Skip to main content
Reactive Resume doesn’t include an AI model of its own. To use the assistant, AI imports, drafting and the other AI features, connect a provider such as OpenAI, Anthropic or Google Gemini. Everything else in the app works without AI. On a self-hosted instance, the administrator can configure a shared provider. Settings then show that AI is enabled globally, and personal provider controls are hidden. You can use AI features without entering your own key. The steps below apply when personal providers are available.

Before you start

  • Get an API key from your provider’s dashboard. You pay the provider directly for what you use. Local Ollama and OpenAI-compatible servers such as LM Studio can connect without a key.
  • Know which model you want to use, such as gpt-6-luna or claude-haiku-5-5. Reactive Resume suggests a default for each provider, but your account may not have access to every model.
Personal AI settings show a billing reminder before and after you connect a provider. Your provider may charge your account for requests made with your API key, according to its plan. Server-managed connections show Managed by your server instead.

Add a provider

1

Open AI settings

Select your name at the bottom of the sidebar to open Settings, then choose AI. You can also go to https://rxresu.me/dashboard/settings/ai directly. Self-hosted instances use their own address.
2

Start adding a provider

Under Connections, select Add a connection.
3

Choose the provider and fill in the fields

Pick your provider from the Provider list. Reactive Resume suggests its usual Model. Use Get an API key to open that provider’s key settings, then paste your key into API key. Switching providers clears the typed key. Change the model if you want a different one.For a named provider, expand Name and base URL only when you want a custom Name or Base URL. An existing custom address is shown when you edit the connection. For OpenAI-compatible, Base URL (required) stays visible. Enter an http:// or https:// address; an invalid or missing address gets an inline explanation and cannot be saved. The server separately checks which addresses it allows.
Add a connection dialog with provider, key, model, task models, name and base URL, and Save and test

Add an AI connection in one dialog

4

Save and test

Select Save and test. Reactive Resume saves the provider, then sends it a tiny request to check that the key, address and every model you configured work. If the test passes, the dialog closes and the provider is ready to use.
If the test fails, the dialog explains the cause and recovery in your language. Retrying Save and test updates the same saved connection. You can also close it and use Edit later. After your first successful connection in settings, AI is ready offers five next steps on web and mobile:
  • Improve selected line: choose a resume, then select the text you want to improve in Write.
  • Review writing: choose a resume and open its Writing check.
  • Check fit: choose an application and open Fit.
  • New cover letter: open letter creation, with an optional resume.
  • Ask the coach: open the Career coach.
If you have no resumes or applications yet, the pickers offer to create a resume or save a job. Opening a next step does not generate text or send your documents to AI. Each AI action still asks for approval before sending. You can dismiss these suggestions; adding another working connection does not show them again. Setup opened from an editing or coaching task keeps you in that task.
AI is ready offers Improve selected line, Review writing, Check fit, New cover letter and Ask the coach, with a dismiss button

Choose what to do after connecting your first AI provider

The fields

Choosing a model

Use Choose model, the arrow beside Model, to see the built-in choices even when a model is already selected. Recommended offers suggestions for everyday work and a More capable alternative. All models holds the additional built-in choices. On the web, typing filters the list; on mobile, the picker has a search field. OpenAI starts with gpt-6-luna, a fast, lower-cost model for everyday work. Defaults are starting suggestions for new connections. Saved models stay as you set them until you change them.
OpenAI model picker with Recommended and All models groups

Choose a built-in model or enter your own model ID

The built-in list is curated for text and tool use. Your endpoint and plan decide which models are available. You can enter any model ID your endpoint supports, including a private or custom model. Some providers have no built-in suggestions; their model field remains editable. Check tool-calling support with your provider when choosing a custom model for the assistant.

