01
Skills
Markdown playbooks the agent loads on demand.Write once, share across teams. Triggers, workflows, and reference files travel together. Update a skill and every agent picks it up on the next turn.
Skills. Sandboxed tools. Git-versioned memory. The agent stack that ships code — not demos.
You've hit at least three of these. The seventh is telling your team it was worth it.
GPT-5 lands, your carefully-tuned prompts break. Skills version cleanly and stay portable.
Memory is stored in a git tree — inspectable, revertable, and never silently overwritten.
Every run is a per-session container. Files don't escape, packages don't pollute, secrets don't linger.
Configure once, activate in any chat. Containers pool per-connector across users and sessions.
Keyword → embedding → LLM tie-break. Three stages so the right playbook lands even when the user drifts topic.
Anthropic, OpenAI, Gemini, Ollama, Azure Foundry — same code, same skills, swap in Settings.
Skills load the how-to. Memory keeps the who and why. Sandbox runs the code. Connectors bring the outside world in.
Write once, share across teams. Triggers, workflows, and reference files travel together. Update a skill and every agent picks it up on the next turn.
Nothing silently forgets. Compaction proposes a diff you can review. Delete a chat and the derived facts follow, cleanly.
Files stay in /workspace, pip installs go through a house proxy, packages you approve get cached. No more agents installing left-pad-2 on your laptop.
Browse the registry, configure secrets once, activate with @mention. Containers stay warm; the handshake happens once, not per turn.
Every session gets a fresh, labelled Docker container. Files land in a bind-mounted workspace that persists across turns. Idle sandboxes reap themselves after fifteen minutes.
Every container carries agent.platform.session_id — so a startup GC can find and reap our own containers without touching anything else on the host.
/workspace/uploads, /scripts, /outputs, /tmp map to ./data/sandbox/<sid> on the host. Delete the session to nuke the folder; otherwise the next turn sees every file the previous turn wrote.
pip and npm route through a house proxy. Approved packages get cached, the top-N list turns into a bake-to-image job, and the agent stops installing left-pad-2 on Tuesday.
Sandboxes stay warm across turns. After AGENT_SANDBOX_IDLE_KILL_MINUTES of inactivity the reaper destroys the container; the workspace stays on disk.
One connector, one Docker container, one JSON-RPC pipe — shared across every user and every chat session. Warm handshakes; concurrent tool calls interleave by request id.
Browse the Docker MCP registry, or paste a mcpServers JSON block from another tool. Config is Fernet-encrypted before it lands in Postgres.
The pool is keyed on connector id, not session id. A container spins up on first activation and keeps serving every session, every user, until you delete or restart it.
MCP over stdio multiplexes by JSON-RPC id. Multiple sessions calling tools at once interleave frames cleanly — no per-call queueing, no coordination code.
Editing config kills the container so the next request spawns a fresh one. Restart honors in-flight requests unless you force it. Idle >30 min? The reaper quietly closes it.
A markdown file with a description, trigger keywords, and the workflow you want the agent to follow. Ten minutes; edits are just commits.
Pick from the MCP registry or paste a Docker JSON. Secrets encrypt on the way in; containers pool across your team.
Skills match, tools run in the sandbox, memory writes to git. Your agent hands back real files — deck.pptx, report.csv — not an apology.
Self-host in a Docker Compose stack or spin up on our infra. Either way you keep your keys, your data, and your skill library.