collaboration and mobile preview

• Multi-session collaboration: live sync, presence avatars, and per-session status across everyone working on a project. Session forks landed earlier in the month and are now hardened.
• Mobile preview with a phone-framed viewport and an Expo QR code for native projects.
• Much faster returns to a project: speculative pre-resume from the dashboard and an instant hot path when nothing changed.

2026-05-08

Browser agents join the chat

• browser_agent in the workspace: the planning agent can now drive a real browser mid-conversation, with an inline timeline card showing every step.
• Faster and tougher browser automation: an express path for simple extractions, a form-fill specialist, and stronger stealth for protected sites.
• Queue messages while architecture generation is running; they’re picked up the moment it finishes.

2026-04-30

One-click API key copy

A dashboard page to reveal and copy your API key, with sign-in and return for signed-out visitors. Your key now lives on Integrations → Built-in; see authentication.

Browser Agents: typed billing errors with actionUrl

Non-2xx responses for POST /browser-agent now carry a structured body:

{
  "error": "Need ~0.5 credits, have 0.00. Buy more to continue.",
  "code": "INSUFFICIENT_CREDITS",
  "actionUrl": "https://pre.dev/projects/browser-agents?upgrade=credits"
}

code is one of SUBSCRIPTION_REQUIRED, INSUFFICIENT_CREDITS, RATE_LIMITED, QUEUE_FULL, BATCH_TOO_LARGE. actionUrl (when present) deep-links the user back to pre.dev with the right billing modal pre-opened. window.open(actionUrl) is enough to send the user to the credit-purchase or subscription flow without any extra UI.

The same body is emitted on the SSE error event when streaming.

Node SDK: predev-api@1.1.0

import { InsufficientCreditsError, SubscriptionRequiredError } from 'predev-api';

try {
  await client.browserAgent(tasks);
} catch (e) {
  if (e instanceof InsufficientCreditsError) window.open(e.actionUrl);
  else if (e instanceof SubscriptionRequiredError) window.open(e.actionUrl);
}

New typed exceptions: SubscriptionRequiredError, InsufficientCreditsError, QueueFullError, BatchTooLargeError. All non-breaking: existing instanceof PredevAPIError checks still match. See the Node SDK error-handling docs.

Python SDK: predev-api==1.1.0

from predev_api import InsufficientCreditsError, SubscriptionRequiredError
import webbrowser

try:
    client.browser_agent(tasks)
except InsufficientCreditsError as e:
    if e.action_url: webbrowser.open(e.action_url)
except SubscriptionRequiredError as e:
    if e.action_url: webbrowser.open(e.action_url)

New exceptions mirror the Node SDK with action_url field. See the Python SDK error-handling docs.