The AI Toolchain — issue 019, September 7, 2026

Adds a Select tool in the chatbox that lets you click any preview element and edit text, colors, font size, alignment, opacity, object fit, alt text, links, and list formatting directly in the UI without writing prompts. Edits are batched above the chatbox and only consume tokens when you click Save changes; CMD+click (Mac) or CTRL+click (Windows) selects a child element inside a selected parent, with arrow-key navigation between nested elements. An optional batch note field lets you type cross-page instructions (e.g. apply a style change to matching elements on other pages) before saving, and a saved batch of visual edits can be rolled back via the project's version history when individual undo isn't available.

Adds MCP (Model Context Protocol) server connectors, letting Bolt integrate with external tools like Notion, Linear, and GitHub via built-in connectors or a custom MCP server URL, accessible via the plus icon in the chatbox under Connectors > Manage connectors.

Adds AI image generation directly in the Bolt chatbox, with transparent background support and automatic WebP conversion; requires enabling Image Generation under Personal Settings > Add-on features.

A new branch menu gives access to the latest preview deployment, a code diff against the base branch, PR creation/open/merge, CI check and repository-rule status, pulling the base branch, and in-chat fix requests for preview, CI, or merge issues. The unified Publish action lets v0 create or reuse a pull request, wait for required checks and reviews, merge into the base branch, and deploy to production, pausing to link directly to the PR if a branch protection rule needs human action.

Raises the maximum number of MCP servers per scope from 10 to 100; in v0 settings, navigate to MCP Servers for your scope to add up to 100 servers.

Adds a 'Team or password' visibility option for published apps, requiring Vercel team sign-in while still allowing password-based access for others. Set via Share/Publish settings by choosing this visibility option and saving.

Generated apps get zero-config authentication to the AI Gateway covering text, structured output, embeddings, reranking, images, video, speech, and transcription — no provider or Gateway API keys required.

Introduces Origin, a built-in code hosting service with repos, pull requests, code browsing, and GitHub sync, available in early beta on all paid plans. Adds a Codebase tab as the home for Origin repos with a +New button to create repos whose names become part of every repo URL at cursor.com/codebase/<name>, and lets Cloud Agents start without a connected GitHub or third-party SCM provider via 'Start from scratch' in the repo picker, saving work to a Cursor Origin repo. Adds pull request sync between Origin and GitHub so comments, reactions, and replies made in either surface appear in the other within seconds, app extensions (Vercel, Depot, Buildkite) to Origin repos via a per-repo Apps tab enabling preview deployments and CI on every PR, and per-repo Settings for managing GitHub sync status, access control, and connected apps.

Introduces self-hosted machines so tool execution, codebase, build outputs, and secrets stay entirely within your own network. Adds dynamic pool scheduling via My Machines (single laptop/VM) and Team Pools (named queues of workers) that auto-scale and hibernate idle machines, plus support for running cloud agents on pre-existing sandbox infrastructure including AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B. Also enables computer use (click, type, screenshot, browser control) on Linux and Mac for self-hosted workers, with live desktop viewing and takeover from Cursor.

Adds a Sync Skills for Cloud Agents toggle under Settings → Agents → Context and Tools that lets individuals enable their personal skills for use in Cloud Agents, and a team admin control under Team Settings → Security & Identity to disable this org-wide for Teams and Enterprise plans. Also adds a team skill marketplace where members publish personal skills from ~/.cursor/skills/ to a shared Default marketplace via Customize → Skills → Publish, an Allow Members to Publish toggle in Dashboard → Plugins that restricts who can publish, and the ability for admins to mark a published plugin 'Default On' or 'Required' to control installation across the team.

Adds /goal command to give a cloud agent a long-lived objective that it pursues until fully complete — pair with a custom mode or /loop for recurring check-ins.

Adds Custom Modes, which pin any skill as an always-on mode in chat, activated via ⌥⏎ (Mac) or Alt+Enter (Windows) from the skill picker.

Introduces Builds for Cloud Agents: pre-prepared environment snapshots (repos cloned, dependencies installed) that make agents start 3x faster, with a Builds tab in the Cloud Agents dashboard for inspecting build status, logs, and commit SHAs.

Adds steering improvements allowing mid-run follow-up messages that wait for the next tool call rather than interrupting the agent — send via 'Send now' or by pressing ⏎ twice.

Arcade's /mcp path now supports per-request protocol-version dispatch and a stateless server/discover endpoint, plus stateless tools/list and tools/call against the real catalog. The MCP /list page size is now configurable, and MRTR url-mode elicitation was added for tool OAuth challenges on the stateless MCP path.

