Skip to content

quickstart.md

Latest commit

 

History

History
310 lines (209 loc) · 10.7 KB

quickstart.md

File metadata and controls

310 lines (209 loc) · 10.7 KB
Raw

First 5 Minutes with tinycode

A hands-on walkthrough to get productive fast. You should have tinycode built already -- see install.md if not.

1. First-run path (diagnose + connect)

There is no tinycode setup wizard. First run is two steps:

  1. tinycode doctor — diagnostic only (config, database, providers, agents, plugins). It does not configure models; it tells you what is missing.
  2. tinycode (TUI) — if no model was auto-discovered, the connect dialog opens automatically (same UI as /connect). Esc dismisses; run /connect anytime to reopen. Or set env vars / config before launch.
tinycode doctor          # diagnose; exit non-zero if critical checks fail
tinycode                 # connect dialog opens when no model is available
# From your project directory
./dist/tinycode

# Or if installed to PATH
tinycode

# Or point it at a specific project
tinycode ~/projects/my-app

You see an animated boot sequence that checks each subsystem. Green check marks mean everything is ready. If a provider check fails (red X), tinycode still launches -- you can connect a provider next.

2. Connect to a provider

If no model was auto-discovered, type /connect and press Enter. This opens a two-step dialog:

  1. Select a provider -- Ollama, LM Studio, OpenRouter, or any discovered provider
  2. Select a model -- type to filter the list (e.g., "qwen" or "llama")

For local models, Ollama is the fastest way to start:

# In another terminal
ollama pull qwen3:8b
ollama serve

Then /connect in tinycode and select Ollama.

For LM Studio, start a local server (default http://localhost:1234), then /connect and select LM Studio.

For cloud models via OpenRouter, set your API key before launching:

OPENROUTER_API_KEY=your-key tinycode

Local providers on a non-default host:

export TINYCODE_OLLAMA_HOST=http://your-host:11434   # preferred; OLLAMA_HOST is the fallback
export TINYCODE_VLLM_HOST=http://localhost:8000
export TINYCODE_LMSTUDIO_HOST=http://localhost:1234

3. Send your first prompt

Type a question at the bottom and press Enter:

what does this project do?

The model reads your project files using built-in tools (read, grep, glob, bash) and responds with a summary. Responses stream in real-time. Scroll up with PgUp or mouse wheel to review long answers. When a tool first needs approval, the UI offers Allow once / Allow always / Reject.

4. Interrupt with Escape

Press Escape to interrupt if the model is taking too long. Press Ctrl+C to clear the prompt (or quit if the prompt is already empty).

5. Undo AI file changes

If the model edited files and you want to revert:

/undo

This restores files to their state before the last AI edit. Use /redo to bring the changes back. You can also use Ctrl+X u and Ctrl+X r.

Review what changed before committing:

/diff

This opens an inline diff viewer showing all uncommitted changes in the working directory. Or use Ctrl+X d.

6. Get help

/help

Opens the command palette showing every keybinding and slash command. You can also press Ctrl+P.

For a list of all CLI subcommands:

tinycode help

7. Try @ file references

Type @ to trigger file autocomplete. A dropdown shows files in your project directory:

@src/main.go explain the entry point
  • Directories appear with a trailing / -- select one to drill into it
  • Use Up/Down arrows to navigate, Tab or Enter to confirm
  • The file contents are included as context with your prompt

Reference multiple files in one prompt:

compare @go.mod with @go.sum and check for issues

8. Use an agent

Press Tab to cycle through agents. The current agent name appears in the status bar on the right. Each agent has a different specialty:

  • build (default) -- general coding, delegates complex work
  • architect -- design decisions, system-level analysis (read-only)
  • debugger -- root-cause analysis, stack traces
  • executor -- focused implementation, smallest viable diff
  • code-reviewer -- severity-rated code review

Or invoke an agent for a single question without switching:

/ask architect should we split this into microservices?
/ask debugger why is TestAuth failing?

Press Ctrl+X a to open the full agent list.

Power features

These are powerful but require explicit consent. Prefer the path above for your first session.

/swarm

Swarm mode dispatches parallel subagents for independent tasks:

/swarm review the config, provider, and agent packages

The build agent splits the task into subtasks and runs them simultaneously. Each subagent works independently, and results are synthesized into a single report.

Consent caveat: Swarm auto-approves tool permissions so subagents can work unattended. Only use it when you are comfortable with that.

/goal

/goal all tests pass

/goal runs an autonomous multi-turn loop until a condition is met (e.g. fixing lint errors or getting a build working). It maps conditions to shell commands and can iterate (max 10 iterations).

