High-risk tools in Nodebench

Run the same prompt against multiple LLM providers and compare responses. Returns side-by-side results with latency, token usage, and a summary. Useful for model selection, qual...

Detect whether a target project repo has parallel agent infrastructure and, if not, scaffold everything needed. Scans for task coordination, role configs, oracle testing, contex...

Build a banker-readiness packet from the canonical company packet.

Build a starter PitchBook/Crunchbase-like company profile.

Build the complete founder operating model: execution order, queue topology, packet routing, source trust policy, progression rubric, and benchmark oracles.

Generate a digest of new (unseen) articles from RSS feeds. Compares against previously seen articles via SQLite. Returns only new items grouped by category. After generating, ar...

Build the exact pull/subscription manifest an agent client should use to watch a packet or packet scope.

Build a filtered snapshot/events/pull manifest for one peer, packet class, producer, scope, or subject so clients can subscribe to only relevant shared-context updates.

Build a generic submission export from the canonical company packet.

Build a temporal relationship graph for an entity.

Capture N sequential screenshots at fixed intervals using Playwright.

Invoke a tool on a connected MCP driver. This proxies the call to the external MCP server (e.g. playwright-mcp or mobile-mcp) and returns the result. Use list_driver_tools to se...

Call an LLM model directly and get the response with metrics (tokens, latency). Uses available API keys: Gemini → OpenAI → Anthropic. Useful for eval-driven workflows, model com...

Run an OpenClaw skill safely through security checks.

Invoke a WebMCP tool on a connected origin. The tool is executed in the browser page context via page.evaluate(). Args are validated for suspicious patterns and results are scan...

Capture Android SurfaceFlinger stats and logcat for jank analysis (Layer 0 only). Returns janky frame counts, missed vsync, slow UI thread metrics, and filtered logcat entries f...

Capture a screenshot of a URL using headless Playwright. Returns the screenshot as an inline image that multimodal agents can see and evaluate directly. Also captures browser co...

Compile entity intelligence into a decision-ready packet.

Generate 3-7 future scenario branches for an entity or decision.

Model explicit tensions between forces for a decision or entity.

Compute block-based SSIM analysis on a set of frame images. Uses 8x8 blocks with parallel ProcessPoolExecutor. Returns SSIM scores, adaptive threshold (max(0.70, median - 2*std)...

Connect to an external MCP server and make its tools available through nodebench-mcp. Predefined drivers:

Connect to an OpenClaw agent with a security policy applied.

Connect to a WebMCP-enabled website via Playwright. Navigates to the URL, intercepts navigator.modelContext tool registrations, and makes discovered tools available for invocati...

Run a comprehensive pre-deployment quality gate. Checks: convex/ directory structure, schema.ts validity, deprecated validator usage, auth configuration, recent audit results, a...

Run a configurable quality gate across all stored audit results. Like SonarQube

Dogfood NodeBench Delta on itself. Verifies runtime health, setup friction, distribution surfaces, and compounding return loops, then emits a repair-ready delta.market packet.

Disconnect from an external MCP driver and shut down its child process. Use this to clean up or to reconnect with different settings.

Disconnect an OpenClaw session and generate a safety summary.

Disconnect from a WebMCP origin and close the browser page. Use this to clean up resources or to reconnect with different settings.

Scan the current page DOM and auto-register components in the dive tree. Discovers semantic landmarks (nav, header, footer, forms, modals, tables, sidebars) and counts interacti...

After fixing a bug, verify the fix by re-navigating to the affected route, comparing before/after state, and updating the bug status + changelog. This is the core flywheel step:...

Define and track a structured interaction test for a component. Provide preconditions and a sequence of test steps (action, target, expected outcome). The agent executes each st...

Re-traverse a route after code changes to detect regressions and verify fixes. Compares the current state against previously registered components, bugs, and test results for th...

End an OpenClaw session and generate a safety summary.

Pre-merge validation combining git state, verification cycles, eval runs, test results, and quality gates. Returns a go/no-go merge decision with detailed check results and bloc...

Record screen and extract key frames from an Android device (Layers 1+2). Uses adb screenrecord then ffmpeg scene-filtered JPEG extraction. CRITICAL: uses -pix_fmt yuvj420p for ...

