Leyline: Code Anticipation
Inline AI code completion for VS Code. Bring your own endpoint and API key, or
run models locally — works with
Codestral (Mistral FIM) and
Ollama.
Why Leyline
Most AI completion tools send your code to a fixed cloud service and charge a
subscription. Leyline is different:
- Bring your own endpoint — point it at Codestral, an Azure AI deployment,
or any FIM-compatible endpoint you control.
- Run fully local — use Ollama with
qwen2.5-coder, codellama, or any
FIM-capable model. No API key, no cloud calls.
- No telemetry — only your configured endpoint ever receives code snippets.
Nothing goes to Anthropic, Mistral, or any third party.
- Cross-file context — completions include type definitions and function
signatures from related open files, not just the current buffer.
Installation
Search for Leyline in the VS Code Extensions panel (Ctrl+Shift+X) and
click Install.
Alternatively, download the VSIX from the
Releases page and
run:
code --install-extension leyline-*.vsix
Quick Start
- Open the Command Palette (
Ctrl+Shift+P / Cmd+Shift+P)
- Run Leyline: Select Provider — choose Codestral (cloud) or Ollama (local)
- Codestral: Run Leyline: Set API Key and paste your Codestral API key
Ollama: Make sure Ollama is running (ollama serve) with a FIM model:
ollama pull qwen2.5-coder:7b
- Open any source file and start typing — ghost text appears after a brief pause
Press Tab to accept · Escape to dismiss · Alt+\ to force-trigger
Features
- Inline ghost text — completions appear as you type, accept with
Tab.
- Cross-file context — includes types, functions, and signatures from open
files so the model understands your project, not just the current file.
- Smart multiline — auto-detects function bodies, docstrings, and JSDoc to
generate multi-line completions where it matters.
- Tab override —
Tab accepts completions without conflicting with
IntelliSense. When IntelliSense is open, Tab goes to IntelliSense as usual.
- Force trigger — press
Alt+\ to request a completion on demand.
- Streaming — responses stream in real time; cancellation works mid-generation.
- Bracket-aware — post-processing truncates completions that would duplicate
closing braces already in your code.
- Tree-sitter validation — opt-in syntax check that silently discards
completions introducing parse errors (grammars downloaded on first use).
- Completion cache — recent completions are served instantly on undo/retype.
- Per-language control — enable or disable completions per language or file
glob pattern.
- Secure credentials — API keys stored in your OS keychain via VS Code
SecretStorage; never logged or sent anywhere except the configured endpoint.
- Reactive status bar — shows provider state, disables per language/glob,
and displays error details on hover.
- Multi-provider — Codestral (cloud) and Ollama (local) out of the box.
Commands
| Command |
Shortcut |
Description |
Leyline: Set API Key |
|
Store your API key in the OS keychain |
Leyline: Toggle Inline Completion |
|
Enable or disable completions |
Leyline: Select Provider |
|
Switch between Codestral and Ollama |
Leyline: Trigger Inline Completion |
Alt+\ |
Request a completion on demand |
Leyline: Show Menu |
|
Quick access to all actions |
Leyline: Test Connection |
|
Verify the endpoint is reachable |
Configuration
Settings are organized in categories under the leyline.* namespace. Open
Settings (Cmd+,) and search for "leyline" to see all options grouped by:
General, Filters, Completion, Quality, Codestral, Ollama.
Key settings
| Setting |
Default |
Description |
leyline.provider |
"codestral" |
AI backend (codestral or ollama) |
leyline.tabOverride |
true |
Let Tab accept completions when IntelliSense is closed |
leyline.enable |
{ "*": true } |
Per-language toggle (e.g., { "*": true, "markdown": false, "json": false, "plaintext": false }) |
leyline.disableInFiles |
[] |
Glob patterns to exclude (e.g., "**/*.md") |
leyline.multiline |
"auto" |
auto / always / never |
leyline.treeSitter |
false |
Syntax validation via Tree-sitter |
leyline.crossFileContext |
true |
Include context from related open files |
leyline.crossFileContextTokens |
500 |
Max tokens of cross-file context (0–2000) |
leyline.cacheSize |
50 |
Cached completions (0 to disable) |
leyline.debounceMs |
300 |
Typing debounce in ms |
leyline.requestTimeout |
30 |
Request timeout in seconds |
leyline.prefixLines |
100 |
Lines of context before cursor |
leyline.suffixLines |
30 |
Lines of context after cursor |
Each provider also has endpoint, model, and maxTokens settings — see the
Settings UI for details.
Tested Providers and Models
| Provider |
Model |
Notes |
| Codestral |
codestral-latest |
Default. Cloud, requires API key from console.mistral.ai |
| Ollama |
qwen2.5-coder:7b |
Default local model. Best all-round for FIM |
| Ollama |
codellama:7b |
Good alternative local model |
| Ollama |
deepseek-coder:6.7b |
Also tested, solid results |
| Azure AI |
Codestral-2501 |
Set leyline.codestral.endpoint to your Azure endpoint |
Any provider that implements the Mistral FIM API (POST /v1/fim/completions) or
the Ollama API (POST /api/generate with suffix support) should work.
Supported Languages
Cross-file context includes import-aware type and signature extraction for:
TypeScript · JavaScript · TSX / JSX · Python · Go · Java · Rust · C / C++
Basic inline completion (ghost text, FIM) works in any language. Cross-file
context is language-specific but everything else is language-agnostic.
Privacy
When you request a completion, Leyline sends the text immediately before and
after your cursor — and, if cross-file context is enabled, short type and
signature snippets from related open files — to the endpoint you configured.
That is the only outbound network call Leyline makes.
- No telemetry, no analytics, no usage tracking.
- API keys are stored in your OS keychain via VS Code SecretStorage — never
written to disk, never logged.
- Leyline has no connection to Anthropic, Mistral, or any service other than
your configured endpoint.
Troubleshooting
Status bar shows a warning icon
No API key is configured for the current provider. Click the status bar to open
the Leyline menu, then choose Set API Key.
Completions are not appearing
Open Output → Leyline for error details. Common causes: incorrect API key,
wrong endpoint URL, or a model that does not support FIM.
Ollama completions are empty or very slow
Make sure the model supports fill-in-the-middle (FIM). Not all Ollama models do.
qwen2.5-coder:7b is the most reliable choice.
Tree-sitter keeps rejecting completions
If you enabled leyline.treeSitter and completions rarely appear, the validator
may be too strict for your codebase or language. Disable it to restore normal
behaviour.
Tab conflicts with another extension or snippet
Set leyline.tabOverride: false to remove the Tab keybinding. Use Alt+\ to
accept completions instead.
Using Leyline alongside GitHub Copilot
Both extensions compete for ghost text. If you want Leyline to be the sole
completion source, disable Copilot: Extensions panel → search "GitHub Copilot"
→ Disable.
How It Works
- A debounce timer starts when you type (default 300 ms)
- Prefix/suffix text around the cursor is extracted
- (If enabled) Signatures from related open files are prepended as context
- A FIM prompt is sent to your configured provider; the response streams back
- Post-processing: repetition stripping, suffix overlap removal, bracket
overflow truncation, (opt-in) Tree-sitter syntax validation
- The cleaned text appears as inline ghost text
Requests cancel automatically when you keep typing; recent completions are
cached for instant replay on undo/retype.
Contributing
See CONTRIBUTING.md for development setup, coding standards,
and the release workflow.
License
MIT
