The agent stack,
end to end.
HasteKit bundles every primitive you need to take an agent from prompt to production — a drop-in LLM gateway, durable runtimes, sub-agents and handoffs, skills, RAG, built-in tools, MCP, connectors, workflows, channels, triggers, self-evolving memory, and end-to-end observability. One platform. One SDK. One bill.
One endpoint. Many providers.
A drop-in replacement for the OpenAI API. Point any existing OpenAI compatible SDK at app.hastekit.dev/api/gateway/responses with your sk-uno-… virtual key and ship — no client-library rewrite, no provider lock-in. Rate-limited per project, cost-tracked per request, OpenTelemetry-traced end to end.
- OpenAI-compatible — drop in any existing SDK, no rewrite
- One virtual key routes across every supported provider
- Per-project rate limits, cost tracking, and OpenTelemetry traces
Every agent knob, in one place.
Pick the model and provider. Tune history with summarizer strategies. Force structured JSON output. Add tools and skills. Build knowledge base and memories to ground the agents in context. Cap max iterations. Snapshot every save as an immutable version and route production to any version via aliases. Run in-process for development, or on Temporal for durable execution.
- Tune model, history summarizer, and JSON schema in one place
- Immutable versions with aliases for safe production rollout
- Run in-process for dev, or durably on Temporal in production
A durable runtime, built in.
Long-running agent loops survive process crashes, node restarts, and transient tool failures. Run on Temporal or Restate for replay-safe execution — every step is checkpointed, every retry is automatic, every resume picks up exactly where the agent left off.
- Replay-safe execution on Temporal or Restate
- Every step checkpointed, every retry automatic
- Crash-resume picks up exactly where the agent left off
Agents that work together.
Call a sub-agent as a tool with isolated or shared context. Or hand off the conversation entirely — the user keeps chatting, but a specialist takes over. Compose focused teams of agents instead of overloading one monolithic prompt.
- Call a sub-agent as a tool with isolated or shared context
- Hand off the conversation to a specialist mid-thread
- Compose focused teams instead of one monolithic prompt
Reusable skills, packaged as code.
Author a SKILL.md bundle with prompts, scripts, and reference docs. Pin it to any agent and it mounts at /skills/<name>/ inside the sandbox — accessible from bash, Python, or Node. Share across agents, share across projects.
- Package prompts, scripts, and docs as a SKILL.md bundle
- Auto-mounted at /skills/<name>/ inside the sandbox
- Reusable across agents and projects, no copy-paste
Knowledge bases with first-class RAG.
Drop in your docs. Configure chunk size, overlap, and embedding model. Attach a knowledge base to one agent or share it across many. At call time, the most relevant chunks are retrieved and injected — with the full citation chain preserved in every trace.
- Drop in docs — configurable chunk size, overlap, and embedding model
- Attach a knowledge base to one agent or share it across many
- Retrieved chunks injected at call time with citations in the trace
Built-in tools, ready on day one.
Image generation, speech, transcription, a sandboxed code-execution environment, and a progress-tracker todo — all shipped out of the box. No integration matrix, no API-key shuffle. Toggle each on per agent.
- Image generation, speech, and transcription out of the box
- Sandboxed code execution for bash, Python, and Node
- Progress-tracker todos — toggle each tool on per agent
MCP & OpenAPI, lazy-loaded.
Attach an MCP server or paste in an OpenAPI spec — every operation becomes a tool. Flag any tool as deferred so it stays out of context until the agent reaches for it. Flag any as requires-approval for a human-in-the-loop gate.
- Attach an MCP server, or paste in any OpenAPI spec
- Mark tools as deferred — keep them out of context until used
- Mark tools as requires-approval for human-in-the-loop gates
First-class connectors to the tools your users live in.
Gmail, Google Calendar, Slack, Jira, GitHub — and growing. Each ships with curated actions wrapped as tools (send · list · comment · transition). Users connect their own accounts via OAuth; your agent never touches a token.
- Curated actions for Gmail, Calendar, Slack, Jira, GitHub, and more
- Per-user OAuth — your agent never touches a user token
- Each action wrapped as a typed tool, ready to call
Workflows that are deterministic.
Build a DAG from the same set of tools your agents already reach for. Run it durably on Temporal. Save it — and it becomes a tool any agent can call. Tools, all the way up.
- 30+ node types for branching, loops, delays, and code
- Durable execution on Temporal — replay-safe by default
- Save any workflow and it becomes a tool any agent can call
Meet users in Slack, Telegram, and beyond.
Bind any agent to a Slack channel or Telegram bot. Users chat naturally; the agent replies in-thread with rich formatting, attachments, and approval buttons. One agent, every chat surface.
- Bind any agent to a Slack channel or Telegram bot
- In-thread replies with rich formatting, attachments, and buttons
- One webhook to paste — no infrastructure to run
Wake agents on schedule — or on event.
A cron expression fires the agent every weekday at 9 a.m. A schedule_once wakes it at a future timestamp. A GitHub PR webhook fires it on every open. Triggers are conversation starters that don't need a human in the loop.
- Fire on cron, one-shot timestamps, or GitHub webhooks
- Schedule once or recurring — the agent wakes itself up
- Conversation starters that don’t need a human in the loop
Self-evolving memory.
Each agent maintains its own knowledge as a living wiki — semantic facts and episodic events side by side. Every N turns the agent edits the page: adds new facts, refines old ones, prunes the stale. Recall surfaces them in context, on demand.
- Semantic facts and episodic events in one living wiki
- The agent edits its own memory every N turns
- Recall surfaces what’s relevant in context, on demand
Every call traced. Every dollar attributed.
OpenTelemetry spans for every gateway call, every agent run, every tool invocation, every workflow node. Aggregate by user, organization, or agent. Watch cost roll up by provider and project, in real time.
- OpenTelemetry spans for every call, run, tool, and workflow node
- Cost rolls up by provider and project, in real time
- Filter and aggregate by user, organization, or agent
Ship your first agent in minutes.
Bring your own keys & configure agents with rich capabilities and get going.