Use Desearch MCP in VS Code 1.145.1

    Use Desearch MCP in VS Code 1.145.1

    Tested 17 November 2026. This guide configures a remote Desearch connection in VS Code's native Local chat harness and verifies a real tool response. It covers a temporary local web workbench, not a gallery submission.

    Requirements

    Use VS Code 1.145.1, commit 6bab64e2d12b04d04c0d99c27ff166023efd1dda, with its command-line code available and bundled GitHub Copilot Chat 0.73.0. The test used the official macOS ARM64 installation and matching serve-web server. Other operating systems and versions were not tested. Current documentation describes newer versions; the Local harness is deprecated and scheduled for removal in February 2027. Check current native configuration before using a newer harness.

    You need a Desearch API key with available credits, outbound HTTPS access to https://mcp.desearch.ai/mcp, Chrome, and an existing tool-capable model endpoint. The test used a local llama.cpp b12478 server with Qwen3 0.6B GGUF SHA256 7f4030143c1c477224c5434f8272c662a8b042079a0a584f0a27a1684fe2e1fa, 32,768 context tokens and thinking disabled. This guide assumes the model server already works; it does not install or download a model.

    Local BYOK chat was tested without GitHub sign-in or a Copilot token. Official model documentation permits that path; some utility features require additional model configuration or an account. This does not establish hosted Copilot plan requirements or gallery eligibility. Desearch searches consume credits; check current pricing. Hosted Desearch deployment was unpinned; this is a dated behavior check.

    Create the isolated workbench

    Start in a new empty scratch directory. Supply DESEARCH_API_KEY from your secret manager to the shell, keeping its value out of files and prompts. Save the following JSON as workspace/.vscode/mcp.json, creating its parent folders first:

    json
    { "servers": { "desearch": { "type": "http", "url": "https://mcp.desearch.ai/mcp", "headers": { "Authorization": "Bearer ${env:DESEARCH_API_KEY}" } } } }

    Start the temporary workbench from that same scratch directory and environment. Port 19990 must be free. Keep this terminal open while using the workbench:

    sh
    : "${DESEARCH_API_KEY:?Set DESEARCH_API_KEY in the VS Code server environment}" mkdir -p home/.copilot cli server-data workspace/.vscode env HOME="$PWD/home" COPILOT_HOME="$PWD/home/.copilot" \ XDG_CONFIG_HOME="$PWD/home/.config" XDG_CACHE_HOME="$PWD/home/.cache" \ COPILOT_OTEL_ENABLED=false COPILOT_OTEL_CAPTURE_CONTENT=false \ COPILOT_OTEL_CAPTURE_IDENTITY=false OTEL_SDK_DISABLED=true \ code --cli-data-dir "$PWD/cli" serve-web \ --host 127.0.0.1 --port 19990 --without-connection-token \ --accept-server-license-terms --server-data-dir "$PWD/server-data" \ --default-folder "$PWD/workspace" --disable-telemetry \ --commit-id 6bab64e2d12b04d04c0d99c27ff166023efd1dda

    The CLI may download the matching official server. Open http://127.0.0.1:19990/ in a fresh browser context. This temporary listener has no connection token and is bound to loopback; stop it after testing. Trust only this scratch workspace when prompted, and decline opening a parent repository. The private home keeps the extension's IDE locks and user configuration separate from your ordinary profile. Despite the disabling settings above, this bundled extension logged an enabled file telemetry exporter. These settings do not establish that all extension tracing is disabled; treat the scratch data as private and remove it when no longer needed.

    Configure the existing model

    In the Command Palette (F1), run Chat: Manage Language Models. Choose Add Models, then Custom Endpoint. Set a group name, its API credential and the Chat Completions API type. For a local endpoint that needs no credential, the test used the dummy value local-no-auth. In the editor opened by VS Code, retain its generated apiKey input reference and set one model with id local-qwen3-0.6b, name Local Qwen3 0.6B, url http://127.0.0.1:19976/v1/chat/completions, toolCalling true, vision false, maxInputTokens 24000 and maxOutputTokens 4096. Save it. These id/URL settings match the test's already-running local server; use the actual settings of your own working endpoint.

    Select this model and the Local harness in chat. A utility-model notification does not by itself mean the MCP connection failed. This test used the main local model and inspected actual tool output.

    Start MCP and verify a native search

    Run MCP: List Servers from F1, choose desearch, and Start Server. The MCP: desearch output should show Running and Discovered 39 tools in this dated test; the count can change. This confirms discovery only.

    Open Configure Tools in chat. Clear all selected tools, expand Desearch, filter for web-search and check the checkbox beside web-search (SERP-style web search using Desearch). Clicking its row only highlights it. Confirm 1 Selected, choose OK, reopen to verify it stayed checked, then close with OK. Start a fresh chat after applying the selection.

    Send: Call Desearch web-search with query Python official documentation. Return the exact title and URL in the tool result.

    Require a native Ran Web Search step labeled desearch (MCP Server). If VS Code asks to review the returned tool result, inspect it and accept that result once with Command+Enter. Expand the completed step, then Ran Web Search, to inspect Input and Output. The actual JSON output must contain a nonempty data array with title and link. The clean replay returned Our Documentation | Python.org and Python 3.15.2 documentation. Titles and result counts may change; inspect the arguments the model actually sent.

    A generated answer alone is insufficient. In the test, model prose changed the title and replaced the returned URL; the raw native tool output held the real title and relative /go/v1/ redirect paths. Preserve returned link values. Use the native output's Copy action or select its code block, select all and copy if you need the exact JSON locally. Do not share complete sessions without reviewing them.

    Troubleshooting and cleanup

    If DESEARCH_API_KEY is missing, the shell check stops before the CLI starts. This validates local configuration, not upstream authentication. In the native invalid-key test, discovery still found 39 tools, but web-search returned Web Search error: HTTP 403 with detail Forbidden. Restore a valid credential and restart the server, then require actual titled data before calling the connection successful. Discovery, an HTTP connection or plausible model prose cannot substitute for a successful tool response.

    If no native Desearch call appears, verify the model supports tools, its context/runtime capacity, the persisted checkbox selection and a fresh chat. Earlier zero-tool-selection attempts queried the session store and fabricated search prose; they did not establish search. Allow only the intended search for this check.

    Stop the workbench with Ctrl+C and stop your test model server if you started one. Retain the scratch directory only as needed for local debugging; it can contain session data. This guide does not publish or register a gallery entry.

    Official references