One-call convenience: gathers all local context and produces a complete

Replay a session

Build a starter GTM script for the current founder wedge.

Run an OpenClaw tool safely through security checks.

Invoke a per-view tool on the current or specified view.

Request a retry, re-plan, escalation, or stop for a failed subtask.

Run the 7-criterion LLM judge on a tool

Dynamically load a toolset into the current session. After loading, the tools become immediately available for use. Based on the

Log and optionally auto-execute an interaction step. If the built-in Playwright browser is active (launched by start_ui_dive), the action is automatically executed in the browse...

Manipulate a screenshot using sharp (image processing). Supports crop (extract a region), resize, and annotate (draw colored rectangles and labels to highlight areas). Use after...

Navigate to a specific view in the NodeBench AI frontend.

Start a new agent conversation with an optional initial message.

Switch the research hub tab.

Send a question to the NodeBench AI agent. Returns a structured response with reasoning and sources.

Navigate to a specific view. Returns the target view

Re-run deterministic public/private boundary and source-evidence verification for a stored public claim.

Run or reuse public role research, store public hiring or market claims, and return a compact role dossier.

Run adaptive, evidence-backed research across one or more subjects. Automatically resolves entities, infers scenario facets, selects relevant research angles, reuses precomputed...

Open the NodeBench UI Dive dashboard in a browser. Shows the full flywheel cycle:

Start the local Daily Brief dashboard server if needed, and return the URL. The dashboard shows Brief metrics, Narrative thread lanes, and Ops status — all from local SQLite.

Build 5-year financial projections based on historical data and industry assumptions. Projects: - Revenue with declining growth rates - Operating expenses (gross margin, SG&A, ...

Run a founder readiness scan against the progression and diligence model.

Render flow visualization with colored bounding boxes for each flow group. Supports overlay on a rendered page image or synthetic dark canvas fallback. Handles wide Figma pages ...

Request a human approval gate for a risky execution-trace action. Approval state is written onto the live run so the UI and ledger can show the pending handoff.

Execute autonomous verification loop with stop conditions. Implements Ralph Wiggum pattern with checkpoints, iteration limits, and timeout. Use for multi-step autonomous tasks t...

Run a longitudinal benchmark batch. N=1 is a smoke test (1 founder, 1 session).

Return a BrowserStack/browser-automation benchmark lane payload.

Track a compile-lint-test-debug closed loop iteration. Record the result of each step. Never present changes without a full green loop.

Static analysis on code or text content for security issues, secrets, homograph attacks, ANSI injections, suspicious URLs, and code quality. Returns structured findings with sev...

Return a competitor-signal-to-response benchmark lane payload.

Run a multi-agent scenario simulation with bounded branching and budget controls. Instantiates agents with personas and incentives, varies conditions across branches, and genera...

Execute the priority 3 dogfood scenarios with automatic LLM judge validation.

Run a full DeepTrace entity intelligence mission with optional bounded research cell. Unifies relationship mapping, ownership, supply chain, signals, and causal analysis. Pass r...

Run full 4-layer Android UI flicker detection pipeline: SurfaceFlinger stats + logcat (L0), screenrecord (L1), frame extraction + SSIM analysis with adaptive threshold (L2), opt...

Run the weekly founder reset autonomy benchmark lane.

Generate a knowledge graph from a folder of code, docs, papers, or images.

Execute a full judge-fix-verify loop: calls a tool, judges the output, and if it fails,

Enforce the mandatory 6-step AI Flywheel verification after any non-trivial change. All 6 steps must pass before work is considered done. Skipping is only allowed for trivial ch...

Compare actual output against a known-good oracle reference. Based on Anthropic\

Return a packet-to-implementation benchmark lane payload.

Evaluate content or code against a set of boolean rules. Returns pass/fail with specific failures listed. The agent evaluates each rule and passes boolean results — the tool per...

Start a reconnaissance research session. Use this at the start of Phase 1 (Context Gathering) to organize research into external sources (SDKs, APIs, blogs) AND internal context...

