opens in a new tab

Last updated on October 26, 2026

AI tools for developers

We organize AI tools by the job they do and the limits in their docs, with our five-tool coding and model stack and six documented cross-tool workflows.

October 2026 releases

Hot this month

Model and tool updates that open up a new workflow or local capability, with the caveats to check before you switch.

  • Gemini 4 Argon rolls out in Google Antigravity

    What changed
    Google rolled out Gemini 4 Argon in the Antigravity desktop app and agy CLI (with current model docs also listing IDE extensions) for Google AI Pro, Ultra, and eligible Gemini Enterprise users.
    Why it matters
    Developers using Antigravity for architecture, debugging, or multi-file refactors can switch to Argon inside their existing terminal, desktop, or editor session.
    Limits to check
    Unavailable in the standalone Antigravity IDE. Requires paid Google AI Pro, Ultra, or eligible Enterprise access (or an eligible device Pro trial) and plan limits.
    Antigravity Gemini 4 Argon releaseopens in a new tabDetails

  • Comfy Agent launches in Comfy Cloud for node-graph workflows

    What changed
    Comfy Agent plans, builds, runs, and edits visible ComfyUI node graphs directly on the canvas alongside you, with support for reusable public or private skills and parallel chats.
    Why it matters
    Instead of handing back a single opaque asset, the agent leaves an inspectable node pipeline—models, seeds, samplers, and masks—that teams can tweak and rerun for product variants.
    Limits to check
    Beta on Comfy Cloud with cloud-supported nodes only; local ComfyUI agent and custom-node support are still coming soon, and Comfy MCP is a separate external route.
    Comfy Agent launch announcementopens in a new tabDetails

  • Higgsfield releases Katana motion skill for Claude MCP

    What changed
    The Katana skill lets Claude trigger preset-driven video edits through Higgsfield MCP using up to five character, product, or clothing reference images and optional prompt adjustments.
    Why it matters
    Developers and designers already working in Claude can turn product or character reference photos into short motion clips without switching to a separate timeline editor.
    Limits to check
    Normal edits accept up to 5 reference images (40 is for community presets). Katana's entry documents preset video edits only; verify any alpha or format needs downstream.
    Higgsfield Creator Hub changelogopens in a new tabDetails

  • Langfuse expands experiment charts and baseline regression review

    What changed
    The Experiments view adds side-by-side metric charts, baseline summary statistics, per-item score threshold filtering, and keyboard-driven inline human scores, output corrections, and comments.
    Why it matters
    Teams evaluating a prompt, retrieval, or model change can spot where average gains hide a failing critical case and inspect the linked trace before shipping.
    Limits to check
    Comparisons require matching dataset and evaluator versions, missing scores are not passes, and picking a baseline in the UI does not configure an automated CI release gate.
    Langfuse experiment comparisons changelogopens in a new tabDetails

  • Superwhisper launches bot-free Meetings with local Mac models

    What changed
    Superwhisper 3.0+ on macOS records mic and system audio without a call bot, generates speaker-labeled transcripts and custom summaries, and can run offline with local transcription and LLM models.
    Why it matters
    With local transcription and LLM models selected, Mac teams can turn calls into local Markdown briefs without uploading audio or transcripts, or query saved notes from coding agents via MCP.
    Limits to check
    Requires Mac and Superwhisper 3.0+. Pages conflict on ongoing Free access vs. trial words (exhausted words stop recording); connected agents send what they read to their provider.
    Superwhisper Meetings launch postopens in a new tabDetails

  • Ollama adds local text-to-speech with Phonon on Apple Silicon

    What changed
    Ollama 0.40.11+ can stream 24 kHz WAV or PCM speech from text via the CLI or the OpenAI-compatible /v1/audio/speech endpoint using Gradium's MLX-accelerated Phonon models.
    Why it matters
    Developers already running Ollama on a Mac can pipe local model replies aloud or generate voiceover WAV files through standard OpenAI SDK audio calls.
    Limits to check
    TTS output only on Apple Silicon macOS (Ollama 0.40.11+; Windows/Linux planned). Base English Phonon is Apache 2.0, while 40M language variants are noncommercial CC BY-NC 4.0.
    Speech models in Ollama blog postopens in a new tabDetails

CodeSubmit team stack

The five tools our team uses

Codex, Claude Code, and Antigravity handle our repository work, while Hugging Face Hub and Ollama host and run our own models.

  1. Codex

    Our team uses Codex, especially with GPT-6 Astra. Start here when a repository change spans multiple files, tests, and architecture decisions, and choose Astra when complex work needs sustained reasoning and judgment.