$ curl -X GET 'https://api.arcade.dev/mcp/server/discover' \
  -H 'Authorization: Bearer <token>'

Arcade adds the microsoft_powerbi toolkit for Power BI integrations, a Postman API toolkit with 30 tools and API-key authentication, an Insightly CRM toolkit with 29 tools for sales and account management, a Fireflies.ai meeting-intelligence toolkit with 17 functionalities, a Microsoft OneDrive toolkit with 16 tools for file management, a Freshdesk customer-support toolkit with API-key authentication, an Ashby recruiting/ATS toolkit for core recruiting operations (including a pipeline exit tool to archive applications), a Vercel toolkit for managing accounts via the REST API, a Datadog observability toolkit for log management, and Nimble as a Partner MCP Server for real-time web data in the search integrations catalog.

Google Workspace toolkits gain new tools: GetSpreadsheetEditHistory in the google_sheets toolkit, a secure Gmail unsubscribe tool with multiple flow outcomes and enhanced validation, new Gmail tools for label lifecycle management and conversation trash status handling, a search_directory_people tool for resolving coworkers' emails, a search_drive tool to search the user's entire Google Drive, and a check_file_access pre-flight probe plus a file_types picker filter for Drive access.

Validates SERVER_INTERNAL_URL at worker startup and requires an http or https URL, refusing to start and naming the variable when wrong. Validates HostPort with digit and range checks, refusing supervisor startup on an invalid port. Treats an empty PORT value on agent-bot and agent-langgraph as the shipped default rather than parsing it as NaN and binding an ephemeral port. Refuses malformed IPv6 entries in AGENT_ENDPOINT_ALLOWED_HOSTS at boot instead of silently accepting entries that never match. Reads durationMs units case-insensitively, accepting 30S alongside 30s.

Reports a malformed DATABASE_URL password containing a bare % with a structured error naming the variable and the required %25 encoding, instead of crashing with URIError: URI error, and refuses any DATABASE_URL value that cannot be parsed without echoing the password into the log.

Validates COMPUTER_MEMORY_BYTES at supervisor startup and refuses non-integer values, preventing a misconfigured cap such as 512m — previously parsed via parseInt into 512 bytes — from ever reaching a container.

Validates from, to, and limit parameters on the audit endpoint strictly, returning HTTP 400 for an unparseable date or non-integer limit instead of coercing and querying.

$ curl -G 'https://<openbot-host>/audit' \
  --data-urlencode 'from=2024-06-01T00:00:00Z' \
  --data-urlencode 'to=2024-06-30T23:59:59Z' \
  --data-urlencode 'limit=100' \
  -H 'Authorization: Bearer <token>'

Adds a serve script that builds the app once and serves the bundle, replacing the dev-only dev script as the path to a production deployment.

Adds the computer.action_stopped audit event type to distinguish a user-initiated Stop from a genuine computer-unreachable failure, enabling accurate outage monitoring by type. Previously all stops were recorded under computer.action_failed, so monitoring that counts failures by that type will no longer count deliberate stops after upgrading.

Refuses titles, slugs, and IDs that are entirely whitespace in sandboxed components, skills, custom servers, and the component catalogue, and trims name, title, kind, and description fields before storing them in the catalogue, so a component with a stray space in its name is treated as the same component rather than creating a duplicate row.

Adds a moderation object to POST /router/v1/chat/completions responses, with input and output sub-objects each carrying type, model, flagged, categories, category_scores, and category_applied_input_types fields for content moderation results.

$ curl --request POST \
  --url https://api.perplexity.ai/router/v1/chat/completions \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{"model": "<model>", "messages": [{"role": "user", "content": "<prompt>"}]}' \
  | jq '{input_flagged: .moderation.input.results[0].flagged, output_flagged: .moderation.output.results[0].flagged, categories: .moderation.output.results[0].categories}'

The POST /router/v1/chat/completions response gains reasoning_content on message for chain-of-thought exposure, an annotations array with url_citation entries (start_index, end_index, url, title) for inline source attribution, and an audio object (id, expires_at, data, transcript). usage gains completion_tokens_details (accepted_prediction_tokens, audio_tokens, reasoning_tokens, rejected_prediction_tokens) and prompt_tokens_details (audio_tokens, cached_tokens, cache_write_tokens), plus top-level service_tier and system_fingerprint fields.

