Jev Social
Research Instagram, TikTok, and LinkedIn through Jev-directed operations in a real browser.
Per-dimension scores and reasoning
The tests show that execution is constrained to known CLI commands, supported platforms, observed targets, and strict URL validation; malicious domains, credential-bearing URLs, file URLs, invented actions, low-confidence choices, and cancelled requests fail closed. The README explains the main flow among Jev, OpenRouter, socai, real Chrome, captured results, and reports, while commands and source links are retained; tests also check that private local paths do not leak into events. Deductions apply because Jev autonomously chooses each step and there is no explicit per-operation confirmation, including before a media download. The material does not fully specify which social content is sent to OpenRouter, retention periods, browser-cookie isolation, or key storage. Dependency installation follows latest releases and a remote installation script, workflow actions are pinned only to major tags, and no lockfile, audit, or supply-chain verification is shown. Effects are mostly read-only and the server is loopback-only, but media downloads write locally; no deletion, cleanup, or rollback procedure is documented. Attribution is strong because reports use captured text, comments, source links, and retained evidence commands.
The README, package metadata, and tests consistently describe the three platforms, typed routing, operation loop, step limits, and partial-result behavior. Tests cover access gates, step exhaustion, cancellation, low confidence, invalid actions, and route mismatches, with explicit error codes or stop reasons, supporting high self-consistency and failure-message scores. Dependency availability is reduced because operation requires Node 20, OpenRouter/Jev, real Chrome, and the external socai CLI. Installation and missing-CLI onboarding are documented, but socai is described as current or installed from latest rather than compatibility-pinned, and live platform changes remain an availability dependency.
The README provides concrete creator, trend, people, and comment-research scenarios and supports CLI or HTTP use, automatic or explicit platform choice, result limits, and step limits. Capability boundaries are explicit: only listed Instagram, TikTok, and LinkedIn operations are offered, while unknown platforms, malformed responses, and unsupported requests are rejected. Tests thoroughly cover typed choices, explicit-platform enforcement, and filtering to observed targets, earning strong trigger precision. Deductions apply because audience roles and deployment scenarios are only lightly described, while real Chrome, login state, external services, and a particular CLI constrain environment fit; container, headless, proxy, and enterprise deployment support are not addressed.
The README is clearly organized around architecture, operations, requirements, installation, parameters, and examples; package, command, and platform naming are coherent. Both no-clone and source-checkout installation paths are documented, and MIT metadata matches the complete license. Deductions apply because there is no FAQ and examples focus mainly on successful searches. Login gates, step limits, variable speed, and partial results are acknowledged, but privacy, platform terms, rate limits, and compatibility limitations are not collected systematically. Version 0.1.1 is declared without a changelog or migration history. SECURITY.md, issue tracking, and private vulnerability reporting provide maintenance channels, but support is stated only for latest and main, the license names generic contributors, and no specific maintainer or release-governance responsibility is identified.
Outputs include operation history, streamed events, cards, tables, reports, comments, source links, and evidence commands. Tests demonstrate detail merging, retained comments, selection of a specific video, and honest partial results, giving strong static support for output usability. Combining typed decisions with an iterative real-browser evidence loop offers clear marginal value over a single search or unconstrained generation. Cost-benefit is reduced because the project only notes that speed varies with steps and live sites; it requires an OpenRouter key and external browser runtime but provides no API pricing, typical latency, request-volume, or resource-cost estimates.
Each operation records its choice, confidence, command, observation summary, and elapsed time, and reports retain source links. Tests show report content coming from captured results and exclude unselected or malicious sources, providing strong claim traceability. A deduction applies because there is no explicit mechanism requiring independent sources to corroborate the same fact; retaining source links alone is not cross-source corroboration. Captured evidence, operation history, and compiled reports are separated, and gated or step-limited runs are not labeled successful, but the supplied material does not show systematic labeling of inference, uncertainty, or fact-versus-synthesis boundaries inside reports.
- Jev autonomously selects and executes subsequent read operations and may choose to download TikTok media; the material shows no explicit user confirmation before that download.
- Operation depends on OpenRouter, real Chrome, social-platform login state, and the external socai CLI; latest-based installation introduces compatibility and supply-chain change risk.
- The full data sent to OpenRouter, retention and cleanup of captured data or downloaded media, and browser-cookie isolation are not specified.
- Reports are source-traceable, but the repository does not demonstrate independent cross-source corroboration of important facts.
- This assessment is based only on the supplied static files; the application, tests, and dependencies were not executed or independently audited.
What does this agent do, and when should you use it?
Jev Social is a locally run social-media research agent for Instagram, TikTok, and LinkedIn. Jev selects its next action from a changing list of concrete, read-only operations, while the socai CLI performs navigation, clicks, scrolling, and extraction in a real Chrome browser. Each observed result is returned to Jev for the next decision, without letting the model invent shell commands or arbitrary DOM coordinates. Runs produce streamed post cards, tables, source-linked evidence reports, and a history containing each choice, confidence, command, result summary, and elapsed time. The application runs on Node.js 20+ at the loopback-only address 127.0.0.1:8766 and requires an OpenRouter key with Jev access, a current socai CLI, and Chrome.