Task models (optional)

For lower cost without changing your everyday model, expand Task models under Model on web or mobile and choose:
  • A Fast model for resume import, Improve and reading job postings.
  • A Capable model for the assistant and letter drafts.
Leave either field empty to use the model above. A connection with one model therefore keeps working exactly as before, with the same key, provider and approval. Task models use that same connection; Reactive Resume never switches to another provider. Save and test checks every model you configured. If one is unavailable, the connection stays off until you correct or clear it, so an unsupported model is never used silently. Changing a task model also needs a fresh test, and pauses scheduled AI work until you re-enable each schedule. Each approval dialog names the model that will receive that action.
Add provider dialog with Task models (optional) expanded, showing empty Fast model and Capable model fields above collapsed Advanced settings

Optional fast and capable models on one tested connection

Supported providers

Suggestions favor fast, cost-efficient models for everyday AI tasks. They apply when adding a provider; saved models stay as selected. You can enter another model available to your account. Ollama defaults to a cloud model for its hosted base URL. For a local Ollama server, enter the name of a model installed on that server and change the base URL. Choose OpenAI-compatible for any other service that speaks the OpenAI API, such as a company gateway, LiteLLM or a local LM Studio server. It needs a model and a base URL; the API key is optional for endpoints that don’t require authentication. The base URL must start with https:// and point to a public address. On the hosted instance, that means a model running on your own computer (for example, a local Ollama) can’t be connected. Self-hosters can allow local and http:// addresses; see Environment variables.

Test a provider again

Each saved provider appears as a row under Connections, showing its name, model and the last four characters of the key. When you set task models, the row also shows which model fast and capable tasks use. The row shows Off next to the name when the provider isn’t in use.
AI settings with the fictional Work OpenAI connection

Your AI connections

Select Test on a row to check it again. The result appears separately: Connected with the round-trip time (for example, Connected · 420 ms), or an error explaining what to check. The button returns to Test, so you can run it again. A failed test makes the connection unavailable. A passed test makes it ready unless you deliberately switched it off; Test preserves that choice. Turn it on in Edit when you want to use it again. Each row’s ⋯ menu has Delete provider, so you can remove a connection without opening it.

Edit, turn off or delete a provider

Select Edit on a provider’s row.
Edit provider dialog for Fictional custom OpenAI with expanded Advanced fields, a custom local Base URL and the saved key left blank

Edit a saved provider without exposing its key

  • Change the key, model, task models, name or base URL, then select Save and test. Leave API key empty to keep the saved key. Changing the key, model, task models or base URL turns the provider off until the new test passes.
  • Turn it on or off with Use this provider. Only a provider that passed its test can be turned on. A provider that is off stays saved but isn’t offered anywhere in the app.
  • Delete it from the row’s ⋯ menu or the Delete provider button under Edit, then confirm with the red Delete provider button. Choose Keep provider to keep it. The provider and its stored key are removed only after confirmation.
Delete provider confirmation for Fictional working connection, with Keep provider and a red Delete provider button

Confirm before deleting a saved provider

Deleting a provider can’t be undone. Conversations keep their history and pending edits, but sending pauses until you explicitly choose another destination. Documents and edits you already accepted aren’t affected.

See your AI usage

Open Settings → AI → This week. On native mobile, open Settings → AI providers. The report shows recorded request attempts for your account, grouped by connection, action and requested model. Select Refresh usage to read the latest totals. Weeks run from Monday through Sunday in your saved account timezone; the report shows its dates and timezone. One request attempt means one call to the provider transport after the server’s authorization and context checks. A retry counts again. If an assistant calls a model, runs a tool, then calls the model again, that is two attempts. Stream chunks and tools that make no AI request add none. A provider-native search fallback makes an AI request and counts as an additional attempt. Copying, saving or reopening a result, deterministic checks, browser speech and actions refused before dispatch add none. Provider connection tests and scheduled AI work count too. Ordinary web-service search and page reading are outside this AI report. Explicit Save-job AI search retries appear as Web research; chat fallbacks stay under that conversation’s action. Older app versions group standalone AI search retries under Assistant.
AI settings with the fictional Work OpenAI connection, This week request attempts, Web access and memory preference

