A coding agent in your terminal for the model you choose: local, uncensored or hosted.
Alt runs a full coding workspace (files, real terminal, checks, undo) around any model you bring: a GGUF file, Ollama, LM Studio, llama.cpp, vLLM, or a hosted API such as NVIDIA NIM or OpenRouter. Uncensored, abliterated and standard checkpoints are all first-class. Alt never swaps your model behind your back, and it shows what actually changed instead of trusting the model's word.
Download the latest Linux x86_64 build from Releases (see Installation for checksum and provenance checks), or build it:
# Debian/Ubuntu prerequisites: git build-essential pkg-config curl ca-certificates python3 libgomp1
git clone https://github.com/3d3dcanada/alt-cli.git
cd alt-cli
cargo build --locked --release # Rust 1.99.0 is selected automatically
./target/release/altAlt opens on Home without any configuration. When Home offers Install agent engine, accept it (a pinned, checksummed Goose build). Building Alt downloads no model weights.
Choose Connect a model on Home (or press Ctrl+P and type connect):
| You have | Pick | Address it fills |
|---|---|---|
| Ollama | Ollama | http://127.0.0.1:11434 |
| LM Studio | LM Studio | http://127.0.0.1:1234/v1 |
| llama.cpp, vLLM, SGLang | A hosted API or another server preset → llama.cpp / vLLM | http://127.0.0.1:8080/v1 / :8000/v1 |
| A hosted API | A hosted API or another server preset → NVIDIA NIM, OpenRouter, Groq, Together, DeepSeek, Mistral, Gemini, Cerebras | the provider's endpoint and key variable |
A .gguf file |
A GGUF file on this computer | Alt runs it with its own CPU runtime |
Keys stay in your environment: export the variable the preset names (for
example NVIDIA_API_KEY) before starting Alt; Alt saves only the variable name.
Pick the model from the server's list (type to search long hosted catalogs).
Headless, the same in one line:
alt init --preset nvidia --model z-ai/glm-5.3 # or: --preset ollama --model qwen3:8b
alt probe # one short request; suggests settingsProbe this model (Ctrl+P) sends one bounded request and suggests settings (thinking off, a larger output allowance, compact tools, native context) that you can apply with one choice. Thinking models that spend their whole allowance reasoning are detected and offered a labelled recovery; see Small-model usage.
- Try a practice project on Home (a small broken example with real checks and undo), or choose a project folder.
- Choose access in Settings → Tools & access: Review only, Guided changes (reviewed edits, checks isolated with Bubblewrap) or Full access (any command and network, with your own permissions).
- Type what you want in the Workspace, for example "The tests in test_cart.py fail; find out why and fix it." Review each proposed edit or command.
- Open Task progress to see the diff, run the checks and undo if needed. A model saying "done" does not mark the work verified; the checks do.
Everything is reachable with the keyboard or mouse: Ctrl+P lists every page, setting and action; F1 opens help; the TUI stays usable at 80×24.
Status: 0.7 in development on main (latest release 0.6.0 beta 3) · Linux
x86_64 · Apache-2.0. Model reliability and hardware support are measured
separately from application tests; see Status and evidence.
| Area | Included |
|---|---|
| Models | Hugging Face search, verified/resumable downloads, multipart GGUF, imported files and server model selection |
| Coding workspace | Browse/search/edit files, inspect diffs, checkpoint Alt edits, undo and resume conversations |
| Terminal | Real PTY sessions, interactive input, resize, attach/detach, persistent jobs, stop/restart and health checks |
| Tools | Native project tools; optional build, Git, HTTP, browser and security workflows; selected external MCP tools |
| Project memory | Saved goals and decisions, pinned notes, incremental source indexing, retrieved context and a visible Context page |
| Verification | Tests/build/lint/health/custom contracts, structured reports, independently pinned assertions and stale-result detection |
| Runtime controls | Context, CPU threads, GPU layers, batch/cache, MoE/tensor offload, RoPE/YaRN, draft models, extra llama-server arguments, generation and restart qualification |
| Any server | Thinking off/low/medium/high per server dialect, penalties, seed, stop, extra request JSON, Ollama context, model probe |
| Small-model effort | Explicit output/sampling, actual request receipts, host workflow, one active skill, source/import outlines and serial independently checked candidates |
| Offline improvement | Development-only instruction proposals with review/rollback; pending trajectory capture, SFT resume/export tooling and verified rewards |
| Recovery | State backup/restore, history retention, preserved imported weights, diagnostics and package rollback |
| Interface | Five themes, four layouts, mouse and keyboard; motion (page cascade, dialog pop, toasts, streaming cursor, gliding meters) with Full, Reduced and Off levels |
Full access uses normal host permissions. Guided changes reviews edits and uses Bubblewrap for isolated checks; it reports unavailable isolation instead of silently running on the host. Review only is also available. File checkpoints cover Alt edits; they do not undo arbitrary terminal, database or remote effects. Tool focus is a separate explicit choice; All is the default.