    Configuration
    Select gpt-6-astra for demanding multi-step work; official docs recommend comparing gpt-6.1-sol on recurring tasks where usage cost matters.
    Tradeoffs & limits
    Model options and reasoning controls depend on your plan, client, and workspace settings, and API-key sessions bill at API rates without Codex Cloud or GitHub review.
    Codex model selection guideopens in a new tabDetails

  2. Claude Code

    Our team uses Claude Code. We recommend it for terminal-first repository edits, debugging failing test runs, and reviewing diffs in place when you want a direct loop over local files, Git history, and shell commands.

    Configuration
    Run in your project root with repo rules and local test commands, or pair with VS Code and JetBrains extensions for inline diff review.
    Tradeoffs & limits
    Excluded from Claude Free (requires Pro, Max, Team, Enterprise, Console, or a supported cloud provider), and third-party provider routing is limited to terminal and IDE clients.
    Claude Code quickstart guideopens in a new tabDetails

  3. Google Antigravity

    Dom started using agy at the Argon 4 Hackathon in San Francisco and likes it. With Gemini 4 Argon available in the desktop app, CLI, and IDE extensions, we recommend it for multi-step repository and agent tasks.

    Configuration
    Use the agy CLI, Antigravity App, or IDE extensions to select Gemini 4 Argon and adjust Low, Medium, or High thinking per turn.
    Tradeoffs & limits
    Gemini 4 Argon is not available in the standalone Antigravity IDE and requires paid Google AI Pro, Ultra, or eligible Enterprise access (or an eligible device Pro trial).
    Antigravity model availability docsopens in a new tabDetails

  4. Hugging Face Hub

    Our team uses Hugging Face for our own models. Use the Hub—not the separate Inference Providers router—when you need to inspect model cards and licenses, collaborate in Git-backed repos, or pull weights via hf CLI or Ollama.

    Configuration
    Run ollama run hf.co/{user}/{repo}:{quant} for GGUF repos; add ~/.ollama/id_ed25519.pub to Hugging Face settings for private GGUF access.
    Tradeoffs & limits
    The Hub hosts repos, not local inference; each checkpoint has its own license, format, and memory needs, and gated models require approved access (plus tokens for CLI or scripts).
    Hugging Face Hub documentationopens in a new tabDetails

  5. Ollama

    Our team uses Ollama to run our own models locally. It wraps local GGUF and Safetensors checkpoints behind a single CLI and OpenAI-compatible HTTP server so coding agents, eval runners, and apps can call local weights.

    Configuration
    Run Hub GGUF checkpoints directly with ollama run hf.co/... or import pre-quantized local GGUF files through a Modelfile with ollama create.
    Tradeoffs & limits
    Ollama does not quantize GGUF files during import, local throughput and context length depend on your machine's memory, and full cloud-catalog fallback requires paid credits.
    Use Ollama with Hugging Face Hub GGUFopens in a new tabDetails

Cross-tool workflows

Combining tools for a complete job

Six recipes from official docs showing how one tool's output feeds the next step, and where format or plan limits affect the handoff.

  • Run and evaluate Hugging Face weights locally

    Load a license-cleared GGUF model into Ollama and pass local Promptfoo checks before wiring the endpoint into your app.

    Starting input
    Target task prompt, Hub model card and license, compatible GGUF weights, and representative JSON/YAML test cases.
    Target output
    Pinned local Ollama model tag, Promptfoo assertion report, and an application module calling the local API.

    Tool handoffs

    1. Hugging Face Hub

      Verify the model card license, gated-access terms, and quantization files on Hugging Face Hub, then copy the repo path or download prepared GGUF weights.

    2. Ollama

      Load the model with ollama run hf.co/{user}/{repo}:{quant} (or ollama create from a Modelfile for pre-quantized local GGUF) and confirm it fits memory.

    3. Promptfoo

      Run promptfoo eval -j 1 against ollama:chat:<model> with task assertions, setting think: false when testing JSON schemas and keep_alive: 0 if RAM is tight.

    4. Codex

      Pass the pinned model tag, response schema, and failing edge cases to Codex to implement the application client and regression tests.

    Boundaries & checks
    Not every Hub repo ships GGUF or fits local RAM. Ollama does not quantize GGUF during import, and Promptfoo model-graded checks remain local only if you also set a local grader provider.
    Use Ollama with GGUF on Hugging Face Hubopens in a new tab

  • Animate a transparent image cutout from a voice brief

    Dictate the motion spec, strip a still image's background in Recraft, render alpha video in Remotion, and check playback in the browser.