Manage AI connections and request attempts

Failed and cancelled transport attempts remain in history. These totals are not provider billing, token counts, costs or proof that the provider received a request. The server records an attempt before network I/O, so a server interruption can leave a recorded attempt unsent. Earlier requests are not reconstructed when recording first becomes available. Saved workspace Library results and Check → Writing reviews show their recorded attempt count, connection label, provider type and requested model. Accepted letter body drafts also show their sources in the letter editor and History. Earlier draft selection, revisions and recovery preserve the corresponding references. Reused AI answer sources keep their recorded lineage; saving or reusing them does not make another provider request. Older results say AI request history unavailable for this result. instead of guessing a count. Reviewed AI PDF and Word imports retain their sources in resume History on web and mobile, including owned copies and saved versions. Sources stay out of portable resume JSON and employer downloads. Older-server import results may have no recorded source; reading them again is not automatic. See Importing a resume. Improve shows the reviewed suggestion’s recorded attempts, original connection and model. Replace keeps those sources with the edited resume or letter; Keep mine leaves the document unchanged. Undo and Redo carry the sources with the accepted wording, including a suggestion with identical wording. Earlier import or draft sources remain. Saving, recovery, Undo and Redo never generate again; the weekly report still counts the original attempt if you discard or undo its result. Older-server suggestions remain usable with unknown request history. Accepting an assistant edit or a Writing rewrite also keeps the saved result’s recorded sources with the document. Pending, rejected and out-of-date edits add no sources. Undo restores the earlier content and sources; accepting again reuses the recorded result. If several accepted edits share one reply, undoing only one keeps that reply’s source for the others. These actions make no new provider request, and earlier records without sources remain unknown.
Resume History showing one AI request attempt from Fictional Review AI, OpenAI-compatible provider, review-main model, and the current and earlier saved versions.

History retains the accepted Writing review's original source after its fictional connection is deleted.

Renaming, changing or deleting a connection does not rewrite its historical labels or models. Usage remains readable without an active AI provider. Account export includes usage history, and deleting the account removes it. An older server without reporting shows an update/retry notice.

Which provider is used

You can save several providers, for example one per account or one per model.
  • One provider is your default, marked Default in Settings → AI. To change it, choose Make default on another provider that’s on and tested. Every AI feature where you don’t pick a provider uses the default, and the line under each AI action names it.
  • The first working connection becomes the explicit default. Using another connection in one conversation does not change it. If the default is unavailable, other actions offer the oldest working connection and ask you to approve that destination before sending.
  • Conversations keep the provider they started with. When it stops working, history and pending proposals remain readable. Choose another model and approve sending earlier messages there before continuing. Scheduled preparation uses only its selected connection. Deleting or switching off that connection, or changing its key, model or address, pauses associated schedules and cancels queued work. Choose a tested connection and approve it when you enable the schedule again; it never switches to the default automatically. A request already sent cannot be recalled.
  • Upgrading to recipient-bound scheduling pauses existing interview-preparation schedules once. Review and enable them again to approve their connection. If an administrator changes a server-managed connection’s details, scheduled work refuses that changed destination until you approve it again.
  • In Use this provider, the switch shows the saved result of each change. If turning it on or off fails, the switch stays as saved and the dialog says why.
  • Voice practice uses the default when its provider offers speech models, otherwise another provider that does, otherwise your browser.
  • The assistant lets you switch models from the model menu in its header at any time. See Using the assistant.
  • Check → Writing lets you pick the provider for each review.

