A Neovim plugin that provides local edit completions and cursor predictions.
Note
We're exploring the idea of an open source dataset to improve completions, with real usage data from the community. If you're interested, join the discussion.
Warning
This is an early-stage project. Expect bugs, incomplete features, and breaking changes. Make sure to regularly update the plugin.
- Go 1.25.0+ (for building the server component)
- Neovim 0.8+ (for the plugin)
Using lazy.nvim
{
"cursortab/cursortab.nvim",
-- version = "*", -- Use latest tagged version for more stability
lazy = false, -- The server is already lazy loaded
build = "cd server && go build",
config = function()
require("cursortab").setup()
end,
}Using packer.nvim
use {
"cursortab/cursortab.nvim",
-- tag = "*", -- Use latest tagged version for more stability
run = "cd server && go build",
config = function()
require("cursortab").setup()
end
}The fastest way to get started is with the Mercury API provider (hosted, no local GPU needed):
-
Get an API key from Inception Labs
-
Set the environment variable:
export MERCURY_AI_TOKEN="your-api-key-here"
-
Add the provider to your setup:
require("cursortab").setup({ provider = { type = "mercuryapi", api_key_env = "MERCURY_AI_TOKEN", }, })
If you prefer local inference, use the sweep provider with
llama.cpp:
llama-server -hf sweepai/sweep-next-edit-1.5b --port 8000require("cursortab").setup({
provider = {
type = "sweep",
url = "http://localhost:8000",
},
})See Providers for all available options.
require("cursortab").setup({
enabled = true,
log_level = "info", -- "trace", "debug", "info", "warn", "error"
state_dir = vim.fn.stdpath("state") .. "/cursortab", -- Directory for runtime files (log, socket, pid)
keymaps = {
accept = "<Tab>", -- Keymap to accept completion, or false to disable
partial_accept = "<S-Tab>", -- Keymap to partially accept, or false to disable
trigger = false, -- Keymap to manually trigger completion, or false to disable
},
ui = {
completions = {
addition_style = "dimmed", -- "dimmed" or "highlight"
fg_opacity = 0.6, -- opacity for completion overlays (0=invisible, 1=fully visible)
},
jump = {
symbol = "", -- Symbol shown for jump points
text = " TAB ", -- Text displayed after jump symbol
show_distance = true, -- Show line distance for off-screen jumps
},
},
behavior = {
idle_completion_delay = 50, -- Delay in ms after idle to trigger completion (-1 to disable)
text_change_debounce = 50, -- Debounce in ms after text change to trigger completion (-1 to disable)
max_visible_lines = 12, -- Max visible lines per completion (0 to disable)
enabled_modes = { "insert", "normal" }, -- Modes where completions are active
cursor_prediction = {
enabled = true, -- Show jump indicators after completions
auto_advance = true, -- When no changes, show cursor jump to last line
proximity_threshold = 2, -- Min lines apart to show cursor jump (0 to disable)
},
ignore_paths = { -- Glob patterns for files to skip completions
"*.min.js",
"*.min.css",
"*.map",
"*-lock.json",
"*.lock",
"*.sum",
"*.csv",
"*.tsv",
"*.parquet",
"*.zip",
"*.tar",
"*.gz",
"*.pem",
"*.key",
".env",
".env.*",
"*.log",
},
ignore_filetypes = { "", "terminal" }, -- Filetypes to skip completions
ignore_gitignored = true, -- Skip files matched by .gitignore
},
provider = {
type = "inline", -- Provider: "inline", "fim", "sweep", "sweepapi", "zeta", "copilot", or "mercuryapi"
url = "http://localhost:8000", -- URL of the provider server
api_key_env = "", -- Env var name for API key (e.g., "OPENAI_API_KEY")
model = "", -- Model name
temperature = 0.0, -- Sampling temperature
max_tokens = 512, -- Max tokens to generate
top_k = 50, -- Top-k sampling
completion_timeout = 5000, -- Timeout in ms for completion requests
max_diff_history_tokens = 512, -- Max tokens for diff history (0 = no limit)
completion_path = "/v1/completions", -- API endpoint path
fim_tokens = { -- FIM tokens (for FIM provider)
prefix = "<|fim_prefix|>",
suffix = "<|fim_suffix|>",
middle = "<|fim_middle|>",
},
privacy_mode = true, -- Don't send telemetry to provider
},
blink = {
enabled = false, -- Enable blink source
ghost_text = true, -- Show native ghost text alongside blink menu
},
debug = {
immediate_shutdown = false, -- Shutdown daemon immediately when no clients
},
})For detailed configuration documentation, see :help cursortab-config.
