drive-biomni-lab

Drive Biomni Lab (hosted) via the cookie-broker connector

The cloud-hub biomni backend exposes 25 bm_* tools that drive the hosted Biomni Lab far beyond its official 10-tool claude.ai connector — effort control, skills, @-resources, live HPC/GPU introspection, usage-aware account rotation — while never handling the credential (a local broker injects the session_id cookie; you only ever see masked previews). Setup + architecture: tools/mcp/cloud-hub/biomni-mcp/README.md.

Preconditions (fail fast, don't improvise auth)

• The bm_* tools are live (hub restarted after the backend was registered).
• For each account, its daemons are running — debug Chrome (signed in) + broker + refresher (see the biomni-mcp README / autostart.py).
• If a call returns 401 / no-auth, the account's cookie is stale → tell the user to ensure that account's debug Chrome is signed in and its refresher is running. Do NOT try to obtain or paste a token yourself — auth is the broker's job.

The task loop

1. Pick the account — bm_account_list / bm_account_current; switch with bm_account_switch or rotate with bm_account_cycle. Before a heavy run, check bm_billing_balance and rotate to an account with headroom (runs consume usage).
2. Pick a project — bm_list_projects (or bm_create_project).
3. Start a task — bm_create_session(project_id, content, model_profile=…):
   • model_profile is the EFFORT tier: standard (default) | fast | premium ("max" is accepted as an alias for premium). Use premium for hard/multimodal work, fast for cheap iteration.
   • submit_initial_message=True (default) starts the run immediately.
   • To attach skills or @-resources, instead create with submit_initial_message=False, then bm_send_message(session_id, content, …) + bm_start_processing.
4. Attach skills / resources (via bm_send_message):
   • Skills → attached_skills_json = a JSON array of skills from bm_list_skills ({id,name,category,intent:"use"}); only active skills apply (bm_set_skill_status).
   • Resources (tools/databases/packages) → @name mentions inside content (e.g. "@Addgene @airr analyse …"). There is no separate field.
5. Watch it — poll bm_get_messages (reliable) for the reasoning/tool trace; bm_session_runtime(session_id, kind=…) for live compute (background-processes | gpu-sandboxes | hpc-jobs | hpc-pipelines | machines); bm_get_streaming for a best-effort live peek. Biomni tasks take 30 s–10 min — poll a few times, don't spin. bm_cancel stops a run.
6. Results — bm_list_results lists output file names/sizes; the files are downloaded from the Biomni web app (not directly). Point the user to the run link.

Cautions

• $$$ / billable: bm_create_session(submit_initial_message=True) and bm_start_processing consume the account's usage (real HPC/GPU). Confirm intent for large or repeated runs; rotate accounts with bm_billing_balance before one runs dry.
• bm_api_raw can reach mutating/DELETE routes on the user's own account — confirm before calling those.
• Trust but verify Biomni output. Biomni is an autonomous agent and has produced confident-but-wrong numbers/citations before; treat its results as hypothesis-generation and have it (or a second pass) self-audit before you rely on them.

Escape hatch

bm_api_raw(method, path, body_json, query_json) calls ANY api.phylo.bio endpoint (e.g. /v2/users/me/memory, /v2/mcp/servers) — use it for anything the 25 typed tools don't cover.