$ curl --request POST \
  --url https://api.perplexity.ai/router/v1/chat/completions \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '{"model": "<model>", "messages": [{"role": "user", "content": "<prompt>"}]}' \
  | jq '.usage | {prompt_tokens, completion_tokens, total_tokens, reasoning_tokens: .completion_tokens_details.reasoning_tokens, cached_tokens: .prompt_tokens_details.cached_tokens}'

Operators choosely how models are kept in memory at darkbloom start: 'Always ready' keeps models loaded, 'Free when idle' unloads them after 60 minutes (the default), or 'Custom' lets the user define their own idle window.

$ darkbloom start

darkbloom models catalog now surfaces required_provider_capabilities per model and prints eligibility verdicts based on detected hardware, e.g. requires: Apple M5, NAX runtime or 'not eligible on this machine (missing: apple_m5, mlx_nax)'.

$ darkbloom models catalog

On API version 2026-07, POST /indexes now creates schema-based document indexes: a schema object combines BM25 full-text, dense-vector, and sparse-vector ranking fields in a single index, replacing the previous top-level dimension, metric, vector_type, and spec fields (which will break existing index-creation code on 2026-07). The Documents API supports BM25 token matching, Lucene query syntax, and vector similarity queries with metadata and text-match filters, and adds fetch, update, and delete of documents by metadata filter in addition to by ID.

$ curl -X POST https://api.pinecone.io/indexes \
  -H 'Api-Key: <your-api-key>' \
  -H 'X-Pinecone-API-Version: 2026-07' \
  -H 'Content-Type: application/json' \
  -d '{
    "name": "my-doc-index",
    "schema": {
      "fields": [
        {"name": "body", "type": "text", "ranking": "bm25"},
        {"name": "embedding", "type": "dense_vector", "dimension": 1536, "metric": "cosine"}
      ]
    }
  }'

Reads via query, fetch, list, and search are now metered for egress, with monthly allowances of 1 GB (Starter), 10 GB (Builder), and 100 GB (Standard and Enterprise). Standard and Enterprise plans continue serving reads beyond the allowance at an overage rate, while Starter and Builder reads are blocked until the next billing period.

Helicone AI Gateway is a new open-source Rust-based proxy available via docker run -p 8080:8080 helicone/ai-gateway at github.com/helicone/ai-gateway, supporting smart load balancing, semantic caching, automatic failover, and rate limiting across OpenAI, Anthropic, Azure, and more providers.

$ docker run -p 8080:8080 helicone/ai-gateway

Prompt Management V2 adds version control, typed variables in system prompts, messages, and tool schemas, and instant deployment through the AI Gateway. AI Gateway API calls now accept prompt_id and inputs fields, enabling prompt references by ID with runtime variable substitution (e.g., {{hc:customer_name:string}}) so prompts can be updated in the dashboard without code changes.

Introduces a Go SDK for the Helicone Helpers Package, installable via go get github.com/helicone/go-helicone-helpers, enabling custom LLM integration with the Go Manual Logger.

$ go get github.com/helicone/go-helicone-helpers

Adds self-hosting support via a single docker compose up -d command from the github.com/Helicone/helicone repository, keeping all LLM interactions and analytics within your own infrastructure.

Adds cost tracking for GPT-5, GPT-5-Mini, GPT-5-Nano, and GPT-5-Chat-Latest models across multiple providers with Playground support; for GPT-OSS models (accounts/fireworks/models/gpt-oss-20b, accounts/fireworks/models/gpt-oss-120b, openai/gpt-oss-20b, openai/gpt-oss-120b) across Fireworks, Groq, and OpenRouter; for Claude Opus 4.1 (claude-opus-4-1-20250805, anthropic/claude-opus-4.1 on OpenRouter) and grok-4 / x-ai/grok-4; and for gpt-4o-search with automatic pricing detection.

Adds a country_code:<ISO> filter (e.g., country_code:US) to the requests dashboard advanced filters section for geographic analytics on LLM usage.

Extends Claude Sonnet 4 and Sonnet 4.5 on the AI Gateway to support a 1M token context window by default across Anthropic API, AWS Bedrock, and Google Vertex AI.

The thread evaluator config now shows a locked 'Trace count >= 2' filter for managed thread evaluators, making explicit that they only run on threads with multiple turns. Evaluators can be validated against a real conversation before saving via POST /runs/rules/validate with test_thread_id, session_id, and rule parameters, catching misconfiguration early.