Consent caveat: Auto-continue is off by default (defaultAutoContinueMax = 0). Loops that burn tokens and run tools without your consent only happen if you opt in. Prefer reviewing each turn until you trust the setup.

Explore keybindings

Press Ctrl+X to see the which-key panel -- a floating overlay showing all leader key follow-ups:

Navigation        Edit              Tools             Actions
  b  sidebar        e  $EDITOR        a  agent list      y  copy response
  o  session list   d  diff viewer    m  model list      x  export session
  n  new session    u  undo           t  theme picker
                    r  redo           i  MCP servers

Press any follow-up key within 2 seconds. For example, Ctrl+X d opens the diff viewer, Ctrl+X t opens the theme picker.

Press Ctrl+P to open the command palette showing all keybindings and slash commands.

Search the chat

Press Ctrl+F to open the search bar at the top of the chat viewport. Type a query to find text in the conversation:

  • Ctrl+N or Enter -- jump to next match
  • Ctrl+P -- jump to previous match
  • Escape -- close search

The search bar shows your position (e.g., "3/12") and auto-scrolls to the message containing the current match.

Workflow commands

A few commands that change how tinycode works during a session:

/effort high          # More thorough analysis, more tool calls
/effort low           # Quick concise answers
/branch experiment    # Fork this conversation to try a different approach
/rewind               # Roll back to an earlier turn if something went wrong
/context              # See what's filling up the context window

Export your session

Save the conversation for reference:

/export

This writes a Markdown file (session-<title>.md) to the working directory. Or use Ctrl+X x.

For a formatted version with syntax highlighting:

/export html

This creates a self-contained HTML file you can open in any browser.

Other useful commands

Command What it does
/thinking high Enable extended reasoning (1k/4k/16k/128k token budget)
/theme Pick a color theme with live preview
/paste-image Paste a clipboard image for vision models
/mcp Manage MCP server connections
/scoped-models Mark favorite models so only they appear in the selector
/archive Soft-delete the current session
/shell Drop into an interactive shell, return to tinycode on exit
/editor Open $EDITOR to compose a long prompt
/editor @file Open a file in $EDITOR for direct editing
/auto-approve Toggle auto-approve for tool permissions this session
/diagnostics Show diagnostics (config, providers, agents, system info)

Bundled skills

Seven built-in skills expand into the current prompt. /debug, /trace, /plan, /verify, /test, and /review instead ask the matching agent to do the job. Adding either kind is described in authoring.md.

Command What it does
/debug Ask debugger for one root cause
/verify Ask verifier for proof a change works
/trace Ask tracer to rank competing explanations
/review Ask code-reviewer to review a change
/plan Ask the plan agent for a work plan
/test Ask test-engineer to write tests
/doctor Diagnose the tinycode environment
/mcp-setup Guided MCP server configuration
/remember Triage session findings across memory surfaces
/deepinit Deep project initialization and onboarding
/incident Triage a live system failure
/change Plan one cluster or host change
/host Inspect a machine, local or over ssh

Terminal bell

tinycode rings the terminal bell when a task finishes or when a permission prompt appears. If you work in another window while the model is running, you hear the bell when it needs your attention. Configure bell behavior in your terminal settings (audible vs. visual).

Quick diagnostics

Run tinycode doctor at any time to verify all subsystems (config, database, providers, agents, plugins, skills):

tinycode doctor

Session resume

Resume where you left off without the TUI session picker:

tinycode -c                        # Continue the most recent session
tinycode -r "my feature work"     # Resume by title substring

Other run modes

tinycode run -m ollama/qwen3.5:9b "explain the main function"   # one prompt, then exit
tinycode run --multi-turn --format json -m ollama/qwen3.5:9b    # loop on stdin, NDJSON events
tinycode serve    # headless API and the thin ops console
tinycode web      # browser chat UI (no in-browser PTY; share disabled)
tinycode acp      # IDE integration over stdio

Flags, NDJSON events, and permission handling are in the user guide.

If something fails

  • Model not found. Pull it first: ollama pull <model>.
  • Cannot reach Ollama. ollama serve must be running. On another host, set TINYCODE_OLLAMA_HOST.
  • No tool calls. The warmup probe checks tool-call support. A model that fails it still answers in text. Larger models (9B and up) usually do better.

More cases: troubleshooting.md.

Next steps

  • Read the full User Guide for detailed coverage of every feature
  • See Plugin Development to build custom tool plugins
  • Check Architecture for how tinycode works internally
  • Run tinycode init for optional Red Hat plugin/role setup (models: /connect, OPENROUTER_API_KEY, or Ollama)