omi

A single static Go CLI for the 1min.ai REST API — with a parallel multi-model consensus panel, Unix-pipe composition, and a typed model registry with per-model defaults.

omi consensus -m mini,claude,gemini-pro --synth-model best "Should we ship today?"

Why omi?

• Multi-model consensus, in parallel. omi consensus fans the same prompt out to an N-model panel concurrently, then a synthesis model produces Consensus / Disagreements / Recommendation. Wall-clock = max(latency), not the sum.
• Composes with the rest of your shell. TTY-aware stdin/stdout — pipe git log or Fabric patterns in, pipe streamed output through jq, tee, or another omi call out.
• Typed model registry, not stringly-typed. Capability-tagged aliases (chat / code-only / vision-only) with per-model defaults that auto-apply (e.g. sonar auto-enables web search). Override via ~/.config/omi/models.json.
• Stateful conversations + REPL. Named sessions persist server-side; an interactive REPL drops in when stdin is a TTY and gracefully handles Ctrl+C cancellation of in-flight streams.
• Single static binary, ~6 ms cold start. No Python or Node runtime, no CGO. Production-safe defaults: Retry-After honored, resource caps, terminal-injection sanitizer.

Features

• Streaming chat over Server-Sent Events for sub-second first-token latency.
• Vision and document Q&A via the unified chat-with-ai endpoint with image or PDF/text/markdown/docx attachments.
• Code generation with a dedicated omi code subcommand and code-capable model aliases.
• Consensus mode via omi consensus, an omi client-side feature that asks multiple chat models and synthesizes their answers.
• Audio transcription via omi transcribe.
• Named, persistent sessions with server-side conversation lifecycle (create + delete) and an interactive REPL when stdin is a TTY.

Install

go install

go install github.com/gl0bal01/omi/cmd/omi@latest

Manual download

Grab the archive for your platform from the GitHub Releases page.

OS	Architecture	Archive
Linux	amd64	omi-<version>-linux-amd64.tar.gz
Linux	arm64	omi-<version>-linux-arm64.tar.gz
macOS	amd64	omi-<version>-darwin-amd64.tar.gz
macOS	arm64	omi-<version>-darwin-arm64.tar.gz
Windows	amd64	omi-<version>-windows-amd64.zip

Extract the archive and place omi (or omi.exe) somewhere on your PATH. Verify with omi --version.

Quickstart

omi config set api_key sk-...
omi "explain Go interfaces in one paragraph"
omi -f photo.jpg "what's in this image?"
omi consensus "Should this CLI default to streaming responses?"
omi -s research "Remember this project context"
omi -s research "Use the context from earlier"
omi --mixed "Compare this with prior context"
omi -s research --mixed "Continue with mixed history"
omi session list
omi session clear research

Command reference

Command	Purpose
omi [prompt]	Chat (streaming SSE). Empty prompt + TTY enters REPL.
omi -f <file> [prompt]	Vision (image) or document Q&A (pdf/txt/md/docx) via chat.
omi code <prompt>	Code generation via CODE_GENERATOR.
omi consensus <prompt>	Ask a 3-model panel and synthesize a consensus locally.
omi transcribe <audio>	Speech-to-text via SPEECH_TO_TEXT.
omi transcribe models	List transcribe aliases and raw speech model IDs.
omi quickstart	Print recommended presets and examples by use case.
omi upload <file>	Upload an asset and print the returned asset path.
omi session list	List saved sessions (name → UUID).
omi session clear <name>	Delete a single session server-side and locally.
omi session clear --all	Delete every saved session (prompts unless -y).
omi config set <k> <v>	Persist a config value to ~/.config/omi/config.json.
omi config get <k>	Print a config value (api_key masked unless --reveal).
omi config list	Print every config key (api_key masked).
omi models [chat|code|vision|all]	List model aliases with CAPS/DEFAULTS/NOTES metadata.
omi models explain <id>	Show one model’s capabilities, defaults, and related aliases.
omi doctor	Validate config, effective models, and endpoint compatibility.
omi completion <shell>	Emit a shell completion script (bash, zsh, fish, powershell).
omi --version	Print the binary version and exit.
omi --help	Print top-level help.

Flags (root chat command)

Flag	Description
-m, --model <id>	Chat model alias (or raw API id with stderr warning).
--task <name>	Task preset (chat, code, vision, research) for auto model choice.
-s, --session <name>	Use a named, server-persisted conversation.
-w, --web	Enable web search.
--no-web	Force-disable web search (suppresses per-model defaults).
-M, --mixed	Enable mixed-history context (compat keys: isMixed, historyMixed, history_mixed).
--no-stream	Buffer the response and print at the end.
-n, --num-sites <N>	Cap numOfSite for web search.
--max-words <N>	Cap maxWord on the response.
-f, --file <path>	Attach an image or document to the chat turn.
--api-key <key>	API key (overrides OMI_API_KEY env and config file).
--timeout <dur>	Per-request timeout for non-streaming calls (default 60s).
--verbose	Print effective endpoint/model/settings before request.
--explain-defaults	Explain why defaults were selected (flags vs model defaults).
--debug-http	Print redacted HTTP request/response summaries.

