MCP tools

Start the server with har mcp --repo <path>. Every tool accepts an optional repo; tools that act on a slot require agentId.

Discovery and setup

Tool	Inputs	Result
har_describe_project	repo	Manifest, stack hints, scripts, stages, and slot limits
har_init_harness	repo, force, smoke, profile (default | cli | ios)	Agent scaffold path (do not run har onboard). Returns the scaffold and validation result, plus a docker block (cliInstalled, daemonRunning, version, warning) — Docker is required for Mission Control and harness infra
har_maintain	repo, optional finalize, summary	Validation issues, template drift, and the maintenance bundle report; finalize: true records a completed manual adaptation in .har/manifest.json
har_add_plugin	plugin (bundled id, path, npm package, or git URL), optional force, withCi	Registered stage ids, files written, warnings, and next steps

Session lifecycle

Tool	Inputs	Result
har_preflight_environment	agentId	Readiness, blockers, warnings (including untracked worktree paths), and whether launch is safe
har_launch_environment	agentId, optional worktree, claude, resume, workUnitId, source, sourceUrl, title, parentWorkUnitId, relatedLinks	Work directory, branch, work/attempt IDs, URLs, and normalized stage result

Pass workUnitId, source, and sourceUrl when the task names a tracker issue or ticket (short repo-scoped id such as widget-123, not a provider-prefixed composite). Use relatedLinks or har_add_work_unit_link for secondary URLs (GitHub PR, mirrored issue, Bitbucket). Omit work metadata for ad-hoc work with no tracker identity. | har_add_work_unit_link | workUnitId, source, url, optional label | Append a related external link to an existing work unit | | har_recover_environment | agentId | Resumed failed or partial launch | | har_get_status | optional agentId | Structured slot status (same source as har env status --json): worktree, branch, dirty state, readiness, last run/verify. A pure read — writes no run records | | har_get_logs | agentId, optional service | Recent service output | | har_complete_environment | agentId, optional verify | Reuse last matching full validation (or re-run when verify=true), teardown, and retained branch | | har_teardown_environment | agentId, deleteBranch | Teardown result |

Every launch creates a new session from the main checkout’s current HEAD. Occupied slots always block a new launch: call har_get_status, then har_complete_environment or har_teardown_environment to free the slot, then launch again. Occupied-slot errors include the upcoming base — launch never chooses main for you. resume=true is only for a failed or starting session (or use har_recover_environment); it is not a way to replace an active session.

Execution

Tool	Inputs	Result
har_run_stage	stageId or kind, optional agentId, args	Normalized generic stage result
har_run_verification	agentId, full	Status, timing, and failed-step output (passing steps omit logs)
har_doctor	repo	Harness contract validation (same report as har env doctor --json): pass/fail checks plus findings with remedies; isError when the contract is broken

Stage kinds are setup, launch, verify, test, inspect, reset, teardown, and custom.

Evidence

Tool	Inputs	Result
har_list_artifacts	optional stageId	Files, reports, screenshots, traces, videos, and URLs
har_list_runs	optional stageId, limit	Persisted run records
har_get_run	runId	One normalized run record

Factory lines

A factory line is a multi-station program with a cumulative gate. Line stages are registered in .har/stages.json but are never added to verificationStages — har_run_verification is unaffected by installing one. Verification plugins (which do join verify) use har_add_plugin instead.

Tool	Inputs	Result
har_line_create	id, optional title, description, stations, gateStage, optInEnv, force	Scaffolds .har/lines/<id>/ (manifest, program, gate stage, README) and returns files written and next steps
har_add_line	line (local id, path, npm package, or git URL), optional force	Installed line id, stations, registered stage ids, files written, program path, adaptation prompt path, and verificationStagesUnchanged: true
har_line_status	optional line	Stations with cumulative gate progress from run records, slots in flight, and any stage that leaked onto the verify plan. A pure read — writes no run records
har_run_line_gate	station, optional line, agentId, force	Per-stage results for every gate stage tagged at that station or earlier

har_add_line refuses a verification-plugin bundle, and har_add_plugin refuses a line bundle — the two apply paths are deliberately not interchangeable.

Mission Control

Tool	Inputs	Result
har_control_up	repo, detach	Starts local Mission Control and syncs remembered repositories

Recommended agent sequence

har_describe_project
har_get_status
har_preflight_environment
har_launch_environment
har_run_verification
har_run_verification(full: true)
har_complete_environment

MCP and CLI share the same core execution service. The structured MCP result is not a separate implementation of the harness.