The plugin defines the following highlight groups with default = true, so you
can override them in your colorscheme or config:
| Group | Default | Purpose |
|---|---|---|
CursorTabDeletion |
bg = "#4f2f2f" |
Background for deleted text |
CursorTabAddition |
bg = "#394f2f" |
Background for added text |
CursorTabModification |
bg = "#282e38" |
Background for modified text |
CursorTabCompletion |
fg = "#80899c" |
Foreground for completion text |
CursorTabJumpSymbol |
fg = "#373b45" |
Jump indicator symbol |
CursorTabJumpText |
bg = "#373b45", fg = "#bac1d1" |
Jump indicator text |
To customize, set the highlight before or after calling setup():
vim.api.nvim_set_hl(0, "CursorTabAddition", { bg = "#1a3a1a" })The plugin supports seven AI provider backends: Inline, FIM, Sweep, Sweep API, Zeta, Copilot, and Mercury API.
| Provider | Hosted | Multi-line | Multi-edit | Cursor Prediction | Streaming | Model |
|---|---|---|---|---|---|---|
inline |
Any base model | |||||
fim |
✓ | ✓ | Any FIM-capable | |||
sweep |
✓ | ✓ | ✓ | ✓ | sweep-next-edit-1.5b |
|
sweepapi |
✓ | ✓ | ✓ | ✓ | ✓ | sweep-next-edit-7b |
zeta |
✓ | ✓ | ✓ | ✓ | zeta |
|
copilot |
✓ | ✓ | ✓ | ✓ | GitHub Copilot | |
mercuryapi |
✓ | ✓ | ✓ | ✓ | mercury-edit |
Context Per Provider:
| Context | inline | fim | sweep | zeta | sweepapi | copilot | mercuryapi |
|---|---|---|---|---|---|---|---|
| Buffer content | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |
| Edit history | ✓ | ✓ | ✓ | ✓ | |||
| Previous file state | ✓ | ✓ | |||||
| LSP diagnostics | ✓ | ✓ | ✓ | ||||
| Treesitter context | ✓ | ✓ | ✓ | ✓ | |||
| Git diff context | ✓ | ✓ | ✓ | ✓ | |||
| Recent files | ✓ | ✓ | |||||
| User actions | ✓ |
Details
End-of-line completion using OpenAI-compatible API endpoints. Works with any
OpenAI-compatible /v1/completions endpoint.
Requirements:
- An OpenAI-compatible completions endpoint
Example Configuration:
require("cursortab").setup({
provider = {
type = "inline",
url = "http://localhost:8000",
},
})Example Setup:
# Using llama.cpp
llama-server -hf ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF --port 8000Details
Fill-in-the-Middle completion using standard FIM tokens. Uses both prefix
(before cursor) and suffix (after cursor) context. Compatible with Qwen,
DeepSeek-Coder, and similar models. Works with any OpenAI-compatible
/v1/completions endpoint.
Requirements:
- An OpenAI-compatible completions endpoint with a FIM-capable model
Example Configuration:
require("cursortab").setup({
provider = {
type = "fim",
url = "http://localhost:8000",
},
})Example Setup:
# Using llama.cpp with Qwen2.5-Coder 1.5B
llama-server -hf ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF --port 8000
# Or with Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding
llama-server \
-hf ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF:q8_0 \
-hfd ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF:q8_0 \
--port 8012 \
-b 1024 \
-ub 1024 \
--cache-reuse 256Details
Sweep Next-Edit 1.5B model for fast, accurate next-edit predictions. Sends full file for small files, trimmed around cursor for large files.
Requirements:
- vLLM or compatible inference server
- Sweep Next-Edit model downloaded from Hugging Face
Example Configuration:
require("cursortab").setup({
provider = {
type = "sweep",
url = "http://localhost:8000",
},
})Example Setup:
# Using llama.cpp
llama-server -hf sweepai/sweep-next-edit-1.5b --port 8000
# Or with a local GGUF file
llama-server -m sweep-next-edit-1.5b.q8_0.v2.gguf --port 8000
# With caching and speculative decoding
llama-server \
-m sweep-next-edit-1.5b.q8_0.v2.gguf \
--port 8000 \
--cache-reuse 64 \
--spec-type ngram-mod \
--spec-ngram-size-n 24 \
--draft-min 12 \
--draft-max 64Details
Sweep's hosted API for Next-Edit predictions. No local model setup required.
Requirements:
- Create an account at sweep.dev and get your API token
- Set the
SWEEPAPI_TOKENenvironment variable with your token
Example Configuration:
# In your shell config (.bashrc, .zshrc, etc.)
export SWEEPAPI_TOKEN="your-api-token-here"require("cursortab").setup({
provider = {
type = "sweepapi",
api_key_env = "SWEEPAPI_TOKEN",
},
})Details
Zed's Zeta model - a Qwen2.5-Coder-7B fine-tuned for edit prediction using DPO and SFT.
