AtlasBurn
DocsPlatformIntegrationsOperationsReference

Changelog

Release notes and improvements for AtlasBurn.

Integrate AtlasBurn from Claude · SDK v1.11 — October 2026

  • “Integrate AtlasBurn” in Claude: two new MCP tools, request_integration_key and get_setup_code, let Claude Code set AtlasBurn up in your project. You approve each new key in AtlasBurn with a fresh sign-in (same account that connected Claude); Claude then runs the CLI with a one-time setup code. The key goes straight to your env file and never appears in the chat. New keys are added, never replace existing ones, are capped at 5 active, and show as “via and show as “via Claude” in System Controlslt;agentand show as “via Claude” in System Controlsgt;” (e.g. “via Claude Code”, “via Cursor”) in System Controls. See Integrate AtlasBurn from Claude.
  • SDK 1.11.0 CLI: abn init --code redeems a setup code; --proxy sets up the Edge Proxy for any language (key + per-provider base URLs, no package.json needed); --yes takes ATLASBURN_KEY from the environment. Next.js instrumentation now runs in the Node.js runtime only, and a new .gitignore also covers node_modules/.
  • SDK 1.11.1 fixes plain Node and CommonJS apps: abn init wrote the key to .env, but plain Node never reads .env, so the SDK stayed off on npm start. The generated file now reads it itself (no dotenv; platform-set keys still win). CommonJS apps couldn't load the SDK at all; they now can on Node 20.19+ / 22.12+. Re-run npx abn init to upgrade an older setup file.
  • Complete history in MCP answers: cost tools now cover the whole window you ask for, built from server-side daily summaries. Previously they used only your most recent 500 requests, so busy accounts saw partial totals.

Throttle first on Kill Switch · SDK v1.10.0 — October 2026

  • Kill Switch throttles before it blocks: budget breaches still block at once, but loops and request-rate bursts are throttled first and blocked only if they are still happening 60 seconds later or cost more than your loop-block limit (2% of the daily budget, min $0.25, or your own amount). Projected breaches and spending anomalies only throttle. See How each tier responds.
  • Strict mode: a Kill Switch option that blocks every enforceable trigger at once, for teams that want the old behavior.
  • SDK 1.10.0 makes throttling real: a throttled feature's calls are spaced 2 seconds apart (throttleSpacingMs) instead of passing straight through. A queue deeper than throttleMaxWaitMs (10s) gets a blocked response so your app never stalls. Alert mode never slows a call.
  • Fix: on Kill Switch, projected breaches and spending anomalies used to hard-block for 2 hours; they now throttle unless Strict is on.
  • Fix: a loop that had stopped could be detected again from its last calls when other features sent traffic, extending its block. Detection now only looks at features with new calls.
  • Alerts now say which step you are on, and why a throttle escalated to a block.

Feature-scoped guardrails — October 2026

  • Loops pause one feature, not your whole app: the identical-request and runaway-context detectors now run per feature and stop only the feature that looped. Every other feature keeps running. Budget breaches still stop the whole project. See What gets blocked.
  • Per-feature Resume: Safety Guardrails lists each paused feature with its own countdown and Resume button. POST /api/guardrails/enforce accepts an optional featureId.
  • Feature tags on the Edge Proxy: send x-atlasburn-feature to attribute cost and contain loops per feature; the proxy strips it before forwarding.
  • Clearer blocks and alerts: the blocked response's reason is now the real trigger (it previously always said budget_breach), and alerts plus the MCP get_guardrail_status tool name the paused feature.
  • Works with SDK 1.9.3 as is. No upgrade needed if you already tag calls.

SDK v1.9.3 — October 2026

  • Per-call feature tags: atlasburnFeature("name") passed as request headers tags one call; the header is removed before the provider sees it. metadata.featureId sets a per-process default.
  • Faster guardrail gate: decisions are cached per feature for 10s and refreshed in the background; a check waits at most 2.5s before failing open; a failed check never lifts an active block.
  • Blocked calls fail fast: blocked responses carry x-should-retry: false, so the OpenAI and Anthropic SDKs no longer retry them for ~2 minutes.
  • Scripts exit cleanly: the flush timer no longer keeps short Node scripts (and npx abn init) alive.

