11 JC Lab Ideas for September 10: Local Agent Frameworks, Permission Audits and Two Crypto Attention Checks (2026)

OtoDock, Geiger, and Ctrlb Decompose give you three immediate on-screen demonstrations with measurable outputs. Frigade Assist API and OpenObserve provide two clear workflow evaluations for creators tracking production agent reliability, while FreeScan and Modeinspect turn marketing claims into structured tests. NEAR Protocol and Kaspa offer two disciplined paper-trading scenarios anchored in live market liquidity, and OpenAI's internal research acceleration data supplies an evidence-backed article angle on autonomous agent workloads.

Video topics to film

OtoDock: run persistent Claude Code and Codex agents on your own server

Format: Tutorial

Film today: Spin up a clean Ubuntu 24.04 test instance with Docker. Run curl -fsSLO https://raw.githubusercontent.com/OtoDock/oto-dock/main/scripts/install.sh && bash install.sh. Open http://localhost:8400, complete the initial setup wizard, and link a disposable subscription key in the dashboard. Assign the agent a simple data transformation script inside an isolated workspace. Show the bubblewrap kernel sandbox and network isolation toggles on screen before concluding the walkthrough.

OtoDock version 1.6.0 ships as a self-hosted platform designed to coordinate persistent Claude Code, Codex CLI, and local open-source models across departmental workspaces. The project documentation specifies that agents operate inside isolated kernel sandboxes with network controls, encrypted credential injection, and multi-tenant access for up to five users under its fair-source license. The repository documentation details support for Docker deployments on Linux servers, outbound-only WebSocket links for remote worker nodes, and direct document editing for spreadsheets and presentations. 1

Hook: Can a self-hosted agent platform protect local credentials while giving persistent coding models access to departmental tasks?

Boundary: Keep initial testing inside a disposable virtual machine. The official installer targets fresh Linux environments and halts when existing configurations are present. Telephony features require external provider accounts, while complete model execution depends on active provider subscriptions.

Geiger: audit local AI agent permissions and MCP endpoints in one scan

Format: How-I-tested

Film today: Open a terminal on your development workstation and execute npx geiger-scan. Record the terminal output as the scanner inspects local configurations across Claude Desktop, Cursor, and registered command-line tools. Generate an HTML artifact using npx geiger-scan --html audit.html and open the generated file in a browser. Highlight any unverified MCP servers or elevated filesystem permissions detected during the pass.

The Geiger repository documents a zero-dependency, read-only JavaScript scanner that catalogs local agents, Model Context Protocol servers, IDE extensions, and browser plugins. The tool evaluates configuration files across major developer environments to flag specific exposure capabilities, including code execution rights, credential storage locations, broad filesystem access, and outbound network permissions. The documentation specifies that all scans run locally without network transmission or secret exfiltration, outputting structured JSON and standalone HTML reports. 2

Hook: Which local coding assistants and MCP connectors quietly hold system execution privileges on your main machine?

Boundary: The utility reads known static configuration paths rather than active runtime memory. Unconventional binary locations, containerized runtimes, and nested virtual machines remain outside the scan radius. The resulting report reflects configuration posture rather than malicious intent.

Ctrlb Decompose: compress noisy production logs into typed LLM prompt structures

Format: Review

Film today: Download the release binary or open the browser-based WebAssembly demo. Feed a 50,000-line sample server log into ctrlb-decompose --llm app.log. Display the token savings on screen by comparing the raw log line count against the compressed template output. Pass the reduced pattern summary to a local language model to identify the root cause of a simulated error cascade.

The Ctrlb Decompose repository describes an open-source Rust tool designed to reduce large log files into compact semantic templates. The engine applies CLP byte encoding, Drain3 clustering, and statistical typing across timestamps, IP addresses, durations, and categorical variables. The project documentation reports that the compression pipeline operates in a single streaming pass without external network calls, producing human-readable tables, structured JSON, or token-efficient Markdown representations suitable for LLM reasoning windows. 3

Verdict frame: Compare raw log token consumption against the clustered pattern view across ingestion speed, variable extraction accuracy, and model diagnostic speed.

Boundary: The tool groups messages by syntactic similarity rather than semantic meaning. Rare one-off errors can merge into generic wildcard buckets if similarity thresholds are set too broadly. Keep raw logs archived for compliance verification.

