Agent runtime

Agents from folders of markdown.

Define an agent in a folder. Cap what it spends. Gate what it does. Replay every run.

project / agents / researcher

Folder

project/ ├ agents/ │ ├ researcher/ │ │ ├ agent.md │ │ └ skills/ │ └ writer/ ├ tools/ │ ├ web.fetch/ │ └ notes.write/ ├ providers/ └ roles/

agent.md

--- name: researcher cost: $5/day tools: - web.fetch - notes.write approve: notes.write output_schema: brief --- # Researcher Read the sources you're given. Write a one-page brief with citations. Ask before saving.

Trail · live

Everything an agent needs to run unattended.

The runtime around the model: limits, approvals, a record of every step, and a way back when things go wrong.

Spend caps that add up

Set a budget anywhere in the tree. Every dollar counts toward every cap above it — once, never twice.

Whole project$6.20 / $20

All models$4.85 / $10

researcher$3.90 / $5

One gate, with a human on call

Every action passes the same door. Risky ones wait for a tap on your phone — the run pauses, it doesn't fail.

researcher · now

Save brief-q3.md to notes?

DeclineApprove

Crash-safe replay

Kill it mid-run. It resumes where it stopped — no repeat charges, no asking you twice.

runcrashresume

Secrets the model never sees

Keys live in a vault, handed only to the tools that asked — scrubbed from every prompt and log.

token: sk-•••••••••• prompt: clean ✓

Every change, signed

Each file write or commit records who did it, why — and how to undo it.

$ trail explain ↳ agent:researcher ↳ revert ready

Drive it from the tools you use

It speaks MCP. Your coding agent sees every agent as a tool — long-running calls included.

Claude CodeCodexAny MCP clientCLI

Typed from end to end

Agents and tools declare what they return. Mismatched wiring is refused at boot, not found in production.

ClaudeGrok

Talk to it. It remembers.

Every agent is someone you can talk to. One thread follows you from the web to your phone to your voice.

One thread, everywhere. Chat, Telegram, voice — same conversation, same memory.
What it said is what it did. Replies, tool calls and approvals in order, exactly as they happened.
Long memory, short prompts. Recall pulls only what matters into each turn.

researchertoday

Summarize this week's support tickets.

3 tool calls

142 tickets, three themes: Android login, slow exports, billing emails. Brief is ready.

researcher · asks

Save support-w41.md?

DeclineApprove

WebPhoneTelegramVoice

Runs on your machine.

One container. Your keys, your data, your hardware. Nothing leaves unless an agent you wrote sends it — through the gate.

Three commands from clone to a live runtime.
Keys in a vault, never in the environment, never in a prompt.
The whole record in one file on your disk.

zsh

~ docker compose up
✓ built  ✓ vault mounted  ✓ trail ready
→ serving on :3000 · hello agent ready

~ claude mcp add cortex --transport http \
    localhost:3000/mcp
✓ added · agents now appear as tools

~ cortex secrets set API_KEY_ANTHROPIC
value: ••••••••  applied live, no restart

Everything is a service.

A service is a folder. An agent is a service that thinks. Every leaf is a tool — for agents, for MCP, for you.

commAgents ask, tell, delegate and escalate — to each other and to you.

knowledge:memoryRetain, recall, consolidate. Ranked by recency, meaning and use.

webFetch with retries and a trust policy.

repoRead the project: glob, grep, read.

trailAsk the record what happened, and why.

specAgents edit agents — with your confirmation.

attachmentsFiles in and out of any conversation.

your-service/Drop a folder in. It hot-reloads, typed and gated.

comm verbsasktelldelegateconfirmescalatelistennegotiate

Federation.Soon

Runtimes that call each other. Your agent uses a service someone else runs — through the same gate, the same budget, the same trail.

Private by default. Nothing is reachable from outside until you say so.
Publishing is a price. Set a fee to share a service — zero counts, but it's always a choice.
Every call metered on both sides of the wire.

Any model. Swap it in a line.

The engine doesn't care who answers. Pick a model per agent, per step — budgets and traces stay the same.

Claude Grok OpenRouter soon Local models soon

From folder to running agent.

Write

A folder with an agent.md. Prompt, tools, budget.

Run

Hot-reloads on save. No build, no deploy step.

Approve

Risky steps reach your phone. One tap to go on.

Review

Ask the trail why anything happened. Undo it if needed.

Quietly building. Shipping soon.

Stealth