From 07e661e8e63f00fc1b33d0461884125018608a9e Mon Sep 17 00:00:00 2001 From: yashbudhia Date: Tue, 20 Oct 2026 16:54:40 +0530 Subject: [PATCH] docs: add Dreamtune MCP GPU planning cookbook --- docs/docs.json | 1 + docs/guides/dreamtune-mcp-cookbook.mdx | 140 +++++++++++++++++++++++++ docs/guides/overview.mdx | 6 ++ 3 files changed, 147 insertions(+) create mode 100644 docs/guides/dreamtune-mcp-cookbook.mdx diff --git a/docs/docs.json b/docs/docs.json index b7a1d83f13a..f46422780d5 100644 --- a/docs/docs.json +++ b/docs/docs.json @@ -176,6 +176,7 @@ "guides/configuring-models-rules-tools", "guides/codebase-documentation-awareness", "guides/plan-mode-guide", + "guides/dreamtune-mcp-cookbook", "guides/ollama-guide", "guides/instinct", "guides/running-continue-without-internet", diff --git a/docs/guides/dreamtune-mcp-cookbook.mdx b/docs/guides/dreamtune-mcp-cookbook.mdx new file mode 100644 index 00000000000..723d2448d97 --- /dev/null +++ b/docs/guides/dreamtune-mcp-cookbook.mdx @@ -0,0 +1,140 @@ +--- +title: "Plan GPU Workloads with Dreamtune MCP" +description: "Connect Dreamtune to Continue, inspect GPU options, and prepare a workload plan without launching compute." +sidebarTitle: "Dreamtune MCP Cookbook" +--- + +Use [Dreamtune MCP](https://docs.dreamtunelabs.com/connect/mcp) with Continue's +Agent mode to inspect tool definitions and GPU options while working on your +application. This cookbook connects an existing Continue agent to Dreamtune's +tools; your configured Continue model still runs the conversation. + +The example workflow compares GPU options for a development workload and produces +a plan for review. It does not launch a job. + +## Prerequisites + +- Continue installed with a model that supports [Agent mode](/ide-extensions/agent/how-it-works). +- A Dreamtune workspace API key. See [Dreamtune authentication](https://docs.dreamtunelabs.com/authentication). +- For the optional local server, Node.js 22.20 or later on the machine where Continue runs. + +## Store your workspace key + +Add the key to your private `~/.continue/.env` file: + +```dotenv title="~/.continue/.env" +DREAMTUNE_API_KEY=YOUR_DREAMTUNE_WORKSPACE_KEY +``` + +Continue resolves the secret reference in the configuration below. You can also +use a private workspace `.continue/.env` file; exclude that file from version +control. See [Working with secrets](/guides/configuring-models-rules-tools#working-with-secrets) +for resolution order. + +## Connect the hosted server + +Create `.continue/mcpServers/dreamtune.yaml` in your workspace: + +```yaml title=".continue/mcpServers/dreamtune.yaml" +name: Dreamtune +version: 1.0.0 +schema: v1 +mcpServers: + - name: Dreamtune + type: streamable-http + url: https://mcp.dreamtunelabs.com/mcp + connectionTimeout: 60000 + requestOptions: + headers: + X-Dreamtune-API-Key: ${{ secrets.DREAMTUNE_API_KEY }} +``` + +Dreamtune uses the `X-Dreamtune-API-Key` header. Keep the header configuration +rather than setting Continue's `apiKey` property, which sends a bearer token. + +### Alternative: run the local stdio server + +To use the published local server instead, replace the contents of the same YAML +file with this configuration. Choose one transport so tools are not registered +twice. + +```yaml title=".continue/mcpServers/dreamtune.yaml" +name: Dreamtune +version: 1.0.0 +schema: v1 +mcpServers: + - name: Dreamtune + type: stdio + command: npx + args: + - "-y" + - "--package" + - "@dreamtune/sdk@0.1.2" + - "dreamtune-mcp" + env: + DREAMTUNE_API_KEY: ${{ secrets.DREAMTUNE_API_KEY }} + DREAMTUNE_API_ORIGIN: https://api.dreamtunelabs.com + DREAMTUNE_REQUEST_TIMEOUT_MS: "60000" + connectionTimeout: 60000 +``` + +The pinned version makes the local tool definitions reproducible. Review the +[local MCP setup guide](https://docs.dreamtunelabs.com/connect/local-mcp) before +updating it. Continue starts and manages this process; you do not need to start a +separate server in a terminal. + +## Verify tool discovery + +Switch Continue to Agent mode. Confirm that Dreamtune's tools appear in the tool +list, including `list_tools` and `compute_catalog`. + +Tool discovery verifies the MCP connection. It does not run a model, create a +compute job, or prove that your key can execute every tool. The local stdio server +can expose its definitions before making an authenticated API request. + +Use Continue's [tool policies](/ide-extensions/agent/how-to-customize#how-to-configure-tool-policies) to review +calls before they execute. For this workflow, keep model execution, compute +creation, and other write tools disabled or subject to approval. + +## Compare GPU options for your project + +Ask the agent to inspect the current catalog: + +```text +Use Dreamtune's compute_catalog tool to find GPU options for testing a 7B model. +Compare up to three options using the returned VRAM, region, and price data. +State when a requested field is missing, and distinguish hourly price from a +total cost estimate. Do not create, start, or stop any compute jobs. +``` + +Expand the tool call in Continue and inspect the returned data. Check that the +comparison uses actual catalog results rather than assumed hardware or prices. +An empty result is useful too: ask the agent to explain which requirement could +be relaxed instead of inventing an available machine. + +Then use your repository as context to prepare a plan: + +```text +Using the GPU options you just retrieved and this repository's dependencies, +draft a setup plan for one short development experiment. Include the selected +hardware, runtime requirements, expected duration, and a shutdown step. Label +cost estimates as estimates and list any missing information. Return the plan +in chat for review; do not execute commands or launch compute. +``` + +To add tool discovery to the same workflow, ask the agent to use `list_tools` and +`inspect_tool` to find an appropriate API tool and inspect its input schema before +proposing a call. See the [Dreamtune MCP tool reference](https://docs.dreamtunelabs.com/reference/mcp-tools) +for the distinction between catalog reads and execution tools. + +## Troubleshooting + +- **No tools appear:** Check Agent mode, the YAML file's location, and Continue's MCP connection errors. +- **Unresolved secret:** Verify that `DREAMTUNE_API_KEY` is in one of Continue's supported private secret locations. +- **Authentication fails:** Use a workspace API key and the exact `X-Dreamtune-API-Key` header for hosted MCP. +- **The local command cannot start:** Confirm Node.js 22.20 or later and `npx` are available to the IDE process. Restart the IDE after changing its environment. +- **Discovery succeeds but a call fails:** Check the workspace key's permissions, balance, and the tool's required inputs. Do not repeatedly retry a paid write to diagnose connectivity. + +Model and tool execution can incur usage charges, and running compute can +continue to accrue charges until it is stopped. Review the specific operation +and its inputs before extending this planning workflow to execution. diff --git a/docs/guides/overview.mdx b/docs/guides/overview.mdx index b336acf68cc..608a946a7cb 100644 --- a/docs/guides/overview.mdx +++ b/docs/guides/overview.mdx @@ -64,6 +64,12 @@ Step-by-step guides for integrating Model Context Protocol (MCP) servers with Co Build AI-powered data pipelines with dlt MCP for pipeline inspection, schema management, and debugging + + + Inspect GPU options and prepare a workload plan with Dreamtune MCP before launching compute + + + ## What Advanced Tutorials Are Available