Connect from the assistant instead

If no provider is connected yet, opening the assistant in the editor shows Connect an AI provider with OpenAI, Anthropic and Other · OpenAI-compatible. Enter your key and model there and select Connect. It saves and tests the provider the same way as settings. Every gated task also has Set up AI, which opens the same setup without leaving your work. Existing saved connections show repair controls instead of first-run copy. For the other providers, use AI. If connection lookup fails, retry it before setting up another provider. Your saved connections have not changed. Interview preparation schedules show a retry action while connections are unavailable; setup returns after a successful lookup. Interview preparation keeps its form and offers Try again when AI connections cannot be loaded
Assistant connection choices beside the editor, with no AI provider connected

The assistant needs a provider; manual editing remains available

Provider and model suggestions ship with the app. Opening settings does not request logos or a catalog from models.dev. You can still enter a custom model ID.

How your key is stored

Your personal API key is encrypted on the Reactive Resume server with AES-256-GCM before it’s saved, and it’s never shown again. Settings only ever show the last four characters. When a feature needs the key, the server decrypts it just long enough to call your provider. Self-hosted instances encrypt keys with the ENCRYPTION_SECRET environment variable. Until it’s set, AI shows “AI providers aren’t available on this server until ENCRYPTION_SECRET is set.” If the secret changes later, saved keys can no longer be decrypted. The error asks you to enter the key again: open Settings → AI, select Edit on the provider, paste the key, and select Save and test.

What is sent to your provider

Requests go from the Reactive Resume server to the provider you chose, using your key. Each interactive action asks you to approve its provider, model and data before sending. Canceling keeps the task available. Changes to the destination while approval is open require another approval. Scheduled requests run only after you enable the schedule. Each action explains its inputs and connection: Career adds these uses: Workspace text actions use your default connection. Voice settings can use another speech-capable connection or browser speech. Speech usage can be billed separately from text coaching. Automatic role discovery uses your web search connection; it does not need an AI model. See Career controls and data. A direct OpenAI, Anthropic or Google Gemini connection can retry search after the web service rejects its key or exhausts its quota, if you approve that possibility in the operation’s sharing dialog. The fallback receives the approved query and search country, without adding document or Knowledge context. It may add AI requests and charges. Custom endpoints and compatible gateways are not assumed to support it; individual models may still refuse. Scheduled discovery keeps using its web service. If web research fails in Coach or the document assistant, the message names a rejected key, exhausted quota or temporary request limit. Set up web access opens personal connection repair without losing your unsent message. A server-managed service directs you to the server administrator instead. If enhanced page reading fails and the built-in reader succeeds, the message says which reader was used; a working server search remains separate from your personal page reader.
Coach names a rejected web-search key, offers Set up web access, and keeps the next unsent message in the composer.

Repairing a rejected web-search key in Coach; fictional conversation

Your provider handles that data under its own terms and privacy policy. Reactive Resume keeps assistant conversations and their attachments in your account until you delete the conversation.

Using the assistant

Ask for changes and review the edits it proposes.

Assistant tools

What the assistant can read and do in a conversation.

Checking your resume

Get an AI review of your wording in Check mode.

Importing resumes

Turn an existing PDF or Word resume into an editable one.
Fictional compatible provider connected successfully while Off and Default remain visible, with instructions to turn it on in Edit.

A passed test preserves a deliberate Off choice.

Phone-width AI connection controls with an Off provider after a successful test.The same connection controls in dark appearance.

Connection controls at phone width, in light and dark appearance.

Phone-width provider list with the default connection first and a rejected fictional key below it, with Test still available.

A rejected key keeps Test available and explains what to check.

Before each generation, the server checks the selected documents and shared Knowledge again, then checks that the captured AI connection is still available. A provider turned off or deleted, or Knowledge excluded, before that authorization prevents the request. Authorization releases database locks before contacting the provider. A change committed after authorization may race an already authorized request; it cannot recall data already sent. Changing the default does not redirect an approved request to another connection. On upgrade, accounts without an explicit default use their oldest working provider. This can change the destination used by legacy API and MCP clients that previously followed the most recently used provider. Choose an explicit default in AI settings to keep a specific destination.