Pollinations CLI v1.4.0

Pollinations CLI v1.4.0

What was added?

A Massive Features and Agents Expansion. This major release introduces a comprehensive suite of new CLI commands (ranging from audio transcription and local QR generation to full settings and key management), deploys three specialized sub-agents into the autonomous workspace, upgrades session saving with AI-generated context summaries, and establishes a robust API resilience layer with intelligent local quota management.
Nine new utility commands out of the box. The CLI expands far beyond standard content generation with powerful new workflow tools:

• pollinations transcribe — Fast speech-to-text supporting 8 audio formats, backed by automatic model validation.
• pollinations upload — Uploads local media with a 14-day TTL, adding a --copy flag to instantly grab the URL for cross-command piping.
• pollinations search — Performs live web searches powered by grounded models.
• pollinations qr — Generates offline QR codes (PNG/SVG/ASCII) directly in your terminal with zero API calls.
• pollinations remove-bg — Instantly isolates subjects and outputs transparent PNGs.
• pollinations diagram — Generates 10 types of AI-designed Mermaid diagrams saved directly to disk.
• pollinations settings / keys / usage — Complete interactive suites for locally auditing your API keys, configuring global defaults, tracking granular per-request costs, and visualizing daily spending via terminal bar charts.
• pollinations quota [limit] — Sets a strict, local hourly call cap to protect your budget independently of your network Pollen balance.
  The autonomous workspace gets three new sub-agents. The pollinations assist loop is now backed by a collaborative trio that automates project context, file insights, and execution safety:
• The Indexer Agent runs silently on startup to construct a structured summary of your project codebase. Using file-change watching, it performs diff-aware partial updates so altered files are patched instantly without triggering a heavy, race-condition-prone re-index.
• The Analyser Agent intercepts file paths mentioned in chat, processing them in parallel. It automatically downscales heavy vision tasks to 512px, localized fallbacks protect instance states on validation errors, and it emits clean, single-line terminal summaries to stop concurrent output interleaving.
• The Executor Agent safely materializes code. Triggered after user confirmation, it automatically detects 16 project types to install dependencies, run lints, and stand up local servers (printing clickable terminal hyperlinks for web previews or capping environment build failures to a strict 3-retry ceiling).
  Complex tasks trigger the Auto-Architect. The system now monitors user messages; requests hitting over 25 words or utilizing complex multi-step keywords automatically route to the Architect role before landing on the Coder. An internal tracking flag ensures this orchestration layer never double-triggers if the Coder calls for a consultation mid-run.
  Context compression gets ruthless. When long sessions breach maximum history boundaries, the history engine strips heavy Base64 image payloads and discards granular reasoning paths entirely. It compresses past states into an 8-bullet snapshot summarizing explicit actions, file edits, and resolved errors—though it safely pauses compression if the last event was a build failure the Coder still needs to analyze.
  Sessions get smart summaries and cross-platform resume. On exit, your active window is reviewed by a fast model that generates an immutable, contextual title (e.g., "Discord bot rate limiter") alongside an 8-bullet recap of your current state. Restoring progress via pollinations continue utilizes os.homedir() for rock-solid cross-platform stability (fixing broken Windows paths) and mirrors that context block back to you so you know exactly where you left off.
  The API Resilience Layer safeguards generation. Built to protect active routines from 402 payment errors or structural timeouts, a new middleware layer coordinates 3-attempt exponential backoffs. It checks a built-in model fallback map reconstructed directly from official platform lists, routes unique audio-stt tasks securely, and uses a circuit-breaking flag to guarantee an unexpected error never triggers an infinite fallback loop.
  The workspace system prompt is fully restored. The specialized buildSystemPrompt layout has been brought back to full strength. The agent once again receives the comprehensive Conversational Rulebook (complete with explicit anti-patterns), wrong-format tool examples, the complete step-by-step Quality Protocol, and specific failure-handling routines for tool crashes and maximum iteration ceilings.

Changed

• pollinations audio — Dropped the --instrumental flag. The engine now determines TTS vs. music arrangements purely based on your model selection, falling back to your persistent local settings.
• pollinations image & video — Corrected default imaging targets to zimage and fixed the video rendering routing to point to /video/{prompt}. Both commands gain native --image mapping for image-to-image/animation loops alongside automated post-generation --upload hooks.
• pollinations profile — Rebuilt around the live three-bucket /account/balance payload. It scales the terminal visual tier bar against real hourly grant limits, handles deduction order visualization, and gracefully flags negative reset-window times as "resetting now".
• Template Interpolation — The template system has been completely rewritten to swap old {{variable}} configurations out for clean single-brace {variable} substitution. It guards against path traversal tricks, warns you if custom template variables collide with reserved CLI flags, and extracts live properties if a saved template's variable array is empty.