After receiving a research goal, Jev chooses a platform and then selects an operation actually exposed by the installed socai CLI. On Instagram it can search, open profiles and their post cards, inspect a selected post or Reel and read comments, and inspect page state. On TikTok it can search, open an author, read a selected video and its comments, optionally download that video's media, and inspect page state. On LinkedIn it can search people, content, or companies; read a selected profile, company, or post; inspect experience or education; and inspect page state. Targets must come from captured results or URLs explicitly supplied by the user. Unsupported, malformed, or low-confidence decisions are not executed, attempted operations are removed from later choices, and socai's result is returned to Jev until it selects finish or reaches the step limit. The final output includes cards, tables, captured text and comments, and source-linked evidence reports; access gates, decision failures, and exhausted step limits yield partial results instead of a false success.
- A social-media researcher wants to discover creators on Instagram, open selected profiles or posts, read comments, and retain links supporting the findings.
- A trend analyst needs to search a TikTok topic, inspect selected videos and comments, and optionally capture the media for a chosen video.
- A recruiting or market-research team needs to search LinkedIn people, companies, or content and inspect selected profiles, experience, education, or posts.
- An OSINT analyst wants evidence gathered through a real browser session while preserving an audit trail of operations, confidence values, commands, and observations.
- A developer wants to run bounded social-platform research from a local web interface, HTTP search endpoint, or CLI with explicit result and step limits.
What are this agent's strengths and limitations?
- Jev chooses only from concrete, read-only operations exposed by the installed socai CLI, using captured targets or explicit URLs instead of generating arbitrary shell commands or DOM coordinates.
- Execution happens in a real Chrome browser and covers distinct research operations across Instagram, TikTok, and LinkedIn.
- Every run records choices, confidence, commands, observed-result summaries, and elapsed time, then compiles cards, tables, and source-linked evidence.
- Access gates, low-confidence decisions, decision failures, and step exhaustion are handled conservatively, with partial results rather than unsupported success claims.
- It supports a local web interface, HTTP search endpoints, npm-based CLI searches, and direct socai CLI usage when autonomous routing is unnecessary.
- The core workflow depends on an OpenRouter key with Jev access, Jev's decision model, and the socai runtime, so it is not provider-agnostic.
- Adopters must supply Node.js 20+, a current socai CLI, and Chrome; runtime speed varies with the number of selected operations and the state of the live sites.
- Login and access gates on the supported social platforms can stop a run and leave only partial findings.
- Available actions are restricted to commands exposed by the installed socai CLI; the agent cannot freely perform undocumented browser actions or cover unlisted platforms.
- The decision loop is capped at 30 steps, and both operations and the finish decision consume steps, which may constrain complex investigations.
How do you install or deploy this agent?
Prerequisites are Node.js 20+, an OpenRouter API key with Jev access, a current socai CLI, and Chrome. The fastest path does not require cloning the repository:
npx --yes github:socai-io/jev-social onboard
npx --yes github:socai-io/jev-socialThe onboarding command prompts for the key and offers to install the official socai CLI if it is missing.
For a source checkout:
curl -fsSL https://github.com/socai-io/socai/releases/latest/download/install.sh | sh
git clone https://github.com/socai-io/jev-social.git
cd jev-social
npm install
cp .env.example .envSet OPENROUTER_API_KEY in .env, then start the application:
npm startIt opens at http://127.0.0.1:8766 and listens on loopback only.
How do you use this agent?
In the local interface, leave the platform set to “Jev · auto,” enter a research goal, and watch the operation history. Equivalent CLI examples are:
npm start -- search "find emerging design creators on Instagram" --platform auto --limit 4
npm start -- search "find AI wearable trends on TikTok" --platform auto --limit 4
npm start -- search "find AI product managers in San Francisco on LinkedIn" --platform auto --limit 4
npm start -- search "find handmade art on Instagram and read the comments" --limit 4 --max-steps 12--limit sets the target result count and per-search or per-profile collection size from 1 to 100. All captured records, including intermediate profile cards, are retained. --max-steps bounds the decision loop from 1 to 30 and defaults to 12; every selected operation or finish decision consumes a step. The HTTP search endpoints also accept maxSteps. For predetermined operations, socai can be called directly:
socai instagram search "AI wearables" --num 10 --pretty
socai tiktok search "AI wearables" --num 10 --pretty
socai tiktok get-videos --video <url> --download-media --pretty
socai linkedin search "AI agents" --type content --num 10 --prettyHow does this agent compare with similar options?
Compared with calling socai directly, Jev Social adds a per-observation decision loop in which Jev selects the next operation from changing candidates and compiles cards, tables, and an evidence report. Direct socai commands are the simpler option when the platform and exact operation are already known. Jev Social explicitly does not hand browsing to another research agent or depend on socai research.