Requirements:
- vLLM or compatible inference server
- Zeta model downloaded from Hugging Face
Example Configuration:
require("cursortab").setup({
provider = {
type = "zeta",
url = "http://localhost:8000",
model = "zeta",
},
})Example Setup:
# Using vLLM
vllm serve zed-industries/zeta --served-model-name zeta --port 8000
# See the HuggingFace page for optimized deployment optionsDetails
GitHub Copilot completions using the official
copilot-language-server
LSP server, enabled with vim.lsp.enable. Can be installed in multiple ways:
- Install using
npmor your OS's package manager - Install with mason-lspconfig.nvim
- copilot.lua and copilot.vim both bundle the LSP server in their plugin
- Sign in to Copilot:
:LspCopilotSignIn
Requirements:
- GitHub Copilot subscription
copilot-language-serverinstalled and enabled viavim.lsp.enable
Example Configuration:
require("cursortab").setup({
provider = {
type = "copilot",
},
})Details
Mercury is Inception Labs' hosted Next-Edit prediction model.
Requirements:
- Mercury API key (set via
MERCURY_AI_TOKENenvironment variable)
Example Configuration:
require("cursortab").setup({
provider = {
type = "mercuryapi",
api_key_env = "MERCURY_AI_TOKEN",
},
})Details
This integration exposes a minimal blink source that only consumes
append_chars (end-of-line ghost text). Complex diffs (multi-line edits,
replacements, deletions, cursor prediction UI) still render via the native UI.
require("cursortab").setup({
keymaps = {
accept = false, -- Let blink manage <Tab>
},
blink = {
enabled = true,
ghost_text = false, -- Disable native ghost text
},
})
require("blink.cmp").setup({
sources = {
providers = {
cursortab = {
module = "cursortab.blink",
name = "cursortab",
async = true,
-- Should match provider.completion_timeout in cursortab config
timeout_ms = 5000,
score_offset = 50, -- Higher priority among suggestions
},
},
},
})- Tab Key: Navigate to cursor predictions or accept completions
- Shift-Tab Key: Partially accept completions (word-by-word for inline, line-by-line for multi-line)
- Esc Key: Reject current completions
- The plugin automatically shows jump indicators for predicted cursor positions
- Visual indicators appear for additions, deletions, and completions
- Off-screen jump targets show directional arrows with distance information
:CursortabToggle: Toggle the plugin on/off:CursortabShowLog: Show the cursortab log file in a new buffer:CursortabClearLog: Clear the cursortab log file:CursortabStatus: Show detailed status information about the plugin and daemon:CursortabRestart: Restart the cursortab daemon process
To build the server component:
cd server && go buildTo run tests:
cd server && go test ./...To run the E2E pipeline tests (ComputeDiff → CreateStages → ToLuaFormat):
cd server && go test ./text/... -run TestE2E -vTo record new expected output after changes:
cd server && go test ./text/... -run TestE2E -updateUpdated fixtures are marked as unverified. After reviewing the generated
expected.json files and the HTML report, mark a specific fixture as verified:
cd server && go test ./text/... -run TestE2E -verify <name>Verification state is tracked in server/text/e2e/verified.json (a SHA256
manifest). In the HTML report, verified passing tests are collapsed while
unverified or failing tests are shown expanded.
Each fixture is a directory under server/text/e2e/ with old.txt, new.txt,
params.json, and expected.json. Both batch and incremental pipelines are
verified against the same expected output. An HTML report is generated at
server/text/e2e/report.html.
To add a new fixture manually, create a directory with the four files and run
with -update to generate the initial expected.json, then review and
-verify-case=<name> to mark it verified.
Why are completions slow?
- Use a smaller or more heavily quantized model (e.g., Q4 instead of Q8)
- Decrease
provider.max_tokensto reduce output length (also limits input context)
The copilot provider is known to be slower than the rest.
Why are completions not working?
- Update to the latest version, restart Neovim and restart the daemon with
:CursortabRestart - Increase
provider.completion_timeout(default: 5000ms) to 10000 or more if your model is slow - Increase
provider.max_tokensto give the model more surrounding context (trade-off: slower completions)
How do I update the plugin?
Use your Neovim plugin manager to pull the latest changes, then restart Neovim.
The daemon will automatically restart if the configuration has changed. You can
also run :CursortabRestart to force a restart.
Why isn't my API key or environment variable being picked up?
The plugin runs a background daemon that persists after Neovim closes.
Environment variables are only loaded when the daemon starts. If you add or
change an environment variable (e.g., SWEEPAPI_TOKEN in your .zshrc), run
:CursortabRestart to restart the daemon with the new environment variables.
Note: If you change plugin configuration (e.g., switch providers), the daemon
will automatically restart on the next setup() call.
Contributions are welcome! Please open an issue or a pull request.
Feel free to open issues for bugs :)
This project is licensed under the MIT License - see the LICENSE file for details.
![@github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=64&v=4){kind=small}