Frigade Assist API: compare static documentation lookups with live UI guidance

Format: Comparison

Film today: Register a test project and install @frigade/ai. Configure a basic support tool call using frigade.assist({ query }) inside a sample web application. Submit two distinct queries: an informational question regarding pricing policies, followed by an actionable task requesting a workspace setting change. Capture the difference between the textual explanation and the interactive step-by-step UI highlight.

Frigade's official documentation describes the Assist API as a developer toolkit that equips autonomous agents with real-time application navigation capabilities. Rather than relying on static knowledge bases, the platform crawls product workflows to build an operational state map, updating its route model across releases. The documentation details support for Vercel AI SDK integration, SOC 2 Type II compliance standards, and automated user handoffs whenever a requested operation exceeds configured agent autonomy thresholds. 4

Hook: Can an autonomous agent guide a user through a live web interface without breaking the underlying application flow?

Boundary: Production deployment requires an active commercial subscription starting at entry enterprise tiers. The underlying route crawler requires staging access to map multi-step flows accurately before production handoff.

Tools and apps to review

AI Observability by OpenObserve: trace agent tool loops and token expenditure

Format: Review

Review today: Deploy the OpenObserve open-source container image or log into the cloud sandbox. Instrument a simple multi-step agent workflow that calls an external search tool and a database query. Trigger a simulated retry loop, then navigate the waterfall trace view to isolate where latency and token costs accumulated.

OpenObserve's official product announcements describe an OpenTelemetry-native observability suite built to monitor large language models and multi-agent sessions. The platform records prompts, completions, latency metrics, and computed token expenses across each span, linking model invocations directly to underlying database queries and infrastructure metrics. The product documentation outlines built-in evaluation scoring, OpenTelemetry Protocol compatibility, and self-hosted storage options designed to maintain log retention without external cloud egress. 5 6

Verdict frame: Evaluate the platform across span visualization clarity, cost-per-call calculation accuracy, and setup friction compared to standard application performance monitoring tools.

Boundary: Setting up custom evaluators requires dedicated model credentials. The open-source edition requires administrative maintenance of local storage volumes and ingestion pipelines.

FreeScan.app: examine automated public website audits across visibility and security

Format: Scam check

Review today: Submit a live public web URL into the scanner input without logging in. Capture the execution of the audit across SEO, accessibility, public headers, and design structure. Cross-check the reported security findings against manual browser developer tools to verify whether the flagged issues represent practical vulnerabilities or automated checklist warnings.

FreeScan's official portal presents a free public web scanner that evaluates single pages across search engine optimization, answer-engine readiness, accessibility standards, and visible security headers. The platform reports 38 individual checks divided into four performance categories, generating prioritized remediation recommendations alongside prompt snippets for AI coding tools. The site outlines an optional subscription tier for automated recurring monitoring across multiple web properties. 7

Verdict frame: Inspect the boundary between legitimate technical debt and inflated audit scores. Check whether the free recommendations provide actionable technical remedies before considering paid monitoring plans.

Boundary: The public audit evaluates static client-facing markup and server response headers. Deep application logic, authenticated database layers, and private backend configurations remain invisible to the scan.

Modeinspect: test real UI state manipulation directly inside a local codebase

Format: How-I-tested

Review today: Connect a test React or Next.js repository to the Modeinspect canvas. Open an existing dashboard screen and use the visual controls to adjust component spacing, color tokens, and responsive layout breakpoints. Review the resulting Git diff in your terminal to confirm whether the platform produced clean, type-safe pull requests using your existing component definitions.

Modeinspect's official documentation describes a code-connected design platform that renders user interfaces directly from the underlying project repository. The application synchronizes design tokens, component variants, and dynamic application states onto an interactive canvas, allowing designers and engineers to edit interfaces on live data. The platform documentation states that visual modifications translate directly into type-checked codebase changes, reducing manual translation between design files and production code. 8

Verdict frame: Score the experience on component library synchronization fidelity, style token enforcement, and the cleanliness of generated pull requests.

Boundary: The visual canvas requires a structured, modern component architecture. Non-standard CSS configurations or legacy monolithic styling setups may require initial configuration adjustments.

