Jev Social — browser-grounded social research

Turn Instagram, TikTok, and LinkedIn into source-cited reports through your own browser.

Jev chooses each next operation: search, open a particular post or profile, read comments, explicitly requested TikTok media download, or finish. The local socai CLI executes the selected command in your real Chrome. Each result goes back to Jev before the next decision.

⭐ Star Jev Social · OpenAI Decisions API: where Jev Social fits · Watch the recorded run · Live site · Social research guide · Security and data flow · Real 64-second Jev report · Recorded TikTok video evidence · Discord · Jev

Try it

With Node 22+ and Chrome already signed in to Instagram, TikTok, or LinkedIn:

npx github:socai-io/jev-social#v0.1.13 onboard
npx github:socai-io/jev-social#v0.1.13

With the default OpenRouter provider, onboarding prompts for the OpenRouter key. On macOS and Windows, onboarding can also install the official socai CLI when it is missing. On Linux, install a current socai CLI from source first, then put it on PATH or set SOCAI_BIN. Unattended or non-interactive installation requires the explicit --install flag. The second command opens the loopback-only demo.

The recording above shows the earlier routing-only prototype. Current runs include a history of every operation chosen by Jev.

Decision APIs are becoming a category

OpenAI introduced the Decisions API in its DevDay 2026 recap. That makes narrow, machine-consumable decisions a broader API category. Jev Social demonstrates the layer that still has to exist after a decision: rebuild choices from live browser state, execute one bounded operation, and retain the observed evidence.

Jev Social currently uses TypeSafe Jev through OpenRouter. OpenAI Decisions API support is not yet wired; compatibility will only be claimed after a public request and response contract can be tested end to end.

Why this pairing

Jev chooses from a changing list of concrete, read-only operations. The list includes exact targets discovered in previous results, so Jev decides which post to open and which socai CLI operation to run. The local CLI handles the underlying navigation, clicks, scrolling, and extraction; Jev does not choose arbitrary DOM coordinates or generate shell commands.

"find handmade art on Instagram"
        │
        ▼
   Jev chooses the platform
        │
        ▼
   Jev chooses an operation ◄──── observed results
        │                              ▲
        ├─ search                      │
        ├─ open a selected profile     │
        ├─ open a selected post        │
        ├─ read comments / download    │
        │         └──── socai CLI ─────┘
        └─ finish → cards · table · evidence report

https://github.com/user-attachments/assets/4849e0f3-87d5-4a0d-8e0b-2a58e3d0267a

This video also predates the per-operation Jev loop.

Available operations

Only commands exposed by the installed socai CLI are offered. Targets come from captured results or explicit URLs in the user's request. Unsupported, malformed, and low-confidence decisions do not execute. Previously attempted operations are removed from the next choice set.

The run stores each choice, confidence, command, observed result summary, and elapsed time. Login/access gates, decision failures, and step limits produce partial results rather than a success claim. After collection, Jev Social can use the existing OpenRouter key to synthesize a concise report from a bounded, sanitized evidence payload. Every accepted finding must cite a captured source; foreign URLs, missing citations, raw JSON, and local paths fail validation and fall back to the deterministic evidence report. Report sections stream into the fixed panel while captured cards stay visible, and the downloaded report.md is the same Markdown shown in the browser. The synthesizer has no browser or shell tools and does not depend on socai research.

The report model defaults to openai/gpt-4o-mini and may incur normal provider usage. Set OPENROUTER_REPORT_MODEL to another available model, or set it to off to keep report generation fully deterministic. Speed varies with the number of chosen operations, the live site, and the selected report model.

Privacy and local data

• Jev Social does not read or copy the browser cookie store directly. The installed socai CLI uses the Chrome profile you selected; use a separate profile or test account for sensitive research.
• OpenRouter is the default decision provider. Decision requests contain the full research goal plus the requested platform, current action labels, source URLs, earlier action summaries, and short visible-text excerpts. The application does not intentionally add cookies, downloaded media, raw CLI JSON, or filesystem fields, but user input and visible social content are untrusted and can themselves contain personal or path-like text. Optional report synthesis receives a separate bounded, sanitized evidence payload; set OPENROUTER_REPORT_MODEL=off to disable that second model call. A configured Kev or Simple Jev endpoint keeps decision requests on the explicit loopback server and never receives the OpenRouter key.
• Jev Social has no analytics endpoint of its own and sets SOCAI_TELEMETRY=0 for every spawned socai CLI process by default. Set SOCAI_TELEMETRY=1 only to opt in explicitly. In the audited socai v0.6.1 contract, enabled CLI events can include search text, action targets, device context, and available cloud-account fields before the socai.io proxy forwards them to Axiom; the OpenRouter key itself is filtered from the child environment. A separately running socai desktop app has its own process-level setting. Exact fields, narrower controls, version scope, and the undocumented server-side retention period are detailed in the security policy.
• Recoverable checkpoints live under ${JEV_SOCIAL_HOME:-~/.jev-social}; socai stores its own evidence and downloaded media separately. There is no automatic cleanup schedule. Stop the app, inspect the run, and remove only the specific checkpoint or evidence directory you no longer need.
• TikTok media download becomes an available action only when the research goal explicitly requests an offline copy. That intent is retained in the operation history; there is no second confirmation after the explicit request.

The complete field limits, file permissions, provider-retention boundary, browser-session boundary, and recovery procedure are documented in Security and data flow.

Recorded evidence