$ curl -X POST 'https://api.smith.langchain.com/runs/rules/validate' \
  -H 'X-Api-Key: <your-api-key>' \
  -H 'Content-Type: application/json' \
  -d '{"test_thread_id": "<thread-uuid>", "session_id": "<session-uuid>", "rule": { ... }}'

Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; existing HTTP routes continue to work only for LangSmith UI clients, so integrations relying on the SDK methods must migrate.

TypeScript SDK v3.30.0 adds openaiFilesCreateTraced, openaiBatchesRetrieveTraced, and completeOpenAIBatchTrace to trace asynchronous OpenAI Batch API requests end-to-end, so batch inference jobs appear as spans in Braintrust experiments.

import { openaiFilesCreateTraced, openaiBatchesRetrieveTraced, completeOpenAIBatchTrace } from 'braintrust';

const file = await openaiFilesCreateTraced(openai, {
  file: fs.createReadStream('requests.jsonl'),
  purpose: 'batch'
});
const batch = await openai.batches.create({
  input_file_id: file.id,
  endpoint: '/v1/chat/completions',
  completion_window: '24h'
});
await openaiBatchesRetrieveTraced(openai, batch.id);
await completeOpenAIBatchTrace(openai, batch.id);

Loop now runs in a Braintrust-managed runtime, persists threads, and operates across logs, experiments, and datasets project-wide. It can create and edit prompts, scorers, datasets, facets, custom views, dashboards, and automations, with a per-write approval step unless auto-accept is enabled. Loop also runs built-in on GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna models with no AI provider configuration required.

Calling auto_instrument() automatically instruments all Agno eval types (AccuracyEval, AgentAsJudgeEval, ReliabilityEval, PerformanceEval) so their runs appear as Braintrust experiments without manual tracing setup.

from braintrust import auto_instrument
from agno.evals import AccuracyEval, AgentAsJudgeEval, ReliabilityEval, PerformanceEval

auto_instrument()

eval_suite = [
    AccuracyEval(agent=my_agent, questions=questions),
    AgentAsJudgeEval(agent=my_agent, rubric=rubric),
    ReliabilityEval(agent=my_agent, scenarios=scenarios),
    PerformanceEval(agent=my_agent, tasks=tasks),
]
for e in eval_suite:
    e.run()

GitHub eval action v2.1.0 adds report_scores and report_metrics inputs, each accepting a comma- or newline-separated list of names to filter what appears in the PR comment.

Patterns is a scheduled automated investigation of traces that surfaces recurring failure modes, cost trends, and cohort-level problems. Each finding includes supporting traces and a suggested fix, and can be used to create a scorer, continue in Loop, or copy as a prompt.

Debugger reports likely failure modes for a single trace, citing spans, tool calls, and model outputs, and groups a run into 'Work' sections via 'Analyze trace' to explain what the agent did.

Python SDK v0.37.0 adds invocation parameters (temperature, max_tokens, and similar model settings) to Pydantic AI LLM span metadata when model settings are provided.

Adds an MCP (Model Context Protocol) provider to connect Promptfoo evals to MCP servers, plus an MCP Agent example for red team testing that demonstrates how to test AI systems using tool call results via MCP.

Kits can now declare arguments in an args: block and receive values with --kit-arg name=value or --kit-arg kit.name=value to target a specific kit; a kits: entry in sbxenv.yaml can carry per-kit arguments under kits[].args.

$ sbx run docker.io/myorg/pentest-kit:latest --kit-arg target=192.168.1.0/24 --kit-arg mode=aggressive

sandbox:
  name: pentest-env
kits:
  - ref: docker.io/myorg/pentest-kit:latest
    args:
      target: 192.168.1.0/24
      mode: aggressive

Adds a --json flag to sbx secret ls, sbx version, sbx mcp ls, sbx skills ls, sbx policy inspect, and the sbx kit verification commands for machine-readable output; sbx version --json also reports a server.state field (running or unavailable) so scripts can check daemon reachability without parsing error text.

$ sbx version --json | jq '.server.state'

sbx ports --publish and kit-declared ports now default to tcp4 instead of dual-stack tcp, so published ports no longer listen on ::1; pass the protocol explicitly (e.g. --publish 8080:3000/tcp) to restore dual-stack behavior.

sbx mcp auth now requests only the chosen scopes (or those the resource requires) instead of a broad default set, accepts a new --no-scope flag to suppress server-required scopes, explains refused scopes with a narrower retry command, and reports granted/requested/supported scope sets in sbx mcp auth status.

sbx create can now omit the workspace path to create a sandbox with no workspace bind mount; correspondingly, an environment file that declares no workspace: now creates a sandbox with no bind mount instead of mounting the directory holding the file — add workspace: . to restore the previous behavior.

