Cinto

Cinto is a local terminal coding-agent harness for OpenAI-compatible model servers. It gives local and open-weight models a focused workspace loop: read files, search code, propose edits, keep todos, inspect prompts, and stay inside explicit safety rails.

Português · Architecture notes

Why Cinto

Most coding agents hide the harness. Cinto keeps it visible.

• Local-first: works with LM Studio, Ollama, and other OpenAI-compatible servers.
• Two tool modes: Harmony prompt rendering for gpt-oss models and native OpenAI tool_calls for Qwen, Llama, and similar tool-capable models.
• TUI setup: first-run greeter, model/server presets, settings, chat, and path suggestions.
• Workspace tools: list_files, read_file, search, write_file, and delete_file.
• Safety controls: edit approvals, /diff, /checkpoint, protected .git/.cinto paths, and no shell execution in the current milestone.
• Persistent project context: optional AGENTS.md instructions at the workspace root.
• Context management: large tool outputs and older transcript history are compacted before they overwhelm the model context window.

Install

Source install

Use this today from a source checkout or a public GitHub repo:

cargo install --git https://github.com/Jshebb/cinto

Shell installer (Linux / macOS)

After the first v* release is tagged, GitHub Releases will publish precompiled binaries:

curl -fsSL https://raw.githubusercontent.com/Jshebb/cinto/main/install.sh | sh

The installer detects your platform, downloads cinto-<target>.tar.gz from the latest GitHub Release, verifies the SHA-256 checksum, installs cinto into ${XDG_BIN_HOME:-$HOME/.local/bin}, and adds it to your shell profile.

Override the install directory:

curl -fsSL https://raw.githubusercontent.com/Jshebb/cinto/main/install.sh \
  | CINTO_INSTALL_DIR="$HOME/bin" sh

PowerShell installer (Windows)

Invoke-WebRequest -Uri https://raw.githubusercontent.com/Jshebb/cinto/main/install.ps1 -UseBasicParsing | Invoke-Expression

This downloads and extracts the latest Windows release to ~\.local\bin and automatically adds it to your User PATH.

npm wrapper

After the npm packages are published:

npm install -g @cinto/cli
npx @cinto/cli

The npm package is a small launcher that depends on the matching optional platform package, such as @cinto/linux-x64 or @cinto/darwin-arm64.

Uninstall

cinto uninstall
cinto uninstall --purge-config

--purge-config also removes ~/.config/cinto.

Quick Start

1. Start an OpenAI-compatible local model server.
2. Run Cinto:

cinto

3. In the setup TUI, choose a preset, confirm the endpoint/model/workspace, and save.
4. Try a focused request:

Summarize this repository and list the files you inspected.

Useful setup commands:

cinto setup          # reopen the first-run setup TUI
cinto --skip-setup   # go straight to chat
cinto --print-prompt # inspect the empty rendered prompt
cinto --config ./config.toml

Model Setup

The default endpoint is the LM Studio local server base URL:

http://127.0.0.1:1234

Cinto normalizes base URLs to /v1/chat/completions. You can still pass an explicit /v1/completions endpoint for text-completion servers.

Server	Example endpoint	Recommended format	Notes
LM Studio with gpt-oss	http://127.0.0.1:1234	harmony	Use the model id shown by LM Studio.
LM Studio with Qwen/Llama	http://127.0.0.1:1234	openai-tools	Set thinking_effort = "none".
Ollama	http://127.0.0.1:11434	openai-tools	Pull a tool-capable model such as qwen2.5-coder:7b-instruct.