Fixed

• Fixed an explicit issue where text.js vision payloads were accidentally stringified instead of structured and delivered as compliant API message arrays.
• Resolved a settings bug where global type coercion accidentally forced Number() formats onto native string parameters like safety modes.
• Fixed a file-state issue where runChanges ran structural path-existence checks after writing, which caused edited items to always report as newly created.
• Fixed an architecture bug where the Analyser ran prior to initial greeting checks, accidentally leaving orphaned system messages scattered through historical logs.
• Fixed structural shell_exec routines to ensure non-zero exit codes are explicitly appended to terminal feeds even when standard stdout data is present.
• Corrected a regex parser bug within the Ghost Runtime syntax-checker (quickLint) where unescaped special symbols like $ or . caused validation crashes on valid variables.
• Patched an environment loader bug where .env files using explicit export VAR=val syntax had their prefixes ignored; values are now cleanly stripped and loaded.
• Fixed a local key management bug where keys.js attempted to parse an unavailable expiresIn property, swapping it out to compute true expiration deltas using expiresAt.
• Resolved an automation problem where the background upload spinner initialized before file validations completed, ensuring missing files or invalid extensions fail fast before drawing UI elements.

Pollinations CLI v1.3.1

Pollinations CLI v1.3.1

What was fixed?

The Auth Reliability Patch. A focused hotfix that resolves the BYOP login flow failing silently across all platforms, the agent not connecting to Pollinations MCP without an AGENTS.md, and a permission denied crash on first run after install — most notably on Termux.

---

BYOP login now actually works. The browser-to-CLI key handshake was broken on every platform due to fetch using no-cors mode, which silently swallows connection errors — the browser always reported success even when the CLI never received the key. The server also had an IPv4/IPv6 bind mismatch and no error handling for port conflicts. All of this is fixed: the handshake is reliable, port errors print a useful message, and keys with special characters no longer arrive corrupted.

No more infinite SSO loop. After signing in through the Pollinations gateway, your key was never saved to localStorage. So the next time you ran pollinations login, the bridge page found no session and bounced you through SSO again. Fixed — the key is now persisted immediately on sign-in.

pollinations assist connects to Pollinations MCP by default. If you had no AGENTS.md, the agent launched with zero MCP tools because the default config had an empty mcp_servers list. The Pollinations MCP server is now included out of the box. AGENTS.md YAML also no longer silently fails to parse on Windows due to CRLF line endings.

permission denied after install is fixed — including on Termux. npm install -g didn't reliably mark the binary as executable on all platforms. A prepare script now runs automatically on install and sets the correct permissions, so the CLI works immediately without any manual intervention.

Pollinations CLI v1.3.0

Pollinations CLI v1.3.0

What was added?

The Swarm Maturity Update. Version 1.3.0 is less about introducing Pollina and more about making her actually good at her job. The swarm gets a memory, a conscience, a researcher, and the discipline to not spiral into chaos every time you say "sup." This is the update that takes the autonomous agent from impressive demo to reliable development partner.

---

Sessions — the hive remembers. You can now save any chat or assist session and pick it up exactly where you left off. Type exit, answer the prompt, and the session is stored under a numbered ID in ~/.pollinations/sessions.json. pollinations session lists everything you've saved. pollinations continue 3 drops you straight back in. For chat sessions, the last 10 messages are printed on screen the moment you resume so the conversation is immediately readable. For assist sessions, Pollina gets a recap of the previous work and — critically — is summoned back into the exact directory she was working in when you saved, regardless of where you're running the command from. If that directory was deleted, she falls back gracefully.

Surgical file editing. The new edit_file tool lets Pollina make targeted changes without rewriting entire files. Five operations: insert_after, insert_before, delete_lines, replace_lines, and replace_text. No more token-heavy full rewrites to change two lines of logic. Cleaner diffs, lower cost, less Critic noise.

Ghost runtime. Before any .js or .json file touches disk, the content is validated via node --check in a temp file. A SYNTAX_ERROR blocks the write entirely and feeds the exact error back to the Coder to fix — no bad code ever reaches your project. JSON is validated through JSON.parse. The Critic never even sees it until the syntax is clean.

The Researcher role. Pollina now has access to a grounded search model embedded directly in the orchestrator loop — not a separate file, not a separate process. When the Coder is uncertain about a library version, an API signature, or any factual technical detail, it calls consult_researcher and gets a current, cited answer. The findings are then injected into the Critic's next validation context so the Critic verifies against up-to-date information rather than whatever was in its training data. The researcher model defaults to gemini-search and is fully configurable per project via AGENTS.md under researcher.model. It uses your Pollinations API key automatically — no separate credential needed.

Context compression for long sessions. Once a session exceeds 26 messages, the older history is summarised into a compact factual state snapshot by the Architect model — files written, commands run, errors resolved, current project state — keeping the 8 most recent messages verbatim. The agent stays sharp across hours-long sessions instead of degrading into context bloat.

MCP credential injection. Your Pollinations API key is now automatically forwarded to every MCP subprocess as POLLINATIONS_API_KEY before the connection is made. For the official @pollinations_ai/mcp server this means zero additional configuration — run pollinations login once and every MCP tool has access. For other servers, the env: block in AGENTS.md supports ${VAR_NAME} placeholders resolved from your host environment at runtime, so credentials like GITHUB_TOKEN or POSTGRES_CONNECTION_STRING never get hardcoded into config files.

capture_asset tool. MCP image generation returns transient URLs that expire. capture_asset downloads any remote URL to a local path immediately, making generated assets persistent. Pollina is instructed to call it automatically after any MCP image generation.

