TASKS
Natural-language intake
Create tasks from the web app or by mentioning the Lark bot. Every task keeps its creator, workspace, and notification channel attached.
Task-first agent operations
Corgi is the task orchestration layer for small teams using coding agents. Submit a task, review the proposed jobs, then let your own runtimes execute with status, artifacts, and Lark updates in one workspace.
Let admins create reusable invite links with expiry and max-use limits. Notify the team in Lark when setup is ready.
Corgi proposes agents, job types, dependencies, and deliverables before any runtime starts execution.
TASKS
Users describe the outcome in plain language. The orchestrator runs on your daemon, decomposes the work into jobs, and brings the plan back for human approval before anything executes.
TASKS
Create tasks from the web app or by mentioning the Lark bot. Every task keeps its creator, workspace, and notification channel attached.
APPROVAL
Rename jobs, adjust descriptions, swap assigned agents, and inspect what deliverables each job must produce before approving.
JOB TYPES
Global templates stay read-only. Team-specific templates live in the workspace and evolve as the orchestrator discovers new patterns.
RUNTIMES
Each runtime is a machine plus provider. Jobs prefer the agent's best executor and can fall back only when allowed.
The CLI validates proposed jobs against the Job Type schema before sending them back to Corgi for approval.
When the orchestrator proposes a new agent persona, approval creates a durable workspace agent that can be reused later.
Disconnected runtimes pause tasks instead of hiding failures. Humans can fix the machine and restart from the failed job.
Workflow
Create a workspace, invite members, and connect the places where tasks start: the web app and, when configured, Lark.
Install the CLI on a machine you control. The daemon discovers supported executors and registers them as workspace runtimes.
The orchestrator decomposes the task into jobs. Review the proposed agents, job types, dependencies, and deliverables before work begins.
Jobs run in order, artifacts accumulate, blockers ask humans for input, and completion summaries arrive in web and Lark.
Runtime and Lark control
The server never calls an LLM directly. Orchestration and job execution happen through connected runtimes, while Lark keeps the creator in the loop when tasks finish, fail, pause, or need input.
Talk to salesExecutors use the credentials already configured on the runtime machine. Corgi does not store model provider API keys.
Members can manage their own runtimes and tasks. Admins manage invitations, roles, Lark, and workspace-level configuration.
Mention the bot to create tasks. Blocked jobs ask questions in Lark and route the reply back into the job context.
Generated agents are saved to the roster with instructions, executor preferences, and chat history for future work.
FAQ
Corgi manages tasks, proposed jobs, approvals, agents, runtimes, artifacts, and notifications. The actual LLM execution happens on runtimes you connect.
Not before approval. The orchestrator proposes the execution plan, and a human approves it in the web app before jobs are dispatched.
Direct coding agents are good at one session. Corgi adds the operating layer around them: task intake, decomposition, approvals, runtime dispatch, artifacts, blockers, and team notifications.
On the runtime machines you connect. Code jobs can use per-runtime repo clones and per-task worktrees so separate tasks do not trample each other.
The job moves to blocked, asks a human-readable question, and notifies the creator in web and Lark. The reply becomes context and the job resumes.
Yes. A workspace can bind a Lark bot. Mentions create tasks, terminal states send summaries, and blocked jobs can receive replies from the original thread.