Environment variables

Variable	Effect
OMI_API_KEY	Overrides config.api_key. Required if api_key is not configured.
OMI_MODEL	Default chat model. Overridden by -m/--model.
OMI_CODE_MODEL	Default model for omi code. Overridden by --code-model.
XDG_CONFIG_HOME	Base for config directory ($XDG_CONFIG_HOME/omi/...). Falls back to ~/.config.

Models

omi ships with an embedded registry of capability-typed aliases.

omi models (no arg) shows chat-capable aliases by default. Use:

• omi models code
• omi models vision
• omi models all
• omi models --json
• omi models explain <alias-or-apiId>

Output columns are:

• ALIAS (shortcut)
• APIID (raw upstream id)
• CAPS (capability tags)
• DEFAULTS (auto-applied per-model defaults)
• NOTES (quick usage hints)

Chat aliases (default mini → gpt-4o-mini)

Alias	API id	Notes
best	gpt-5-chat-latest	Best general chat quality.
mini	gpt-4o-mini	Default chat model (cost/speed).
gpt54	gpt-5.4	Strong general reasoning.
gpt54-pro	gpt-5.4-pro	High-end reasoning.
o3	o3	Deep reasoning.
claude	claude-sonnet-4-6	Great long-form synthesis.
gemini-pro	gemini-3.1-pro-preview	Long-context + strong analysis.
deepseek	deepseek-chat	Fast, low-cost general chat.
grok	grok-4-0709
sonar	sonar-pro	Auto web search, numOfSite=5.
deep-research	sonar-deep-research	Auto web search, numOfSite=5.
qwen-max	qwen3-max

Code-only aliases

Alias	API id	Notes
codex	gpt-5.1-codex	Default code model.
codex-mini	gpt-5.1-codex-mini	Faster/cheaper code generation.
qwen-code	qwen3-coder-plus	Strong alt for code-heavy prompts.
qwen-code-fast	qwen3-coder-flash	Fast code-focused responses.
grok-code	grok-code-fast-1

Vision-only aliases

Alias	API id
qwen-vl	qwen3-vl-plus
qwen-vl-fast	qwen3-vl-flash
pixtral	pixtral-12b

A model alias not found in the registry is passed through to the API verbatim and a stderr warning is emitted. Known API IDs from the embedded registry are accepted without warning.

Refresh check against live 1min.ai catalogs:

scripts/sync-models.sh --strict

models.json override

The override file at ~/.config/omi/models.json (or $XDG_CONFIG_HOME/omi/models.json) fully replaces the embedded set. There is no merge. To add a single alias, copy the embedded JSON, append your entry, and save the result. On parse error, omi logs a warning to stderr and falls back to the embedded registry.

Sessions and ephemeral REPL

Unnamed REPL sessions are ephemeral. The conversation persists on the 1min.ai server but is not tracked locally. Pass -s <name> to make a session persistent and reusable.

When the REPL exits (exit, quit, Ctrl+D) without -s, the auto-created server-side conversation continues to exist on the 1min.ai backend but is dropped from local tracking.

Session examples:

omi -s research "Remember this project context"
omi -s research "Use the context from earlier"
omi session list
omi session clear research

Mixed-history examples:

omi --mixed "Compare this with prior context"
omi -s research --mixed "Continue with mixed history"

Consensus

omi consensus is an omi feature, not a separate 1min.ai endpoint. The CLI sends the same prompt to a 3-model panel of normal chat models, then sends those answers to a synthesis model to produce a consensus report.

Default panel:

omi consensus "Should we use SQLite or Postgres for this local CLI?"

Custom panel and synthesis model:

omi consensus -m mini,claude,gemini-pro --synth-model best "Should this release go out today?"

The output includes each model's answer followed by a synthesis with Consensus, Disagreements, and Recommendation sections.

--no-web and per-model defaults

Some models (e.g., sonar, sonar-reasoning) have webSearch=true as a per-model default. omi auto-enables web search for these and emits a stderr notice. The explicit --no-web flag suppresses this. The explicit -w/--web flag forces web search on (no double notice). User-provided -n/--num-sites and --max-words always override per-model defaults.