Environment files (sbxenv.yaml) can now declare their own arguments in an args: block, referenced as ${{ env.args.NAME }} and supplied with sbx env --env-arg; this replaces the previous ${VAR} interpolation syntax, which is no longer expanded.

The built-in droid, copilot, and kiro agents are removed and must now be run as community kits, e.g. sbx run docker.io/sbx/droid-kit:latest, sbx run docker.io/sbx/copilot-kit:latest, or sbx run docker.io/sbx/kiro-kit:latest.

sbx env now shows a plan of all changes an environment file will make to the host (lifecycle commands, credentials, bindings, MCP servers, workspaces, kits, ports) and prompts for approval unless env.rememberHostCommands is set.

Sandbox default Docker volume size is reduced to 10 GB (from 50 GB); configurable via the sandbox.disk.dockerVolume setting or the DOCKER_SANDBOXES_DOCKER_SIZE environment variable.

sbx mcp add gains a --skip-auth flag (replacing --skip_auth) and now allows --url values pointing at private, loopback, or cloud-metadata addresses, printing a warning instead of refusing.

Adds an sbx run <sandbox-kit-ref> positional form so kit references can be passed directly as the agent argument; the --kit flag is now reserved for mixins.

The Admin API is now available in the ant CLI and Python, TypeScript, C#, Go, Java, PHP, and Ruby SDKs under client.beta.organization, covering organization info, members, invites, workspaces, API keys, rate limits, service accounts, workload identity federation issuers and rules, and customer-managed encryption keys; it reads credentials from ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN.

With the mid-conversation-output-config-2026-07-01 beta header, a role: "system" message inside messages can include output_config.effort to change effort mid-conversation on Claude Fable 5.1, Claude Mythos 5.1, and Claude Opus 5 while preserving the prompt cache.

The mid-conversation-system-clear-at-2026-08-21 beta header lets a mid-conversation role: "system" message set clear_at: "next_user_message" so it applies only to the current turn and remains in history at no token cost.

The ant CLI (v1.30.0) adds ant apply to create and update agents, environments, skills, memory stores, and deployments directly from repository files, writing a claude-lock.json lockfile to keep CI runs reproducible.

The thinking-binding-controls-2026-08-01 beta header reports dropped thinking blocks in an input_transformations response field and adds thinking.block_binding.prefix_mismatch_behavior to choose between rejecting or dropping blocks whose conversation history changed.

The Compliance API's local session endpoints (beta, Claude Enterprise) now return transcripts for Claude Science sessions (product_surface value claude_science) and Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook (product_surface values beginning with office_agents), using the existing Compliance Access Key and read:compliance_user_data scope; the same session endpoints for Cowork and Claude Code sessions are promoted out of beta.

In Python SDK 1.2.0, TypeScript SDK 0.122.0, Go SDK 1.68.0, Java SDK 2.59.0, Ruby SDK 1.67.0, and C# SDK 12.44.0, client.beta.files and client.beta.skills no longer send the files-api-2025-04-14 and skills-2025-10-02 beta headers and return the same response shapes as client.files and client.skills; client.beta.skills.delete() now deletes a Skill together with all of its versions, and the beta Messages type BetaSkill is renamed BetaContainerSkill.

claude-fable-5-1 launches with a 1M token context window, 128k max output tokens, always-on adaptive thinking, priced at $10/$50 per MTok input/output with prompt cache reads at $0.25 per MTok (0.025x base input price); claude-mythos-5-1 ships the same context, output, thinking, and pricing for Project Glasswing participants.

With the thinking-display-updates-2026-08-18 beta header, thinking.display accepts the value "updates", returning short progress updates Claude writes between tool calls as text, with at most one thinking block before each tool call.

The computer_toolset_20260801 and browser_toolset_20260801 toolsets are now available on Google Cloud for Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8.

On Claude Fable 5.1 and Claude Mythos 5.1, tool_choice types any and tool are not supported and return a 400 error; only auto and none are accepted.

For accounts created on or after August 31, 2026, replaying a thinking block on Claude Fable 5.1 after the system prompt, tools, or an earlier message has changed now returns a 400 error, and thinking blocks produced by Claude Fable 5.1 and Claude Mythos 5.1 are dropped by the API if replayed to an earlier model.