itsy

A coding agent for small local LLMs (8B–35B parameters). For when you've got a model running locally and want it to actually edit files without it spinning itself into a loop.

_{Composed from actual run logs (a real cobol-modernization solve on Qwen3.6-35B IQ2_XXS); the source ANSI lives at docs/itsy-tui.ansi and is re-rendered via docs/regen-screenshot.sh.}

The core trade-off is tokens and speed for reliability. itsy is slower and chattier than a thin wrapper. The payoff is that it actually finishes things. Local models at 4-bit quant drift, repeat themselves, forget what file they were editing — itsy burns the extra tokens on guardrails so that stays contained. The tool-call parser accepts JSON, XML-ish, and code-fence wrappers. Edits are search-and-replace patches, not full rewrites. Multi-step work decomposes into a small TODO list so the model has somewhere to look when it loses the thread. There's a SQLite + FTS5 code graph and per-project memory so it doesn't reintroduce itself every turn.

Works well with Qwen3.5/3.6 and Gemma4 at IQ2–Q4. Talks to anything OpenAI-compatible: LM Studio, Ollama, vLLM, llama-server.

Install

You need Rust 1.85+ (edition 2024) and a running OpenAI-compatible endpoint. Everything else is bundled.

git clone https://github.com/Doorman11991/itsy
cd itsy
cargo build --release

The binary lands at target/release/itsy. Copy it onto your $PATH or invoke it from the repo. There's no cargo install recipe yet because the build picks up native-feature flags from Cargo.toml that cargo install doesn't propagate cleanly.

If you're running against bench images or older base distros that ship a too-old glibc, build statically against musl:

rustup target add x86_64-unknown-linux-musl
apt-get install -y musl-tools
cargo build --release --target x86_64-unknown-linux-musl

First run

itsy

On the very first launch (no ~/.config/itsy/config.toml) the wizard asks for provider, endpoint, model name, and the safeguards worth turning on for your quant tier. It probes /v1/models for the real context window so you don't have to guess. Re-run the wizard any time with itsy --init.

After that, every launch drops straight into the REPL:

• itsy — interactive fullscreen REPL (default)
• itsy --classic — line-based REPL, useful over SSH or in docker exec where the alternate screen breaks
• itsy -p "fix it" — one-shot, prints the answer and exits
• itsy --mcp — talk to itsy from another tool over MCP

Slash commands inside the REPL: /help, /model, /endpoint, /memory, /plan, /diff, /git, /checkpoint, /rollback, /sessions, /share, /quit. Anything tied to a feature that's disabled in your config is a no-op.

Configure

Everything lives in ~/.config/itsy/config.toml. The wizard writes a sensible default; you usually only touch this file to swap models or flip off a feature flag that's eating your context.

version = "2"

[model]
provider = "openai"
name = "Qwen3-Coder-30B-A3B-Instruct"
base_url = "http://localhost:1234/v1"
timeout = 300

[context]
detected_window = 32768
max_budget_pct = 70

[tools]
bash_timeout = 30
tool_routing = "direct"   # or "two_stage" for tiny-context models
shell_persist = true

[limits]
max_tool_calls = 50
max_tool_calls_per_turn = 250
max_output_tokens = 0     # 0 = auto (thinking_budget + 4k headroom)

[security]
allow_outside_paths = true   # /data, /tmp ok; sensitive paths still blocked

[features]
plan = true
snapshot = true
write_guard = true
clarifier = true
semantic_merge = true
error_diagnosis = true
context_retrieval = true

[tui]
auto_approve = false
classic = false
theme = "dark"

Every field can also be overridden by a CLI flag — see itsy --help. For the long tail of less common knobs (dedup similarity, code-graph DB path, snapshot dir, etc.) use --set key=value, e.g. --set dedup.similarity=0.85 or --set code_graph.disable=true.

itsy no longer reads ITSY_* env vars; everything is config + CLI. The standard cloud-provider keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY) are still consulted when cloud escalation is enabled.

The version field is honoured on load and older files get migrated forward automatically. Don't remove it.

Benchmark

Scored against terminal-bench-2 — 11 tasks, Qwen3.6-35B IQ2_XXS, n-concurrent=1. Compared against smallcode (JS agent, same model, same endpoint).

Task	smallcode 5×	itsy (current)
fix-git	80%	100%
multi-source-data-merger	100%	100%
git-leak-recovery	100%	100%
pypi-server	100%	100%
kv-store-grpc	40%	100%
prove-plus-comm	60%	100%
cobol-modernization	80%	67%
regex-log	0%	67%
overfull-hbox	0%	33%
break-filter-js-from-html	0%	0%
filter-js-from-html	0%	0%
Overall	50.9%	69.7%

Run pinned at .claude/skills/benchmark-driven-development/baselines/scoreboard-3x-20260525-streaming/ (3 attempts per task, Qwen3.6-35B IQ2_XXS, Gemma4 26B second-opinion enabled).

The two tasks both agents score 0% on (break-filter-js-from-html and filter-js-from-html) are model-capability failures at IQ2 quant — neither agent solves them regardless of framework.

License

MIT — see LICENSE.