No-spend health check: auth, API status, credits, free-run pass, defaults, and MCP version.

No-spend, client-aware guided onboarding. Local agents receive the full setup-first versus jump-in flow; hosted clients receive a research-first flow with no filesystem setup unless the user explicitly requests approval-gated workspace-rule guidance. Budget maps to about 15 minutes per $1: $5 is about 75 minutes and $20 is about 300 minutes (5 hours). The $2/$5/$10/$20 tiers are starting points, not caps.

No-spend topic-aware guide for explaining Hound, budgets, completion, setup, reports, datasets, sources, billing, troubleshooting, or general Webhound behavior.

No-spend guidance for removing Webhound MCP config and Webhound-specific local rules from an agent workspace. Does not revoke keys automatically.

Read the saved MCP defaults for budget, product, and free-run use. The MCP always uses Hound.

Update only the provided default budget/product/free-run fields for future MCP runs; omitted fields keep their saved values. The MCP always uses Hound. Do not use this for private workspace-derived rules; save those locally in the agent workspace.

Start a private long-running report with Hound, Webhound's DeepSeek V4 Pro + GPT-5.4 research harness. Budget controls research depth; watch until done=true. Do not force finalization before done=true.

Start a private long-running dataset with Hound, Webhound's DeepSeek V4 Pro + GPT-5.4 research harness. Budget controls extraction depth; watch until done=true. Do not force finalization before done=true.

Authoritative session watcher. done=true means the run is terminal. output_ready=true without done=true can still be intermediate; keep waiting unless the user explicitly asks for a partial update.

Bounded wait wrapper around webhound_watch. Max 110 seconds, then returns still_running if not terminal. still_running is normal for budgeted research; call wait/watch again unless status is awaiting_input or a blocking alert is present. Do not finalize or stop a healthy running session.

Save concrete source-backed notes or hypotheses found by the calling agent while Webhound keeps running. No spend. Does not interrupt the Planner/Executor/Verifier cycle and does not change session status. Do not use for user intent changes.

Read shared sidecar notes for a session. No spend. Use this to see what the calling agent has already shared with Webhound before adding, correcting, or dismissing notes.

Edit, restore, or dismiss one shared sidecar note. No spend. Does not interrupt the Planner/Executor/Verifier cycle and does not change session status.

Send user-provided guidance to a session. Use reason="awaiting_input" to answer a checkpoint and resume. Use reason="user_guidance" only for a real user objective/scope/constraint/deliverable change. Use webhound_add_sidecar_notes for source suggestions.

Pause/stop a running Webhound report or dataset without deleting it. Use only when the user explicitly asks to stop, pause, or cancel. Do not use for healthy long-running sessions, normal budget use, warning-level tool errors, or because partial notes look sufficient.

Resume a paused/completed/awaiting-input session with optional additional budget and guidance.

Add research budget and optional guidance/context to a session.

Lower a running or paused report budget only after the user explicitly asks to reduce the remaining research scope or finish with the research already gathered. This changes the report stopping boundary; Webhound then performs normal final assembly. Never use this because partial notes look sufficient, the run is taking time, or the agent wants an earlier result.

Read final report/working document or dataset rows. By default this is for terminal sessions only; do not read or summarize partial working notes while a healthy run is still running.

Export a completed report or dataset as Markdown, HTML, TXT, JSON traces, CSV, JSONL, or PDF. Does not spend credits. Wait for done=true before exporting unless the user explicitly asks for a partial artifact.

Read the full evidence payload for a completed Webhound session: final output, working docs, claim traces, sources, and export links. Use this before serious follow-up answers so Webhound value is not reduced to only the polished output document.

Make a report or dataset accessible to anyone with the link and return the share URL. This is share-only: reports use /document/:id, datasets use /dataset/:id, and it does not publish to Explore or create a /p/:slug publication.

Read normalized claim traces and provenance for a session.

Read source inventory and citation counts for a session.

Semantic search across prior Webhound sessions.

List recent Webhound sessions.

Read the complete canonical session in one uncapped response: prompts, messages, phases, tasks, agents, final and working documents, dataset rows, claims, sources, notes, diagnostics, usage history, and artifact links. Nothing is paginated, truncated, or omitted.

Upload a CSV, XLSX, PDF, DOCX, TXT, Markdown, or VTT ChatGPT attachment, local file, text, or base64 content for use in a report or dataset. Convert legacy XLS/DOC files to XLSX/DOCX first.

Read credits, recent usage, free-run status, and defaults. Does not spend.

Explain whether a session is healthy, done, usable, and what to do next. For a healthy running session, the correct next action is to keep waiting; budget use is the point of the run. Do not stop or force finalization because the run is slow.