DocumentationAI Assistant

AI Assistant

A built-in conversational sidebar that can drive the workspace for you — list your libraries, find a request by name, create or edit endpoints across all four protocols (HTTP / gRPC / Kafka / SOAP), set variables, and run requests — all from natural language.

AI Assistant

Bring your own model

ForgeReq is BYOK (bring-your-own-key). No data passes through a ForgeReq backend, because there is no ForgeReq backend. Conversations and tool calls go directly from your machine to the LLM provider you configured.

Supported providers:

  • OpenAI (gpt-4o, gpt-4o-mini, …)
  • Anthropic (claude-sonnet-4, claude-opus-4, …)
  • Google Gemini (gemini-2.5-pro, gemini-2.5-flash, …)
  • Google Vertex AI — via the OpenAI-compatible gateway path
  • Azure OpenAI — via your deployment’s /openai/responses endpoint
  • xAI (Grok)
  • Ollama — fully local; nothing leaves your machine at all

Every provider lets you override its baseURL, which makes it trivial to point the assistant at a corporate gateway, a custom proxy, or a self-hosted compatibility layer.

Setup

  1. Open Settings → AI Assistant.
  2. Click Add provider, choose the provider type, paste your API key, and (optionally) a custom baseURL.
  3. Add the model names you want to use — each provider supports multiple models, switchable per-conversation from the sidebar header.
  4. Click Save.
  5. Click the AI button in the title bar to open the sidebar.

API keys are stored inside the encrypted vault — same protection as the rest of your data.

What the assistant can do

The assistant has tools for the entire workspace. A few of the things it can do for you in one sentence:

  • Discover: “What’s in my Drafts library?” → calls listLibraries + listRequests and summarises.
  • Search: “Do I have a login request?” → fuzzy-searches by name + URL.
  • Create: “Create a GET /users/{{id}} request in the Users shelf” → calls createRequest with the right library, shelf, method, URL.
  • Create across protocols: “Make a Kafka consumer for orders.v2 on localhost:9092” → createRequest with protocol="kafka" plus the right broker / topic config. The same works for gRPC (proto + service + method) and SOAP (WSDL URL + service + operation).
  • Edit: “Add a Bearer {{token}} Authorization header to this request” → calls addHeaderToActiveRequest on the active tab.
  • Run: “Send the login request, then send GET /me” → resolves the targets, checks variables, sends, and surfaces the responses.
  • Manage env: “Switch realm to staging and set apiKey to demo_xyz” → setActiveRealm + upsertStashVar.

Privacy and persistence

  • Local-only. The sidebar talks to the LLM provider over the network you configured. ForgeReq never sees the conversation.
  • Persisted in the vault. Your single conversation history rides inside the encrypted vault, so closing and reopening ForgeReq picks up where you left off — same provider, same model, same scrollback.
  • Auto-compaction. Long conversations are summarised in the background by the same model to keep within its context window without losing earlier context. The conversation summary is also stored in the vault.
  • Clearable. The trash icon in the sidebar header wipes the conversation immediately.

Rich rendering

Assistant replies render as full Markdown with syntax-highlighted code blocks (powered by Shiki), so curl snippets, JSON examples, and tables come through cleanly. Copy buttons on every code block make it one click to lift a snippet into a terminal.

Tips

  • The header pill always shows the active provider · model — use the model picker right next to it to swap mid-conversation if a different model handles a specific question better.
  • When the assistant isn’t sure where to put a new request (e.g. you have multiple libraries), it will ask rather than silently pick a default. Confirm explicitly.
  • For destructive actions (delete library / shelf / request), the assistant will confirm with you before executing — don’t be surprised when it pauses for your “yes” before deleting anything.
  • The assistant can do searchRequests before sending a request by name, so “run the login request” works without you ever quoting an exact endpoint id.
Vault Backup