    Starting input
    Spoken animation brief, a commercially licensed source image, and target light/dark UI backgrounds.
    Target output
    Downloaded alpha PNG/WebP cutout, transparent WebM (yuva420p) plus opaque fallback video, and browser QA screenshots.

    Tool handoffs

    1. Wispr Flow

      Dictate the animation timing, layout, and fallback requirements into your active Claude Code terminal or editor prompt as a plain-text brief.

    2. Recraft

      Call Recraft POST /v1/images/removeBackground on a rights-cleared raster asset and download the temporary PNG or lossless WebP cutout into the repo.

    3. Claude Code

      Have Claude Code wire the cutout into a React composition and app component with a transparent prop for alpha and opaque fallback renders.

    4. Remotion

      Render single-pass local WebM with --image-format=png --pixel-format=yuva420p --codec=vp8 (or ProRes 4444 for editors) plus an MP4 fallback.

    5. Playwright MCP

      Inspect the running localhost preview with Playwright MCP across light, dark, and reduced-motion states to confirm alpha compositing and playback.

    Boundaries & checks
    Recraft Free-plan images are public and noncommercial, API URLs expire in ~24 hours, Remotion preserves existing alpha rather than matting opaque video, and WebM alpha needs an opaque fallback.
    Remotion transparent videos guideopens in a new tab

  • Turn a PDF or Word spec into scoped code changes

    Parse the spec into auditable Markdown and JSON, locate the right symbols via LSP, and verify the code changes in a local browser.

    Starting input
    Specification PDF or DOCX, target repository with language-server support, and explicit acceptance criteria.
    Target output
    Source-linked Markdown/JSON spec, symbol-scoped repository diff, and verified test and browser results.

    Tool handoffs

    1. Docling

      Convert the specification PDF or DOCX with Docling into Markdown for reading and lossless Docling JSON to audit tables, formulas, and page references.

    2. Serena

      Initialize Serena MCP over the repository so the agent can look up symbol definitions, references, and file outlines instead of grepping raw text.

    3. Claude Code

      Supply the reviewed requirements to Claude Code to edit the target symbols and run unit tests against the specification's acceptance criteria.

    4. Playwright MCP

      Run Playwright MCP against the local preview to verify that the rendered UI states and form flows match each requirement in the spec.

    Boundaries & checks
    Check OCR and extracted tables against the source PDF before coding, treat document text as untrusted data rather than scope authority, and ensure Serena's language server is initialized.
    Docling supported formats documentationopens in a new tab

  • Test an AI feature locally and trace it in production

    Check your application handler with local Promptfoo assertions, trace live calls in Langfuse, and feed failing inputs back into your test suite.

    Starting input
    Application handler or prompt module, representative pass/fail examples, and model-provider credentials.
    Target output
    Versioned Promptfoo config and wrapper, OpenTelemetry trace spans in Langfuse, and updated regression cases.

    Tool handoffs

    1. Codex

      Use Codex to expose your application's actual prompt and tool-handling function through a thin Python provider wrapper (call_api) and add trace spans.

    2. Promptfoo

      Run npx promptfoo@latest eval against file://provider.py with deterministic schema and rubric assertions to catch regressions before deployment.

    3. Langfuse

      Instrument backend calls with the Langfuse SDK, inspect failing traces by version, and add redacted failure inputs to your Promptfoo regression suite.

    Boundaries & checks
    Promptfoo runs locally, but hosted targets or graders still receive data and bill tokens. Langfuse v5/v6 SDKs require Langfuse v4.59.0+, and turning traces into Promptfoo tests is a manual handoff.
    Langfuse observability SDK overviewopens in a new tab

  • Move a marimo notebook into a durable background job

    Validate data logic in a live marimo notebook, extract a tested Python function, and wrap each side effect in checkpointed Inngest steps.

    Starting input
    Sample dataset or schema, a running marimo notebook, and an event trigger or retry specification.
    Target output
    Validated marimo notebook, tested standalone transformation module, and an Inngest step.run() workflow.

    Tool handoffs

    1. marimo

      Start uvx marimo@latest edit notebook.py and connect your agent via marimo pair so it can inspect live dataframe schemas, test scratchpad logic, and run cells.

    2. Claude Code

      Have Claude Code extract the verified notebook cells into a pure repository function with unit tests and fixture checks that run without notebook state.

    3. Inngest

      Wrap each fetch, transform, and idempotent write in Inngest step.run() checkpoints so failed steps retry without rerunning completed earlier steps.