Run a bounded re-analysis cell for a DeepTrace entity investigation. Queries existing DeepTrace state through parallel branches (evidence gap analysis, counter-hypothesis, dimen...

Run a local-first autonomous delivery loop across exploratory research, planning, implementation commands, dogfood, verification, and judge. Persists one durable run in SQLite a...

Autonomous self-healing for detected drift issues. Fixes orphaned verification cycles

Run autonomous self-maintenance cycle. Checks TypeScript compilation, documentation sync, tool counts, test coverage, and more. Can auto-fix low-risk issues. Scope: quick (1hr c...

Run a live signal sweep across all data sources (HackerNews, GitHub Trending, Yahoo Finance, ProductHunt). Returns signals sorted by relevance with severity tiers (FLASH/PRIORIT...

Open the outbound websocket bridge, pair or resume the local device, and flush pending approved operations to the paired web account.

Execute a shell test command with timeout, capture stdout/stderr, and return structured results. Useful for running test suites, linters, or build commands as part of verificati...

End-to-end visual QA pipeline: burst capture → SSIM stability analysis →

Trigger the full DRANE newsroom pipeline (Scout > Historian > Analyst > Publisher) for a given topic or entity. Returns the generated narrative analysis.

Scan for common operational issues: stale missions, blocked tasks with met deps, old sniff checks.

Execute multiple commands, index all outputs, and run multiple search queries — all in ONE call. This is the highest-efficiency tool: one sandbox_batch replaces N sandbox_execut...

Run a shell command, automatically index the output into the sandbox, and return only a summary. The raw stdout/stderr stays in SQLite — only line counts and a preview enter con...

Start a multi-page spider crawl with extraction. Crawls from start URLs, follows links matching a CSS selector, extracts data per page. Returns a session_id to poll with scrapli...

Stop a running crawl session. Pass the session_id from scrapling_crawl. Items collected so far are preserved. Use when you have enough data or need to abort. Requires Scrapling ...

Self-implement missing agent infrastructure. Generates implementation plan and code templates for: agent_loop, telemetry, evaluation, verification, multi_channel, self_learning,...

Run Monte Carlo simulation for founder decisions. Generates multiple random paths to visualize possible future outcomes. Shows average payoff, success/failure rates, best/worst ...

LLM-powered tool selection: sends your task description + a compact tool catalog to a fast model (Gemini 3 Flash, GPT-5-mini, or Claude Haiku 4.5) to pick the best 5-10 tools. M...

Start a secure OpenClaw session with safety rules applied.

Start an autonomous capability benchmark. Defines a complex build challenge and tracks agent progress through milestones. Inspired by Anthropic\

Start a new dogfood session for one of the 3 canonical loops (weekly_reset, pre_delegation, company_search). Returns sessionId for subsequent recording.

Start a new eval run. Define the test batch upfront with test cases (input, intent, expected behavior), then record results as each case is executed. Rule: no change ships witho...

Start a live Convex-backed execution trace run for a workflow. Creates a task session and trace together so later steps, decisions, evidence, verifications, and approvals all la...

Initialize a UI/UX Full Dive session. Auto-launches a headless Playwright browser if installed (zero setup). Navigates to the app URL and optionally auto-discovers page componen...

Start a new 6-phase verification cycle for a non-trivial implementation. Returns the cycle ID and Phase 1 instructions. Call this before declaring any integration, migration, or...

Switch between research hub tabs

End-to-end Thompson Protocol pipeline orchestrator. Takes a complex topic and runs it through all 4 agents (Writer → Feynman Editor → Visual Mapper → Anti-Elitism Linter) with q...

Transcribe a local audio file (MP3/WAV/etc) to text using faster-whisper via Python. Deterministic, no network.

Run a scheduled task right now instead of waiting.

When an eval run shows regression, trigger a new verification cycle to investigate. This is how the outer loop feeds the inner loop: regressions trigger 6-phase investigations.

Run triple verification on agent implementation. V1: Internal codebase analysis. V2: External authoritative source validation (Anthropic, OpenAI, LangChain, etc.). V3: Synthesis...

Run the agent validation harness — simulates how AI agents (Claude Code,