Webhound research

Webhound research

Use Webhound when the cost of missing information is greater than the cost of doing more research. Use a quicker tool for simple factual lookups.

Research contract

• The prompt says what to investigate. The dollar budget says how much effort the investigation deserves.
• Hound is Webhound's research harness built with DeepSeek V4 Pro and GPT-5.4 across planning, execution, verification, and assembly.
• Hound is not a selectable model or mode, and it is not a direct pass-through to one provider.
• A larger budget should buy more searching, reading, comparison, verification, and evidence—not just a longer answer.
• Every user authorizes Webhound with their own account and API key. Never embed, share, or reuse a publisher's, teammate's, or other user's credential.

Before starting

1. If the hosted Webhound server needs authentication, ask the user to complete Webhound sign-in and consent in the client's OAuth flow. Each person connects their own Webhound account. Local stdio installs instead use that user's own generated Webhound API key.
2. Confirm the research question, desired artifact, and dollar budget. Starting a report, dataset, or budget extension can spend money, so do not invent or silently increase a budget.
3. Use a report for a cited argument, comparison, map, or recommendation. Use a dataset for rows, fields, and per-field provenance.
4. If the user wants guidance, recommend $2 for quick scouting, $5 for normal cited research, $10 for deep work, or $20 for exhaustive/highest-stakes work. These are starting points, not caps or permission to spend; the user can choose a larger custom budget for longer, deeper research. Roughly, $1 buys about 15 minutes, so $5 is about 75 minutes and $20 is about 300 minutes (five hours). Actual runtime still varies.
5. A new account may have one indivisible free run for one exact $5 report or dataset. Check account or onboarding status rather than promising it. Honor the saved preference: use it automatically only when enabled, or after the user explicitly consents to use it for this exact $5 run.

When the user asks to onboard, call webhound_onboarding once, send its immediate_next_message exactly once, and treat agent_playbook.conversation_flow as the canonical sequence. The entry whose say matches the message already sent is consumed; after the user replies, continue with the next unconsumed entry. Treat setup_flow as reference-only and next_action as the entry instruction, not additional sequences. Do not repeatedly call onboarding to advance it. Continue the first run through done=true, then return its output with sources and provenance. Drop onboarding immediately if the user changes the subject. Use webhound_account or webhound_help when account state or product behavior is unclear.

Run the investigation

1. Start with webhound_start_report or webhound_start_dataset.
2. Let Webhound work in the background. Do useful independent work instead of repeatedly polling.
3. Check progress with webhound_watch or webhound_wait. Follow runtime_estimate.recommended_next_check_seconds when the client can schedule a later check.
4. If you find a concrete source-backed lead while Webhound works, save it with webhound_add_sidecar_notes. Sidecar notes are leads, not a reason to interrupt the current planning, execution, or verification cycle.
5. Use webhound_send_message with reason="awaiting_input" when the run asks for information. Use reason="user_guidance" only when the user changes the objective, scope, constraints, or deliverable.
6. Do not lower the budget or ask Webhound to wrap up because partial notes look usable. Change the stopping boundary only when the user explicitly asks.
7. If a call returns billing_required or credit_exhausted, send the user to https://www.webhound.ai/billing, wait for them to act, then confirm account state before retrying the spend-bearing action.

Finish honestly

• Treat done=true as the authoritative completion signal.
• output_ready=true can describe an intermediate artifact. Do not present the investigation as finished until done=true, unless the user explicitly asks for a partial update.
• Once complete, call webhound_get_session for the canonical session and webhound_get_evidence_pack when the answer depends on the research trail.
• Inspect the final output together with working documents, claims, sources, traces, limitations, and gaps. The report is an entry point into the work, not a substitute for the work.
• Use webhound_get_output for the polished result and webhound_export_session when the user needs a file.
• Create a public share link only when the user explicitly asks. Sharing changes the report or dataset from private to accessible by link.

When presenting the result, separate supported findings, uncertainty, limitations, and useful next investigations. Never describe a run as resolving to one backend model.