Example Ollama flow:

ollama pull qwen2.5-coder:7b-instruct
cinto setup

Then set:

[model]
endpoint = "http://127.0.0.1:11434"
model = "qwen2.5-coder:7b-instruct"
format = "openai-tools"
thinking_effort = "none"

TUI Workflow

Inside the TUI:

• Type a request and press Enter.
• Use /tools to inspect the tool catalog exposed to the model.
• Use /prompt to inspect the exact prompt/messages being sent.
• Use /settings, Tab, or F2 to edit model and harness settings.
• Use /diff before and after risky work.
• Use /checkpoint [label] to save a patch snapshot under .cinto/checkpoints.
• Use /git, /stage, /unstage, and /commit for explicit git actions.

Keyboard shortcuts:

Key	Action
Tab / F2	Switch between Chat and Settings
Enter	Send message or edit/apply a setting
Right	Accept the first workspace path suggestion
Up / Down	Move through settings
Space	Toggle boolean settings
s	Save settings to TOML
Ctrl-C	Quit

Workspace Instructions

Cinto reads AGENTS.md from the configured workspace root and injects it into the model-facing developer instructions. Use it for:

• project context
• coding conventions
• common commands
• anti-patterns to avoid
• release or testing expectations

The file is optional and bounded so it cannot dominate the prompt. Reopen /prompt to inspect the exact instructions being sent.

For untrusted repositories, disable it:

[harness]
load_workspace_instructions = false

Or toggle workspace instructions off in Settings.

Safety Model

Cinto is a local agent harness, not a sandbox.

• Read tools can expose workspace contents to the configured model endpoint.
• File writes and deletes require TUI approval by default.
• Workspace paths cannot escape the configured root.
• .git and .cinto internals are protected from model tool access.
• Shell execution is intentionally not exposed in the current milestone.
• Release archives are checked with SHA-256 by the installer, but they are not code-signed or notarized yet.

Avoid pointing Cinto at repositories that contain .env files, private keys, or production credentials unless you trust the model server and understand what the agent can read.

Configuration

Cinto stores config at ~/.config/cinto/config.toml unless --config is provided.

[model]
endpoint = "http://127.0.0.1:1234"
model = "openai/gpt-oss-20b"
format = "harmony"            # harmony or openai-tools
api_key_env = ""
max_tokens = 4096
temperature = 0.2
thinking_effort = "medium"    # none, low, medium, high
stream = true
stop = ["<|return|>", "<|call|>"]
request_timeout_secs = 600
context_window = 8192

[harness]
workspace = "/home/you/project"
allow_shell = false
require_edit_approval = true
max_tool_turns = 16
auto_context_compression = true
context_compression_threshold = 80
context_compression_keep_recent = 18
load_workspace_instructions = true
system_prompt = "You are Cinto, a local coding agent running in a terminal UI."
developer_prompt = "Use concise reasoning, ask before destructive actions, and prefer small verifiable edits."

When api_key_env is set, Cinto reads the secret from that environment variable and sends it as a bearer token. The TUI stores the variable name, not the secret.

Supported Tool Formats

Format	Tool-calling shape	Best fit
harmony	Tool calls embedded in Harmony-style assistant text	gpt-oss-20b, gpt-oss-120b, and Harmony-compatible servers
openai-tools	Native OpenAI-compatible tools and tool_calls fields	Qwen, Llama, Ollama, LM Studio, and other OpenAI-compatible chat servers

If the model keeps requesting tools without answering, raise max_tool_turns or ask for a narrower step. If the model returns neither text nor a tool call, Cinto shows an Empty Model Response note with the active model and format.

Development

cargo fmt --check
cargo clippy --all-targets --all-features -- -D warnings
cargo test
cargo run -- setup
cargo run

Release and packaging checks:

cargo package --allow-dirty --no-verify
sh -n install.sh
node --check npm-package/bin/cinto.js
npm pack --dry-run ./npm-package

Project Status

Cinto is early-stage and intentionally small. The current release focuses on the local agent loop, prompt/tool adapters, setup flow, install paths, and safety controls. Shell tools, provider-specific adapters, richer persistence, and signed binaries can layer on later.

License

Licensed under either of MIT or Apache-2.0 at your option.