PlannerPlanQUEUED
Three steps: locate reconnect logic → add backoff → add regression test
PUBLIC POOL20,000 tokens a day, ready to run
Agents, skills, sandboxes,one runtimeworkflowteam
Create sessions, attach skills and run sandboxes in one workspace. Set it up once, reuse it everywhere — the left menu is the whole product.
ORCHESTRATORLIVE
LATENCY 42ms
AGENTS2 / 4
TURN 008
01 One turn
Four roles,
same room.
You describe the outcome once. Planning, retrieval, coding and verification each bring their own tools and skills — and write back into the same context.
- APlanBreak down the goal, pick skills
- BRetrieveRead the repo, docs, context
- CCodePatch files and run commands in a sandbox
- DVerifyRun tests, write results back
↓ keep scrolling to advance the turn
YOU
Add SSE reconnect to src/api and run the tests.
RetrieverRetrieveQUEUED
read sse_client.py · 2 related issues · loaded skill: async-retry
CoderCodeQUEUED
sandbox patch +12 −3 · running make test-unit
VerifierVerifyQUEUED
132 passed · 0 failed · results written to the turn
Holomer
Done, tests are green. Want me to push the diff to GitHub?
CONTEXT WINDOW8,204 / 32,768
async-retrycode-reviewrepo · web
tool calls · 14sandbox · isolatedtokens · 8,204
02 Capabilities
Six things,
one entry point.
Stop moving context between tools. Models, skills, MCP, sandboxes and repos all hang off the same runtime.
- 01Agentsroles & tools
- 02Skillssaved workflows
- 03MCPexternal world
- 04Sandboxisolated runs
- 05Exchange Spacesecret injection
- 06GitHubcode collaboration
Role, prompts, tools and skills, set once
Pick it in any session. An agent is not a prompt — it is a reusable unit of work.
PPlanner4 skills
CCoder7 tools
VVerifier2 skills
Turn a workflow into a skill
Loaded on demand, no repeated briefing.
code-reviewdebug-loopdeployresearch
Connect the outside world
Every command, isolated
Secrets injected into the runtime
Never in prompts, never in the transcript.
ANTHROPIC_API_KEY••••••••••••3f9aINJECTED
GITHUB_TOKEN••••••••••••c21bINJECTED
Work on code directly
repo · webhook · token, all in one place
03 Skill library
Workflows
as assets.
Judgement and steps, written once and attached next time.
SCROLL →0public pool · tokens / day
¥0both tiers · start now
0capability modules
0steps to first session
READY WHEN YOU ARE
Turn the next
repetitive job
into a config.
Start freeNo credit card · bring your own key · 20,000 tokens a day
04
Blog
Trade-offs, written down
Your first real task in five minutes
From sign-up to an agent running tests inside a sandbox: four clicks, one paste, plus the three most common ways it fails.
Five trade-offs from building the runtime
Why isolation lives in the database, why streaming uses SSE, why sandboxes are disposable, and where we chose slower but explainable.
Why build another agent workspace
Chat boxes are everywhere. What is missing is a runtime that can finish a task — sessions, tools, sandboxes, skills and MCP without a human ferrying context around.
Docs follow the shipped behaviour; UI changes are recorded in the changelog.