---

The Critic grows up. The Critic role has been significantly tightened. It now skips entirely for read-only operations — list_files, read_file, test_syntax, move_file, delete_file, capture_asset. There is nothing to validate on a directory listing and the old behaviour of demanding source code from a file list was embarrassing. For operations the Critic does review, it receives the last 6 tool results as project context so it can validate imports and file references without demanding to audit the entire codebase. The prompt is now explicit: do not ask for more files, do not reject because output looks brief, a list of filenames is a complete and correct result. Critic output is also silent on the terminal when it passes — you only see REJECTED and the specific reason if something is actually wrong.

Greeting detection. A regex pre-flight check now intercepts casual input — hi, sup, ok, thanks, what can you do, etc. — before it ever reaches the full tool-enabled agent loop. Matched inputs get a single lightweight API call with a "reply naturally" system prompt and nothing else. Pollina no longer responds to "hey" by listing your entire project directory.

Parser hardened. The JSON action parser was rebuilt to scan all candidates in the model's response in order, skipping non-tool objects, rather than only trying the first {" it finds. This fixes a silent failure mode where a model writing reasoning JSON before the tool call would cause the actual tool call to be dropped entirely. The parser also handles { "tool": with a space after the brace, which several models emit. Duplicate candidates from nested objects are eliminated by advancing the scan position past each matched span.

Concurrency guards. A busy flag in both chat and assist prevents a second orchestrator.run() from firing if the user types while a task is still in progress. An exiting flag blocks new input while the save dialog is open. Previously, fast typing could spawn parallel agent runs or fire new tasks mid-exit.

Session save reliability. The sessions file load is now wrapped in a try/catch — a corrupted or partially-written sessions.json no longer crashes every command that touches sessions. The save calls in chat and assist are also wrapped so a disk-full or permissions error prints a clean message and exits gracefully rather than throwing an unhandled rejection at the worst possible moment.

AGENTS.md updated. The generated template now includes the researcher block, correct MCP entries pointing to @pollinations_ai/mcp with the env: credential syntax documented inline, and revised constraints. Every field is commented with what it does.

Pollinations CLI v1.2.4

Pollinations CLI v1.2.4

What was added?

Added the official GitHub repository link to package.json to enable the repository sidebar and links on the npm registry.

Pollinations CLI v1.2.3

Pollinations CLI v1.2.3

What was updated?

The CLI now automatically pulls your Pollen key from the web browser for instant login. If you are not logged in, it will guide you through the standard sign-in process. Updated the navbar links to ensure registration works smoothly and clarified the login instructions in the terminal.

Pollinations CLI v1.2.2

Pollinations CLI v1.2.2

What was added and changed?

This update introduces BYOP (Bring Your Own Pollen), a new authentication flow that lets users log in via enter.pollinations.ai to use their own credits at no cost to developers. The login command now features a Hybrid Login UI, giving you the choice between secure Web Auth and manual API key entry, supported by automatic browser integration for a smoother setup. Additionally, our autonomous agent, Pollina, gains File System Tooling for advanced file renaming and reorganization, while a modular refactor and UI refresh ensure a cleaner, more maintainable experience.

Pollinations CLI v1.2.1

Pollinations CLI v1.2.1

What was fixed?

Stabilized the swarm orchestrator by resolving critical pathing and logic loops. This update fixes the asset placement bug by enforcing absolute path resolution for generated images, ensuring they land in the correct project subfolders. The internal feedback loop is significantly improved by granting the Critic direct visibility into tool arguments, allowing for real-time code validation. Additionally, refined system protocols prevent unnecessary tool triggers during casual conversation, while a streamlined terminal interface hides internal agent conversatio to focus on direct task execution and tool status.

Pollinations CLI v1.2.0

Pollinations CLI v1.2.0

What was added?

The Swarm Intelligence Update. Introducing Pollina, a bee-ast of an autonomous agent that transforms the CLI from a simple bridge into a full-scale development partner. Version 1.2.0 brings AGENTS.md integration, allowing you to define project-specific constraints and context that the AI actually follows. With full Model Context Protocol (MCP) support, Pollina can now bridge local file operations with external tools. Whether you're in a high-end dev environment or grinding in Termux, Pollina now adapts, plans with an Architect persona, and executes with a Coder's precision.

Pollinations CLI v1.1.1

Pollinations CLI v1.1.1

What was fixed?

Squashed a tiny endpoint bug with the chat command and fixed responses parsing. Published on NPM.

Pollinations CLI v1.1.0

Pollinations CLI v1.1.0

What's New?

The Interaction & Gallery Update. Introducing a persistent Chat mode with session memory/context and a professional Gallery to track your batch generations. Featuring enhanced TUI tables, improved file versioning to prevent overwrites, and deeper data persistence. Experience a smarter, more organized way to interact with Pollinations.ai. Published on NPM.