SDK v1.8.0 — September 2026

  • The abn CLI — the SDK now ships a zero-copy-paste setup CLI (npx abn init): it detects your framework, auto-detects providers, takes your key via hidden input, shows a diff and asks before writing, wires instrumentation.ts (Next) or an entry import (Node), and runs a verification test. Also abn test, abn status, and abn doctor. Zero new dependencies.
  • MCP Server — AtlasBurn is now a public, OAuth 2.1 (PKCE) Model Context Protocol server, so Claude, ChatGPT, and Cursor can read your cost, runway, guardrail, and runtime data in-chat. Read-only, content-blind; published in the official MCP registry as com.atlasburn/atlasburn.
  • Verification events — verifyAtlasBurn() now tags its pulse as a verification event (billed the nominal $0.00001) so it's clearly distinct from production telemetry in the dashboard.

SDK v1.7.2 — September 2026

  • Expanded auto-detection (I1) — the SDK now auto-captures Azure OpenAI, OpenRouter, Cohere, Mistral, Groq, Together AI, DeepSeek, xAI, and AWS Bedrock (best-effort) alongside OpenAI, Anthropic, Gemini, and Vertex AI. Only a self-hosted model with no fixed public host still needs the proxy or /api/ingest.
  • Live model pricing — the Model Intelligence catalog refreshes list prices from a maintained provider registry every 3 days, with per-model manual pins; the static registry stays the fallback so a price is never blank.
  • Multi-channel alerts — guardrail breaches now fan out to Email, Slack, Discord, and Telegram with per-channel severity floors (SMS / WhatsApp planned).

SDK v1.6.1 — June 2026

  • Streaming usage capture (I3) — the SDK and Edge Proxy now inject stream_options.include_usage: true on OpenAI streams so the final SSE frame carries real input/output token counts. Pre-I3 streamed calls logged $0; that bug is fixed.
  • Estimated-token flag — events without a usage frame are now flagged est/~est in the Forensic Ledger.
  • License — repackaged under Apache-2.0. See License.
  • Docs — withdrew the legacy 94.2% accuracy figure pending re-derivation against current invoices.
  • Forecasting — engine outputs documented as p5 / p50 / p95 + VaR95 / CVaR95; the spurious P99 reference was removed.

v2.0.0 — May 2026

  • AtlasBurn Edge Proxy — new Cloudflare Worker integration. Zero code changes, native SSE streaming, fails open by design.
  • Autonomous Runtime Guardrails — server-side 5-layer defense engine replaces SDK-level Auto Kill:
    • Layer 1 — Financial Limits (daily / monthly hard cap)
    • Layer 2 — Dumb Loop Detector (identical token fingerprinting)
    • Layer 3 — Smart Loop Detector (monotonic context inflation)
    • Layer 4 — RPS Burst Protection (per-key frequency limits)
    • Layer 5 — Premium Model Concentration (prompt-injection / hijacked routing detection)
  • Cloudflare KV Kill Switch — <5ms edge enforcement with TTL auto-recovery and manual dashboard override.
  • New API endpoints: POST /api/ingest (telemetry collection) and POST /api/guardrails/enforce (KV flag webhook).
  • OpenRouter auto-detect via the Edge Proxy. Layer 4 is the primary safety net due to tokenizer drift from upstream load-balancing.

v1.5.0 — April 2026

  • Synchronous loop guards — added isRouting guard to prevent proxy recursion (RangeError: Maximum call stack size exceeded) in Next.js App Router
  • Universal Framework Support — initAtlasBurnAuto now survives arbitrary globalThis.fetch overrides by host frameworks (Next.js, Nuxt, SvelteKit)
  • Improved SSE stream chunk parsing for complex AI provider outputs (multi-chunk usage metadata, partial JSON frames)
  • Single-init integration via Next.js Instrumentation hook (instrumentation.ts)
  • Streaming and JSON responses captured automatically across Server Components and API Routes
  • Added Vertex AI auto-detect support alongside OpenAI, Anthropic, and Gemini
  • Deprecated route-level patching and custom fetch wrappers — safe to delete

v1.2.0 — March 2026

  • Added OpenRouter auto-detect support
  • Burn forecasting upgraded with tail-risk modeling (VaR95 / CVaR95)
  • Improved retry cascade detection with configurable thresholds
  • SDK: reduced cold-start overhead by 40%

v1.1.0 — February 2026

  • Gemini auto-detect support (1.5 Pro, 1.5 Flash)
  • Reconciliation tool for provider invoice comparison
  • New metadata field for custom context on events
  • Dashboard: per-model cost breakdown view

v1.0.0 — January 2026

  • Initial release
  • Auto-detect support for OpenAI and Anthropic
  • Forensic ledger with append-only, server-authored audit trail
  • Real-time observed spend calculation
  • Fail-silent SDK behavior
  • Manual instrumentation mode