The coding agent that does what it says.
A sovereign coding agent with capability parity across LLMs, self-improvement (skills, persona, memory) and mandatory design-system adherence. In the terminal, on Telegram, wherever you want — without locking your company to a single model vendor.
- LicenseMIT
- LanguagePython 3.11+uv
- Versionv0.15-betapublic beta
- Tests4.081 passando
The two pains that created Okami Agent
Unreliable harness: the agent says “I'll do it” and doesn't; when pushed, it says “hold on” and stalls. You lose the afternoon checking whether the work actually happened instead of reviewing the code.
Vendor lock-in: every provider has its own format, harness and account. Switching models means redoing the integration — and paying twice when one paid agent drives another paid agent.
$ okami "refactor the auth module following the design system" ▸ skills loaded: refactor · design-system · tests ▸ model: claude-opus-4.8 → fallback gpt-5.6-luna ▸ persona: senior engineer · memory: 3 relevant decisions ◆ 12 files changed · DS tokens respected ✓ 4,081 tests passing · diff ready for review $
What Okami Agent does
Same behavior on any LLM
Immutable runtime targets, one resolver and a transport registry. LiteLLM stays available as a compatibility adapter — not as the architecture. Switching models doesn't change what the agent can do.
Tool streaming with atomic history
Structured tool-call deltas, crash-safe resume and compaction that never separates a call from its result. History stays consistent even when the session dies mid-run.
Request-scoped cancellation
Total, TTFB and idle deadlines, cooperative abort and interruptible retry. After a terminal cancellation, no recovery work keeps running behind your back.
30+ built-in skills and self-improvement
The agent evolves through its own skills, persona and memory — and keeps mandatory adherence to the project's design system instead of inventing a new look on every task.
Model control straight from chat
Secure inline model picker, /providers command, persistent per-session selection, live heartbeat and generated media delivered in the chat itself.
Per-response cost footer
Every turn shows context, tokens and time. You know exactly what that response cost — instead of finding out at the end of the month.
Spec sheet and when to use it
- License
- MIT
- Language
- Python 3.11+ (uv)
- Version
- v0.15-beta (public beta)
- Tests
- 4.081 passando
- Compatibility
- LiteLLM (adaptador)
- Interfaces
- TUI · Telegram
- Authentication
- OAuth (6 provedores) · CLI · local
Small team with several AI subscriptions
You already pay for Claude, Codex and others. Okami Agent uses the subscription you already have, with behavior parity — instead of requiring a second API bill.
Project with a strict design system
Design-system adherence is mandatory, not a suggestion. The agent stops reinventing components and starts reusing what already exists in your repository.
Operating outside the editor
Fire and follow tasks from Telegram, with a model picker and heartbeat — useful when work arrives while you're away from the terminal.
Okami Agent FAQ
Do I need an API bill on top of my subscription?
No. Okami Agent works with subscription-based auth (Codex/Claude via OAuth or CLI, local LMStudio, MiniMax/MiMo via Token Plan), with automatic fallback. Bringing your own API key is an option, not a requirement.
Is it production-ready?
It's in public beta (v0.15-beta) with 4,081 tests passing. The command and config surface may still change before GA — before exposing it publicly, run `okami policy check --strict`.
How does it compare to running the provider CLI directly?
The provider CLI gives you that provider. Okami Agent gives you consistent behavior across providers, with atomic history, deadline-based cancellation and its own skills — which an isolated CLI doesn't provide.
The products that carry the consulting
Okami Monitor
Mission control for multi-agent environments: usage and cost per period, sessions, Kanban, logs, API keys and external runtime connections — in one cockpit.
OkamiCode
Local-first desktop cockpit for AI coding CLIs: folder-bound projects, native sessions per provider, chat, email, Kanban, usage/cost analytics and durable memory — without paying a second API bill.
Okami Maturity
OWASP SAMM v2 maturity assessment: 5 functions, 15 practices, 90 questions — with scorecard, radar, prioritized roadmap and a board-ready PDF report.
Okami Tally
An iPhone panel for your AI subscription quotas — Claude, Codex, Grok, Cursor, Minimax, MiMo and any service with a JSON endpoint. In the app, in widgets, on the Lock Screen and in the Dynamic Island. No backend, no telemetry.
Okami Sentinel
Local workbench for OpenAI Codex Security scans: run, inspect, compare up to six runs and gate changes with versioned guardrails — evidence, cost and context preserved in a local workspace.
OkTally
Every AI coding subscription quota in your macOS menu bar — Claude Code, Codex, Cursor, OpenRouter and more. Colored pins, a popover with the full picture and a notification before you hit the limit. All local, no telemetry.
OkamiUNI
Email and calendar together on the Mac. Unified inbox, macOS calendars and AI to summarize threads, draft replies and turn emails into appointments. Local data, your choice of AI provider.
Want Okami Agent running in your context?
The product is open and you can run it yourself. Rolling it out in your environment, with design-system integration and a documented hand-off, is the consulting.