    Boundaries & checks
    marimo pair and Agent/Code modes can inspect live runtime state (and remote LLMs receive shared context). Inngest retries failed steps, so external writes inside step.run() must be idempotent.
    Pair with agents via marimo pairopens in a new tab

  • Matte a generated video clip into a transparent UI loader

    Generate a short motion clip in Higgsfield, strip the video background into alpha WebM in ComfyUI, and render browser-ready outputs in Remotion.

    Starting input
    Full-body key-pose image on a contrasting solid background, motion prompt, UI dimensions, and Comfy account credits.
    Target output
    ComfyUI SaveWEBM (vp9, yuva420p) alpha clip, Remotion transparent WebM + MP4 fallback, and Playwright browser checks.

    Tool handoffs

    1. Higgsfield

      Generate a 4s 720p MP4 with start=end pose, locked camera, and contrasting background; scrub to verify loop wrap (not guaranteed) and download within 7 days.

    2. ComfyUI

      In current/nightly ComfyUI (credits + API access), wire LoadVideo (MP4 <=60s) to Bria Remove Video Background (Transparent) (BriaTransparentVideoBackground).

    3. ComfyUI

      Feed Bria images+inverted mask to JoinImageWithAlpha (RGBA); wire RGBA + LoadVideo->GetVideoComponents fps to SaveWEBM (codec=vp9, yuva420p; av1 drops alpha).

    4. Remotion

      Pass VP9 WebM to <Video> from @remotion/media (no opaque background); render 1-pass local WebM (--image-format=png --pixel-format=yuva420p --codec=vp8) + MP4.

    5. Playwright MCP

      Open the localhost preview with Playwright MCP, emulate light/dark themes and reducedMotion, and check alpha compositing, loop wrap, and MP4 fallback playback.

    Boundaries & checks
    Untested recipe (no .riv/.lottie). Bria Partner Node uploads clips <=60s (needs current/nightly ComfyUI, positive credits, API access). AutoSprite is separate; Remotion keeps alpha (1-pass + MP4).
    ComfyUI Bria Partner Nodes tutorialopens in a new tab

Browse the collection

Evals & observability

Evaluation runners & prompt optimizers

DSPy

Declarative Python framework for building modular language-model pipelines and running algorithms that compile and optimize prompts or model weights against metrics.

dspy.aiDetails

DeepEval

Runs Pytest-style LLM evaluations for outputs, agent trajectories, and traced components, with a skill workflow for coding agents to iterate on failing tests.

deepeval.comDetails

Promptfoo

Evaluates and red-teams LLM apps via a local CLI, library, or CI workflow, running declarative comparisons across prompts, models, and RAG pipelines.

promptfoo.devDetails

Ragas

Evaluates RAG retrieval and answer quality, agent tool calls, and text-to-SQL workflows, with synthetic test-data generation and dataset experiments.

docs.ragas.ioDetails

Tracing, datasets & experiment platforms

Arize Phoenix

Provides open-source local and self-hosted AI observability with OpenTelemetry/OpenInference tracing, versioned datasets, LLM evaluations, and prompt experiments.

arize.comDetails

Braintrust

Connects AI tracing, human annotation, evaluation datasets, and experiments across a web UI, bt CLI, MCP server, and the Loop trace-analysis agent.

braintrust.devDetails

Langfuse

Traces model calls and agent workflows, tracks cost and latency, versions prompts, and runs dataset evaluations via web UI, API, CLI, or MCP server.

langfuse.comDetails

LangSmith

Captures execution traces across supported AI frameworks and SDKs, builds evaluation datasets from runs, monitors quality, and collects human feedback.

langchain.comDetails

W&B Weave

Traces LLM application and agent calls, versions instrumented code, and evaluates outputs using LLM judges or custom scoring functions.

W&B Weave docs route to CoreWeave: Agent Lens is the agent-tracing successor sharing trace ingestion, while Weave function and evaluation workflows remain.

docs.coreweave.comDetails

Incident & telemetry debugging agents

HolmesGPT

Open-source SRE agent that investigates production incidents, telemetry signals, and deployment issues from a standalone CLI or self-hosted operator.

holmesgpt.devDetails

Sentry Seer

Analyzes production issues, plans fixes, and reviews code by combining Sentry application telemetry with connected source repositories.

sentry.ioDetails

How this collection is maintained

We mark the five tools our team uses, while all profile facts, release notes, and workflow recipes are compiled from official docs and pricing pages rather than controlled testing. Check the linked sources for current plans and limits.

Original historical article
Suggest a tool or update