| Connection | Address or input | Notes |
|---|---|---|
| Ollama | http://127.0.0.1:11434 |
Use the server root; Alt sends the connection context as num_ctx and reads back what Ollama loaded |
| LM Studio | Usually http://127.0.0.1:1234/v1 |
Load a model and start its compatible server first |
| llama.cpp / ORA / other compatible servers | The server's OpenAI-compatible /v1 address |
Tool behavior depends on the model and chat template |
| Managed local inference | A compatible GGUF file or complete split set | The pinned download is CPU first; GPU runtime selection is explicit |
| Hosted APIs | NVIDIA NIM, OpenRouter, Groq, Together, DeepSeek, Mistral, Gemini, Cerebras presets | Key from an environment variable; retries and an output guard for flaky providers |
The reference target is Linux, 16 GB RAM and a GTX 1070 with 8 GB VRAM. Physical GTX 1070 performance has not been measured. Your selected 7B/9B Q4 model is a deployment target to qualify. Start with 4K–8K context and CPU or measured partial offload, then test your exact configuration. More context needs additional memory; project retrieval cannot enlarge a model's native context window. Pascal GPUs need a compatible CUDA 12 runtime, not a CUDA 13-only build. CPU mode remains available.
The Hub's variant filter uses publisher labels, not a quality certification. Qwen Heretic, Spark-X2.5 abliterated and MiMo Heretic derivatives have exploratory results in the project. The complete 0.6 CPU matrices passed 7/120 pinned behavioral checks for Spark and 16/120 for Qwen. Manual review and additional edge cases found unsupported claims and incomplete repairs even among passes. These are different model/budget configurations, not a matched comparison. No model preset is promoted. Real LM Studio, private Hub acquisition and GPU qualification remain open acceptance gates. See model/runtime qualification and live evaluations.
| Action | Control |
|---|---|
| Navigate | Alt+1…9; Alt+0 opens Jobs; Ctrl+P lists every page |
| Help / quick actions | F1 / Ctrl+P |
| Send / newline | Enter / Alt+Enter or Ctrl+J |
| New conversation / project brief | Ctrl+N / Ctrl+B |
| Stop work / quit | Esc / Ctrl+Q |
| Detach from an attached terminal | Ctrl+] |
alt --help
alt practice
alt doctor
alt hardware
alt models search 'Qwen3 4B heretic'
alt models import /path/to/model.gguf --uncensored --use
alt run 'Inspect this project and explain its test commands'
alt --access trusted run 'Run this project’s checks' --allow-tools --json
# terminal commands and external tools need their own flag:
alt --access trusted run 'Install and test' --allow-terminal --json
alt state backup /path/outside-alt-state/backup.tar.gzdocs/CLI.md covers setup, headless execution, jobs, verification
and export. State defaults to $XDG_DATA_HOME/alt-cli or
~/.local/share/alt-cli; use --data-dir or ALT_DATA_DIR to choose another folder.
Keys are referenced by environment-variable name, not saved in model profiles.
- Small-model implementation plan · Training handoff
- Small-model usage · Delivered work orders and remaining gates
- Installation and updates · Troubleshooting
- Workspace guide · Motion · CLI reference
- Verification contracts · Runtime qualification
- Architecture · Contributing · Changelog
- Implementation and evidence · 0.7 work orders · PC test guide
- Research and project history · Documentation index
Current source validation includes 162 Rust tests, strict Clippy, both adapter fixtures for both compact edit formats, and actual small-terminal permission, edit/check/undo and conversation-continuation journeys. Compact follow-ups retain the complete original request as context, with the latest user message taking precedence. The retained 0.6 acceptance also includes five PTY walkthroughs, 30 responsiveness journeys, 60 stream-recovery scenarios and packaged install/update/rollback checks on Debian 11 and Ubuntu 24.04. The local PC recorder also passed a live uncensored-model CPU trial. These measurements do not guarantee that a model can solve an arbitrary task. The current evidence separates application results, model outcomes and remaining human/hardware checks.
The 2026-10-08 PC trial (GTX 1070, 8 GB) measured why small models fail: reasoning cut-offs and edit format. Spark-X2.5 4B passed 0/8 tasks at 1,024 output tokens (every token spent thinking) and 3/8 at 4,096; MiMo 9B abliterated 3/8; hosted Kimi K3 4/8 and GLM 5.3 3/8. The 0.7 work orders (board) attack exactly that: provider-aware thinking control and cut-off recovery, open parameters, hosted presets and retries, a model probe and whole-file edits. The matched re-run on the owner's PC (A10) has not been run yet; no improvement is claimed until it is.
Earlier results: the repair follow-up
has independently verified MiMo 9B Q4 repairs (matched CPU pilot 3/4 per interface;
portable build 6/9). The complete 0.6 CPU matrices passed 7/120 pinned behavioural
checks for Spark and 16/120 for Qwen; no model preset is promoted. Physical GTX 1070
qualification, real LM Studio and novice sessions remain open gates. See
compatibility, qualification and
live evaluations. Raw evidence stays in docs/evidence;
new raw evidence over 1 MiB goes to release assets (scripts/check_repo_weight.py).
cargo test --locked
cargo fmt --all --check
cargo clippy --locked --all-targets -- -D warningsSee Contributing for terminal tests, optional tools and the
cloud setup. CI builds and tests a Linux package; successful runs expose a
linux-build-and-evidence artifact in GitHub Actions.
Git does not include prebuilt binaries or model weights.
Alt's Rust interface is original project code using a separately installed, replaceable Goose ACP engine. It is not an OpenCode fork. Alt is licensed under Apache-2.0; upstream components and models keep their own licenses. Third-party provenance records versions, attribution and distribution boundaries.
