FIM Autocomplete

Inline code completion (fill-in-the-middle) for VS Code, backed by any LLM provider you point it at. No chat, no agent, no codebase indexing — just autocomplete.

This is a fork of Continue reduced to its autocomplete engine. Source and issues: dkruyt/FIM-Autocomplete.

Setup

Run FIM: Select Model from the command palette, or click the FIM status bar item and choose Select model…. It walks you through provider → connection → model, and for providers that support it (Ollama, LM Studio, vLLM, and any OpenAI-compatible server) it lists the models actually available on the server rather than making you type a name.

Or set it by hand in settings.json:

{
  "fim.model": {
    "provider": "ollama",
    "model": "qwen2.5-coder:1.5b",
    "apiBase": "http://localhost:11434",
  },
}

Any provider with a fill-in-the-middle endpoint works well — Ollama, vLLM, Mistral/Codestral, DeepSeek, LM Studio, or an OpenAI-compatible server:

{
  "fim.model": {
    "provider": "mistral",
    "model": "codestral-latest",
    "apiKey": "...",
  },
}

The picker offers the eleven providers that make sense for autocomplete, but provider accepts any of the 61 the engine ships with — set it by hand if yours isn't listed.

apiKey is stored in settings.json in plaintext and will sync via VS Code's Settings Sync. Use a local provider, or an environment-scoped key, if that matters to you.

Using it

Just type. Once you pause for fim.debounceDelay (350 ms by default), the suggestion appears inline as grey ghost text after your cursor.

Accept, dismiss and accept-word are VS Code's own inline-suggestion bindings, so if you've rebound them they keep working here. Only one suggestion is offered per position — there's nothing to cycle through with alt+] / alt+[.

cmd+alt+\ is the one to remember: it skips the debounce and asks the model straight away, which is what you want when a suggestion didn't appear on its own. The file-exclusion rules below still apply — forcing won't get you a completion in an ignored file.

The status bar

The FIM item on the right shows what the extension is doing. Click it for a menu with Select model…, an enable/disable toggle, and a link to all settings.

Commands

All are under the FIM: prefix in the command palette.

When nothing appears

In rough order of likelihood:

• The model isn't a code model. A general chat model gets sent raw <fim_prefix>-style tokens it has never been trained on, and whatever it replies gets discarded by the output filters. Use a fill-in-the-middle code model — qwen2.5-coder, codestral, deepseek-coder, starcoder2, codegemma. FIM: Select Model warns you when a name doesn't look like one.
• The server isn't reachable. Errors surface as a notification, with a "Start Ollama" button when that's the problem. Check fim.model.apiBase.
• VS Code's inline suggestions are off. editor.inlineSuggest.enabled must be true (it is by default) or providers never get called.
• The file is excluded. Check fim.disableInFiles, plus .fimignore in the workspace root and ~/.fim/.fimignore (gitignore syntax). *.prompt files are always skipped, as are empty untitled files.
• The file looks like it holds secrets. Paths matching .env*, *.key, *.pem, *.p12, certificates and keystores are refused outright — and so are settings.json, config.json and config.yaml, which surprises people. This isn't configurable.

Suggestions arriving truncated is a different problem: fim.modelTimeout (150 ms) caps how long output keeps streaming after the first non-empty line, so a slow local model gets its multi-line completions cut short. Raise it to 500–1000 for local models on modest hardware.

For anything else, Help → Toggle Developer Tools → Console shows the errors.

How it builds a prompt

Beyond the text around your cursor, completions get context from:

• definitions of imported symbols (tree-sitter + LSP)
• the enclosing class/function signature and the types it references
• recently edited and recently visited ranges
• recently opened files
• optionally the clipboard

Most of these can be toggled under the fim.* settings, and each is raced against its own timeout so a slow one can't stall a keystroke.

Settings

See all options under Settings → Extensions → FIM Autocomplete, or search fim. in settings.json. Useful ones:

• fim.enabled — master switch
• fim.debounceDelay — ms to wait after a keystroke (default 350)
• fim.maxPromptTokens — prompt budget (default 1024)
• fim.multilineCompletions — always / never / auto
• fim.disableInFiles — glob patterns to skip

A global ~/.fim/.fimignore and per-workspace .fimignore files also suppress completions, using gitignore syntax.

What gets sent where

Every keystroke that triggers a completion sends a prompt to the provider you configured in fim.model — and to nowhere else. That prompt contains the code around your cursor plus the context sources listed above, so assume anything in your open files can reach that endpoint. Point it at a local Ollama or vLLM if that's not acceptable.

There is no telemetry, no analytics, and no account. The only requests that don't go to your configured provider are ones you explicitly ask for: clicking Install Model on an error notification looks the model up in the Ollama registry first.

License

Apache-2.0 — see LICENSE.txt. Original work © 2023-2026 Continue Dev, Inc.; see the NOTICE file in the source repository for what changed.