Webhound lets a Dify agent or workflow decide how much work a question deserves. A prompt defines what to research, and an explicit dollar budget defines the research effort. The result includes both the finished artifact and the evidence behind it.
Hound is Webhound's research harness built with DeepSeek V4 Pro and GPT-5.4 across planning, execution, verification, and assembly. It is not a selectable foundation model or mode. This plugin deliberately exposes no alternate model or mode controls.
The plugin is free to install. Research runs use the Webhound account and credits belonging to the Dify workspace that configured the plugin.
- Create or sign in to your own Webhound account at webhound.ai.
- Create an API key at webhound.ai/api.
- In Dify, open Tools → Webhound → Authorize.
- Paste that workspace's key into the encrypted
WEBHOUND_KEYfield.
Every Dify workspace must provide its own key. Do not reuse a publisher, developer, or another customer's credential.
- Call
webhound_start_reportfor a cited narrative orwebhound_start_datasetfor sourced rows. Both tools require an explicit budget andconfirm_spend=true. A budget is a maximum, not a promise that the entire amount will be charged. If an included first run is available, Webhound can use it first. - Call
webhound_watch. A run is terminal only whendone=true.output_ready=truecan describe an intermediate artifact. A watch can wait for up to 60 seconds, but should usually follow the returned recommended check-in interval instead of polling continuously. - When
done=true, callwebhound_get_evidence_packfor serious follow-up work. It returns the final output together with working documents, claim traces, sources, and export links. Usewebhound_get_outputwhen only the polished report or dataset rows are needed.
The current rule of thumb is that $1 buys about 15 minutes of research. A larger budget should buy more evidence, coverage, verification, and decision value—not merely a longer answer.
| Tool | Spend | Purpose |
|---|---|---|
webhound_start_report |
Yes | Start a private cited report with an explicit budget |
webhound_start_dataset |
Yes | Start a private sourced dataset with an explicit budget |
webhound_watch |
No | Read authoritative status or perform one bounded wait |
webhound_get_output |
No | Retrieve a completed report, working document, or dataset |
webhound_get_evidence_pack |
No | Retrieve the result and the evidence behind it |
webhound_help |
No | Explain budgets, Hound, completion, setup, and troubleshooting |
webhound_account |
No | Read credits, usage, included-run status, and defaults |
Only the two start tools can begin spend. The adapter has no tool for adding budget, resuming with more budget, stopping a run, or making an artifact public.
The adapter exposes static native Dify tools and forwards them to Webhound's hosted Streamable HTTP MCP endpoint:
https://api.webhound.ai/api/v2/mcp
The endpoint is fixed in code so a credential cannot be redirected to a user-supplied host. HTTP redirects are rejected. Both JSON and Server-Sent Events responses are supported.
Calls are never retried automatically. That matters for spend-bearing starts: if a connection fails after submission, inspect the Webhound session list before retrying so the same investigation is not started twice.
Use Python 3.12:
python3.12 -m venv .venv
.venv/bin/pip install -r requirements.txt pytest ruff pyyaml
.venv/bin/ruff check .
.venv/bin/pytest
dify plugin package .Authenticated remote debugging requires a Dify debug key and a
workspace-owned WEBHOUND_KEY; neither is committed to this project.
See PRIVACY.md and the Webhound Privacy Policy.
- Support: team@webhound.ai
- Plugin source and issues: https://github.com/WebhoundAI/dify-plugin-webhound
- Webhound MCP source: https://github.com/WebhoundAI/webhound-mcp
- Webhound MCP setup: https://www.webhound.ai/mcp-setup