Give it context
State, history, files, and the right information at the right moment.
Ayjnt is an agent harness framework for building capable agents with a browser interface, a terminal interface, and a safe path to the host.
bunx ayjnt new my-agent › fix the failing bridge test ✓ 3 files updated A model can reason. A harness makes that reasoning useful: it gives the agent tools, context, memory, boundaries, observability, and a place for people to step in.
State, history, files, and the right information at the right moment.
Tools that run near the agent or cross the bridge to the host.
Isolation, explicit permissions, and human approval for meaningful actions.
A useful interface for directing, inspecting, and collaborating with the agent.
Isolation where agents reason. Native capability where work touches the machine. A typed, permission-aware bridge between them.
app.tsx browser cli.ts terminal Agents run in workerd V8 isolates, with durable state, scheduling, workflows, and agent-to-agent RPC.
Host actions run in Bun, where agents can work with files, shells, databases, processes, and local connectors.
app.tsx browser cli.ts terminal Agents reason in isolated V8 environments with durable state, scheduling, workflows, and agent-to-agent RPC.
Host actions work with files, shells, databases, processes, and local connectors.
The interface lives beside the agent, not in a separate product. Create the surface that fits the moment—or provide both.
app.tsxTyped, realtime React UI co-located with the agent it controls.
edit src/core/hostBridge.tscli.tsFull-screen, scriptable, and close to the files and tools developers already use.
you clean up the docs and examples
agent I’ll inventory the current site first.
● Search examples/** 24 results
● Read docs-astro/src/pages/index.astro
agent There are three overlapping scheduling examples. I’ll consolidate them into one practical brief agent.
›
Each example teaches a distinct harness pattern and includes a human interface you can run, inspect, and adapt.
A full-screen terminal coding agent with browser-visible sessions, transcripts, and token usage.
A Gemini-powered conversation with a fluid audio-reactive sphere that turns green while listening.
Play an agent or let two providers play each other. Bring OpenAI, Gemini, Claude, or Ollama.
Check an HTTP endpoint once, on an interval, or with cron and inspect status, latency, and response history.
A useful approval workflow that researches, drafts, pauses for a human, and publishes after approval.
A lead agent delegates investigation, fact-checking, and synthesis to focused agents and shows the handoffs.
Start locally. Add a browser UI, a terminal UI, or both. Deploy the agent runtime only when you need the cloud.
Read the quickstart →bunx ayjnt new my-agent
cd my-agent
bun install
bun run dev