Kiger is Kipple Labs' agent runtime — a hardened AI agent, sandboxed, with a real console. Run it on your laptop, deploy it to your own cloud server, or have us host it — it's the same agent either way. Capability-scoped tools, multi-model routing, memory, and voice, all under your control.
Each skill declares exactly which network hosts, files, secrets, commands, and channels it may reach — six axes, default-deny. Anything it didn't declare is unreachable, even if the model asks.
Skills compiled to WebAssembly run in a wasmtime sandbox with per-skill fuel and memory limits — a hard boundary around untrusted code.
Operators set a maximum capability ceiling per workspace. An agent's config can only narrow it — never widen past what you approved.
Writing files, running commands, or spending money needs a signed approval on file first. Approvals are keyed to the exact capability surface, so a real change re-prompts.
Credentials resolve from your OS keychain or 1Password — never the agent's environment. Credential-shaped paths like .env, *.key, and id_rsa are refused even inside an allowed folder.
Every skill ships with an Ed25519-signed manifest and per-file hashes. Tamper with a skill and the runtime refuses to load it.
Develop in the terminal when you're tinkering. Operate from desktop, browser, or mobile when you're running it. Kiger is both — same agent, same configuration, wherever you are.
A fleet overview shows every agent — its status, skills, and recent activity — without opening a config file.
Approve an action from your laptop. Check the audit trail on your phone. The same console runs in the browser, as a desktop tray app, or on your phone — your agent doesn't stop when you step away from your desk.
Enable or disable a skill without redeploying. It takes effect on the agent's next loop tick — no restart, no config edit.
Actions that need sign-off surface with a live badge — grant or deny in a click. Freeze an agent's risky capabilities with a time-boxed lockdown when you need to.
Per-agent threads grouped into projects with their own instructions and files, plus a memory panel showing what the agent actually knows.
Build and script from the terminal — mint tokens, install skills, run the daemon, open a secure tunnel. Then operate from the console when it's running. Kiger is both.
Anthropic, OpenAI, Gemini and Vertex, local Ollama, and OpenAI-compatible models like DeepSeek, Qwen, and GLM — mixed however you like.
Simple, moderate, and deep tiers map to the models you pick. Each turn is classified and routed — by fast heuristics, a small cloud or local classifier, or a hybrid of both.
Every call is logged with its tier, token counts, and estimated cost — so you always know which agent is expensive and why.
A provider outage falls through to your backup chain instead of dropping the turn — with a breaker that stops hammering a failing endpoint.
A push-to-talk console with spoken replies — usable like a call, metered like everything else.
Speak in the browser with on-device speech recognition, with a Whisper fallback for browsers that don't have it.
Agents speak back with a per-voice character, and a voice overlay turns a chat thread into a back-and-forth conversation.
Speech and voice cost is written to a ledger and shown live in the chat status line — no surprise voice bill.
Voice turns default to a faster, terser model so replies come back quickly — the classifier can still promote a hard question.
Agents save what matters and recall it by keyword or by meaning, using local embeddings — with a graceful fallback when the embedding model is offline.
Each agent keeps a markdown vault with frontmatter and wikilinks it can write to and query.
People, projects, and tasks as first-class entities with relations — so "who else is on this project?" has a real answer.
Ingest documents and papers per agent and search them, so answers can be grounded in real sources.
Every agent carries its own Ed25519 keypair, and high-trust actions are wrapped in signed envelopes that break if tampered with.
When one agent invokes another, the authority chain is recorded — trace any downstream action back to the instruction that started it.
Channel, tier, model, tokens, steps, tools, and time for every turn — persisted outside the agent, so restarting it never erases the history.
Opt in and agents carry portable AXIS Protocol identity across organizations. It's off by default — you decide when cross-org identity turns on.
Two-way chat channels, connectable tools, and secure remote access — all under the same capability rules.
Allowlisted people message an agent directly in Slack, Telegram, or Discord — full routing and memory, with approval cards inline. Fail-closed: unknown senders can't get in.
Agents send email and post to webhooks through the channel router — the skill names a target, and you map where it actually goes.
Point an agent at an MCP server and its tools join the agent's catalog, authenticated per call.
Reach your agents over a Cloudflare Tunnel with a remote-scoped token — nothing exposed to the open internet.
Kiger is the runtime. AXIS gives each agent an identity. Owyhee is the governance layer its calls run through. Each layer stands on its own — together they make an agent you can hand real access to.
The agent itself — capability sandbox, model routing, memory, voice, and the operator console. Deploy it yourself, or have us run it.
Deploy an agent →An open protocol that gives an agent a signed, verifiable identity and a delegation chain back to you. Opt in when you want identity that travels across organizations.
About AXIS →Kiger agents connect through Owyhee for governed access — credentials in custody, grants scoped to exactly what you allow (say, content:comment on one site), and a receipt for every call. Revoke, and it stops.
The guided deploy path walks you through standing up a governed agent on a small cloud server — or your own machine — step by step. The Owyhee-hardened OpenClaw agent is in the catalog now; Kiger is next in.
Start deploying →A native 1-Click that stands an agent up on a droplet you own — nothing to configure. The Marketplace listing is in review.
Get notified →We run Kiger for you; you drive it from the console. While it's early, onboarding is hands-on with us — write and we'll set you up.
Request early access →Free to run. Deploy an agent yourself from the guided path, or ask us about hosted early access. Add Owyhee when you're ready to hand it real access.