Use Desearch MCP in Crush v0.101.0
Tested 17 November 2026.
For developers with a working Crush installation and a tool-capable model who want to search the web through Desearch. This guide adds the remote MCP connection and verifies a real search using Crush's native CLI and saved session output. Complete Crush's official installation and model setup first.
Requirements
Tested with Crush v0.101.0, commit 0dcd3d5fc452343d400b9de876d2dc10883f9519. The official Linux ARM64 release archive had SHA256 ebbc141e80ad19d321b89558c2513fac5bebabd97530570ce749902dc4411f05. Crush ran in a disposable Linux ARM64 container on Docker Engine 29.5.2. Its configured model used an existing host-native llama.cpp b12478 server, Qwen3 0.6B GGUF weights SHA256 7f4030143c1c477224c5434f8272c662a8b042079a0a584f0a27a1684fe2e1fa, 32,768 tokens of context, thinking disabled and the completions provider type. No hosted model account was used. Other models and operating systems were not tested.
You need a configured model that can call tools, a Desearch API key with available credits, and outbound HTTPS access to https://mcp.desearch.ai/mcp. The local open-source Crush path was tested; no subscription requirement or hosted plan is asserted. Searches consume Desearch account credits; consult current Desearch pricing. The hosted MCP server version was not pinned, so this is a dated behavior test.
Configure the connection
Open a fresh scratch project using your working model configuration. Supply DESEARCH_API_KEY to the Crush process through your secret manager or shell environment. If Crush runs in a container, pass the variable into that container as well. Keep its value out of the project configuration and prompts.
Merge this JSON into that project's crush.json, preserving the existing provider and model settings. This version still supports crush.json, although upstream marks the JSON format deprecated in favor of crushrc. If you already use crushrc, do not assume a second JSON file overrides it; use the tested JSON configuration in a separate scratch project for this check.
The tool filter exposes web-search for this check. The permission allows Crush to execute that specific search without an interactive approval prompt. A search can consume credits. Review the configuration before running it; Crush supports shell expansion in configuration values.
Validate with the native CLI
Run this from the same fresh project and environment:
Inspect the last command's JSON. In messages[].parts, confirm a tool_call named mcp_desearch_web-search with the requested query, followed by a matching tool_result with no is_error flag set to true (false is omitted in this version). Parse that result's content as JSON and confirm a nonempty data array containing title and link. The test returned Our Documentation | Python.org through the native tool result. Result titles can change.
A model-written answer alone does not verify the connection. In the test, the small model returned a generic Python.org sentence even though the actual tool result contained titled search data. Use the saved native tool result to validate. A link beginning /go/v1/ is a returned Desearch redirect path; preserve its value rather than inventing a replacement URL.
Troubleshooting
If DESEARCH_API_KEY is absent, the shell check above stops before Crush starts. This checks local configuration only; it does not authenticate the upstream key. An intentionally invalid key returned Web Search error: HTTP 403: Forbidden inside the native tool content. Crush still exited zero and its is_error flag was false. This negative check returned no titled data and was not counted as a successful search. If the actual tool reports authentication or credit failure, check the key received by Crush and the account balance before retrying. A discovered tool or a zero CLI exit code does not prove a successful search.
If the model never calls mcp_desearch_web-search, inspect the actual session JSON and verify tool support, model context capacity, the loaded project configuration and permissions. The test disabled unrelated built-in tools and used an isolated configuration/data directory. Earlier runs in a constrained model runtime produced no titled native result; they were not counted as successful. Use an adequately sized model runtime and verify actual tool output for your model.
Run session last immediately after the check in the same project and with the same CRUSH_DATA_DIR, if you set one. Its JSON is a version-specific diagnostic interface; review it locally before sharing because sessions can contain project or prompt data. This guide covers the remote MCP connection, not installation of a model server or configuration of every model provider.