chetter

name	chetter
description	Use Chetter to submit, track, recover, and manage remote agent tasks, sessions, triggers, definitions, fleet health, and task-created GitHub artifacts.

Chetter Remote Development Runner Fleet

Chetter is a self-hosted MCP control plane for running autonomous development agents in isolated task containers.

• Source: https://github.com/flatout-works/chetter
• Hosted MCP endpoint: https://some.ip.name/mcp, or the endpoint for your own instance

Delegate Requests Addressed to Chetter

When a user clearly addresses an instruction to "Chetter", delegate it to the runner fleet instead of performing it locally.

1. Turn the request into a self-contained task prompt with the desired outcome, repository, constraints, verification, and whether edits or GitHub artifacts are allowed.
2. Determine git_url and git_ref from the request or current repository context. Ask only when the target is genuinely ambiguous.
3. Submit with chetter_submit_task.
4. Return the task ID and, when useful, offer to track it with chetter_task_progress or chetter_task_status.

Use an agent image containing the selected harness, such as ghcr.io/flatout-works/chetter-agent-base:main or ghcr.io/flatout-works/chetter-agent:golang. Never use chetter-runner: it is the fleet daemon image and does not contain the harness CLIs needed by tasks. Omit agent_image when the Chetter instance has a suitable default.

For recurring work, create or change a Chetter cron trigger. Do not use a local machine scheduler as a substitute.

Tool Contexts

Control-plane clients

Interactive MCP clients may expose the full Chetter API, subject to token scope and optional server integrations. Major tool groups are:

Area	Tools
Tasks	chetter_submit_task, chetter_list_tasks, chetter_task_status, chetter_task_progress, chetter_task_events, chetter_task_latest_event, chetter_task_export, chetter_cancel_task, chetter_recover_task
Sessions	chetter_list_agent_sessions, chetter_agent_session_status, chetter_resume_agent_session
Fleet	chetter_runner_health, chetter_drain_runner, chetter_clear_queue
Triggers	chetter_create_trigger, chetter_update_trigger, chetter_list_triggers, chetter_delete_trigger, chetter_run_trigger, chetter_list_trigger_runs
Event callbacks	chetter_create_event_callback, chetter_update_event_callback, chetter_list_event_callbacks, chetter_delete_event_callback
Definitions	chetter_list_definition_sources, chetter_get_definition_source, chetter_sync_definition_source, chetter_sync_definitions, chetter_list_definitions, chetter_get_definition, chetter_create_definition_proposal, chetter_list_definition_proposals, chetter_get_definition_proposal
Git identities	chetter_create_git_identity, chetter_list_git_identities, chetter_update_git_identity, chetter_delete_git_identity, chetter_set_git_identity_default
Administration	chetter_create_token, chetter_list_tokens, chetter_delete_token, chetter_create_team, , ,

Arcane image-scanning tools are present only when the server has Arcane configured: chetter_arcane_scanner_status, chetter_arcane_environment_summary, chetter_arcane_list_images, chetter_arcane_image_summary, and chetter_arcane_list_vulnerabilities.

Admin-only or unavailable tools may not be visible to team-scoped clients. Use the tools actually exposed by the current MCP connection.

Agents running inside tasks

OpenCode task agents receive a restricted subset of the remote Chetter tools. They can inspect tasks, sessions, fleet health, triggers, definitions, audit/usage data, and Arcane data; submit/recover/resume tasks; create definition proposals; and create GitHub artifacts. They cannot administer teams, tokens, identities, triggers, callbacks, the queue, or runners.

Task agents also receive runner-bridge versions of these artifact tools, with task context injected automatically:

• chetter_create_issue
• chetter_issue_comment
• chetter_create_pr
• chetter_pr_review

Use these tools instead of gh issue create, gh issue comment, gh pr create, or gh pr review. They add the canonical Chetter signature and record artifact/audit metadata. Standard harness file tools handle workspace reads and edits; the runner bridge does not provide workspace_* tools.

Submit Tasks

Useful chetter_submit_task fields include:

• prompt: clear, scoped instructions
• git_url, git_ref: committed Git state to clone
• agent_image: optional agent image override, never the runner daemon image
• harness: opencode, claude-code, pi, codewhale, or codex
• agent, provider_id, model_id, variant_id
• skills, mcp_endpoints: names resolved from applicable Git-backed definitions
• env: non-secret environment values only
• timeout_sec
• session_mode: resumable, pause_reason, ttl_hours: retain an eligible gVisor task container for follow-up work

Chetter clones committed Git state. It cannot see local uncommitted changes. Put all required context in the prompt or repository, and never put secrets in prompts or env.

Track, Recover, and Resume

• Use chetter_task_progress for a concise timeline and chetter_task_events for full diagnostics.
• Use chetter_task_export to inspect a completed task transcript.
• Use chetter_recover_task after a failed task when a fresh task should receive the previous session export as workspace context.
• Use chetter_resume_agent_session only for paused or recoverable sessions. Resumable sessions require gVisor and expire according to their TTL.
• Use chetter_runner_health to distinguish task failures from stale runners or old heartbeats.

Triggers

Supported trigger types are:

• cron: requires cron_expr
• pr_review: requires repo; event can narrow webhook actions
• issue: requires repo; supports event and match_labels

Trigger run configuration can include the same harness, model, image, skills, timeout, and resumable-session fields as task submission.

Git-backed trigger definitions are the durable source of truth. Store them under the applicable scope:

• global/triggers/*.yaml
• groups/<team>/triggers/*.yaml
• repos/<owner>/<repo>/triggers/*.yaml

After merging changes, call chetter_sync_definition_source (or the legacy default-source chetter_sync_definitions) to materialize them. Trigger names must be unique across the instance; scope does not namespace the trigger table. Use direct create/update/delete tools for explicit operational changes, not as a replacement for reviewed definition files.

Definitions

Chetter materializes these Git-backed definition types: agent, skill, trigger, task_template, and mcp_endpoint. Definitions resolve by global, team, and repository scope, with the most specific applicable definition winning.

Use chetter_create_definition_proposal for durable changes from a task agent. It accepts complete replacement file contents, creates a branch and pull request, and records the proposal. Track it with chetter_list_definition_proposals and chetter_get_definition_proposal.

Agents that declare an identity require a matching server-managed Git identity. MCP endpoints and skills referenced by a task or agent must exist in an applicable scope; missing skill definitions are not useful hints and should be removed or added properly.

Event Callbacks

Event callbacks react to task event patterns such as task.completed, task.failed.*, or task.failed.model_error. Supported actions are create_task, webhook, and slack. Callback action configuration is JSON. Keep webhook credentials in server-managed configuration, never in definition files or task prompts.

Safety

• State whether edits, commits, pushes, issues, comments, reviews, or pull requests are allowed.
• Prefer a branch and reviewed PR over direct pushes to a default branch.
• Do not submit secrets in prompts or task env.
• Check for an existing issue or PR before scheduled automation creates another.
• Use realistic timeouts: about 600 seconds for quick inspection and 3600 or more for implementation.
• Treat Arcane, GitHub App, managed identities, and resumable sessions as optional deployment capabilities; handle their absence explicitly.

Model Selection

Use chetter_get_model_catalog as the authority for available providers, models, harness defaults, and variants. Do not hard-code model recommendations in automation guidance: the catalog changes independently and explicit trigger/task selections override defaults.