Precedence for the final webSearch value:

1. --no-web → false, no notice.
2. -w/--web → true, no notice.
3. Otherwise → use defaults.webSearch from the registry.

Composition with Unix pipes

omi reads piped stdin as a single-turn prompt when stdout is a TTY but stdin is not, and writes streamed model output to stdout. That makes it composable with any stdin/stdout tool.

Pipe upstream → omi:

cat notes.md | omi "summarize this in five bullet points"
git log --since='1 week' --oneline | omi -m claude "release notes draft"

Pipe omi → downstream:

omi -m best "write a structured incident report for $INCIDENT" \
  | tee report.md \
  | grep -i 'action item'

Compose with Fabric patterns — Fabric brings curated prompt patterns (extract_wisdom, summarize, improve_writing, …), omi brings the 1min.ai model panel. Either tool can drive, the other refines:

# Fabric extracts → omi rewrites for a specific audience
cat article.md | fabric --pattern extract_wisdom \
  | omi -m claude "rewrite this for a non-technical reader"

# omi drafts → Fabric polishes the prose
omi --no-stream "draft a postmortem for $INCIDENT" \
  | fabric --pattern improve_writing

Use --no-stream when piping into a tool that buffers stdin or expects a single complete document.

Resource limits and cancellation

omi bounds memory and surface area against hostile or runaway upstream responses:

Limit	Value	Behavior on overflow
--no-stream buffered output	16 MiB	Truncated; one omi: warning: line on stderr.
Piped-stdin prompt	4 MiB	Fails fast with a usage error.
Non-streaming success body	16 MiB	Read via io.LimitReader; tail bytes silently dropped.
Upstream error body	64 KiB	Read via io.LimitReader; truncated for *api.Error.
Retry-After honored on 429	up to 30 s	Capped to prevent server-side stall attacks.

omi installs a signal.NotifyContext for SIGINT/SIGTERM so Ctrl+C cancels in-flight HTTP and SSE work cleanly. Clean cancel exits 130.

Automation And Quality Gates

Stable JSON output is available for automation on:

• omi models --json: object with filter, count, and entries; each entry has alias, apiId, caps, defaults, and notes.
• omi models explain <id> --json: object with query, alias, apiId, caps, defaults, notes, and relatedAliases.
• omi transcribe models --json: object with default, aliases, models, endpoint, and type.

Local quality gates are repeatable through make:

make fmt-check
make test
make race
make lint
make security
make vuln
make build
make verify

make smoke runs the live smoke suite and requires OMI_API_KEY; set OMI_SMOKE_AUDIO to include transcription of a real audio fixture. omi doctor is the quick local compatibility audit, and omi doctor --live adds a lightweight authenticated model-catalog request.

Release steps and artifact paths are documented in docs/release-checklist.md.

Shell completion

Generate completion scripts with omi completion <shell>.

Load directly into the current shell

No install, lasts until the shell exits.

# bash
source <(omi completion bash)

# zsh
source <(omi completion zsh)
compdef _omi omi   # only needed the first time per shell

# fish
omi completion fish | source

# powershell
omi completion powershell | Out-String | Invoke-Expression

Persist across sessions

# bash (system-wide)
omi completion bash | sudo tee /etc/bash_completion.d/omi >/dev/null
# bash (user-local)
omi completion bash > ~/.local/share/bash-completion/completions/omi

# zsh — drop into a directory on $fpath, then `autoload -Uz compinit && compinit`
omi completion zsh > "${fpath[1]}/_omi"

# fish
omi completion fish > ~/.config/fish/completions/omi.fish

# powershell — append the load line to your $PROFILE
omi completion powershell | Out-String | Invoke-Expression

Completion covers subcommands, the -m/--model flag (filtered by capability for the active command), and the -s/--session flag (saved session names). Verify with omi <TAB><TAB>.

Architecture

cmd/omi/                      // entrypoint; reads -ldflags `main.version`
internal/api/                 // HTTP client, chat (SSE), assets, code, speech, conversations
internal/cmd/                 // cobra commands (root, code, transcribe, upload, session, config, models, completion, repl)
internal/config/              // ~/.config/omi/config.json load + save (0700 dir, 0600 file)
internal/models/              // embedded registry + override loader
internal/session/             // ~/.config/omi/sessions.json
internal/stream/              // 80-LoC SSE parser

The CLI uses Cobra, stdlib net/http, and a hand-rolled SSE parser. No CGO. No vendored deps beyond cobra and golang.org/x/term. All chat traffic — including vision and document Q&A — flows through POST /api/chat-with-ai with the UNIFY_CHAT_WITH_AI type. CODE_GENERATOR and SPEECH_TO_TEXT use POST /api/features.

License

MIT