Athenic AI: evaluate automated pre-market equity screening and catalyst extraction

Format: Comparison

Review today: Enter a plain-language stock screening query specifying volume surge thresholds, market capitalization boundaries, and recent regulatory filing triggers. Compare the generated candidate list against a traditional financial screener. Inspect the synthesized catalyst notes alongside official regulatory filing links to verify factual alignment.

Athenic AI's official product portal outlines an automated equity research platform that combines institutional market data with large language model analysis. The system scans market feeds, financial statements, and regulatory disclosures to highlight trading catalysts and unusual volume movements prior to market open. The company documentation highlights custom thesis screening, pre-market alert generation, and automated daily briefing deliveries. 9

Verdict frame: Place Athenic's natural-language results alongside conventional quantitative screeners to evaluate filter accuracy, data freshness, and catalyst relevance.

Boundary: Financial market screening involves capital risk. Treat all automated catalyst summaries as preliminary research leads, and verify every filing directly with primary sources before making trading decisions.

NEAR Protocol / NEAR: inspect trading volume resilience during search momentum spikes

Format: Trend react

Today's signal: CoinGecko's trending search tracking placed NEAR Protocol at rank 4, reflecting elevated user discovery interest across the platform. Spot market queries recorded a reference snapshot price of $2.43 at the time of verification. CoinGecko documentation defines the trending board as an index of active user search volume rather than trade execution quality. 10 11

Film today: Record the search position on CoinGecko, then pull up real-time market depth and 24-hour exchange volume on a major spot venue. Build a disciplined paper-trading card detailing the spread, resistance levels, and a planned risk-reward entry with an explicit stop threshold.

Boundary: High discovery ranking often coincides with rapid market volatility. The $2.43 price marker represents a fixed snapshot from 10 September 2026. Use paper-trading simulations to study market execution without committing capital.

Kaspa / KAS: measure order-book depth against Proof-of-Work attention signals

Format: Trend react

Today's signal: CoinGecko's trending board placed Kaspa at rank 6, indicating sustained search volume among proof-of-work digital assets. Spot market data recorded a reference snapshot price of $0.03828 at the time of verification. The trending registry tracks search inquiry velocity across the ecosystem. 10 12

Film today: Display the trending ranking, then inspect the 24-hour bid-ask spread across available secondary markets. Construct a paper-trading order with a predefined slippage tolerance and document the execution friction across different order sizes.

Boundary: The $0.03828 price quote reflects a verifiable reference snapshot from 10 September 2026. Secondary market liquidity varies significantly between exchanges; limit all demonstrations to paper calculations.

Article

The reality of the automated research intern: what 3.1 agent-workdays reveal

Format: Article

Write today: Open with the concrete benchmark disclosed by OpenAI on 06 September 2026: research staff now consume 3.1 agent-workdays of automated compute for every standard human workday. Detail the progression from early 2026, when median researchers barely touched coding models, to mid-August, when the median researcher directed over $600 of daily API inference and power users exceeded $7,000 per day. Examine the shift in task distribution across infrastructure debugging, automated test runs, and experimental coordination. Highlight the central operational bottleneck: over half of successful four-to-eight-hour agent runs still required direct human intervention.

OpenAI's published study, titled Research acceleration: The view inside OpenAI, reports that automated coding agents have substantially altered internal research workflows throughout 2026. The paper documents that August 2026 reached an all-time high in experiments conducted per active researcher, driven by concurrent agent deployments and automated tool troubleshooting. The report notes that routine infrastructure assistance inquiries declined as agents resolved environment failures, while high-level strategic planning remained a minor fraction of overall token generation. 13

Angle: Frame the analysis around the practical transition from writing code to orchestrating and auditing autonomous workers. Explore what happens to engineering teams when compute cost replaces typing speed as the primary development constraint.

Boundary: Keep the analysis anchored to the empirical metrics and constraints disclosed in OpenAI's paper. The reported figures describe a frontier research organization with custom infrastructure; individual software teams face distinct tooling and budget constraints.

Filming order

Begin with Geiger because the terminal scan runs in seconds and produces immediate visual findings. Move to Ctrlb Decompose to demonstrate tangible log compression before configuring OtoDock inside Docker. Record Frigade Assist API once application routes are established. Transition to the review tools by testing OpenObserve traces, auditing FreeScan checks, and validating Modeinspect codebase synchronization. Conclude on camera with the NEAR Protocol and Kaspa paper-trading cards, then draft the research acceleration article using OpenAI's verified workload metrics.