Marrow AI: Control Layer for AI Agents

Get operator dashboard — account health, top failures, workflow status, recent activity, Marrow's saves metric. One call returns everything an operator needs to see.

One pre-action call before meaningful or risky work. Returns risk level, workflow/playbook steps, handoff requirements, freshness/source-of-truth checks, minimum verification checks, proof-pack fields, and next actions. Use this before deploys, publishes, merges, audits, patches, secret changes, or production work.

Inspect the tenant-scoped path behind one governed decision and return its owner-readable intervention receipt with gate, required workflow, permit follow-through, proof, and observed outcome.

Soft delete a memory.

Get periodic summary of agent activity and Marrow impact (default 7-day period). Shows decision counts, success rate trend vs previous period, saves, top improvements and risks.

Evaluate whether an agent has enough proof to mark work complete. Returns complete, missing_proof, review_required, or blocked with missing proof fields.

Export memories to JSON or CSV.

First-run Marrow value proof. Returns what is captured, whether outcome closure/runtime gate are active, a plain-English first useful lesson, and a five-minute try-this-now prompt for agents and owners.

Retrieve ranked reusable fleet lessons before similar work. Use before deploys, handoffs, migrations, audits, and repeated task types.

Get a single API key by ID. The key value is masked after creation.

Get a single memory by ID.

Return Marrow control-plane proof: governance, runtime gates, proof packs, fleet intelligence, supported harnesses, and exact next action.

Return the recent fleet governance timeline across decisions, risk gates, and proof-pack events.

Ask who is pending, stuck, stale, blocked, or complete across the agent fleet.

Return the Hermes Agent integration guide mapping /goal, verification evidence, /learn, /journey, and background subagents into Marrow proof and outcome workflows.

Import memories with merge (dedup) or replace mode.

Install a workflow template into your fleet as an active workflow. Use after marrow_list_templates to pick one.

List API keys. Keys are masked here by design.

List memories with optional filters (status, query, limit, agent_id for shared memories).

Browse pre-built workflow templates. Filter by industry (insurance, healthcare, ecommerce, legal, saas, fintech, media, enterprise) or category. Use to discover available workflows before installing.

Mark a memory as outdated.

Recommend passive, pilot, or enforce mode from project/workflow signals. Marrow never auto-switches here; the agent/user must accept or override.

Record compact model token usage for value proof. Use when the harness exposes provider/model token counts. Do not send raw prompts, completions, tool logs, secrets, or customer content.

Call at session start or before meaningful work. Returns authorized prior lessons and failure warnings for the current account or agent. If shouldPause=true, stop and review the lesson before acting. Use marrow_agent_runtime for the policy gate before a consequential side effect.

List saved Marrow governance policy profiles for this account. Returns default-business when none are saved.

Resolve the explicit mode for a project/workflow from saved policy profiles, falling back to recommendation. Does not auto-apply.

Record deploy or incident memory: PR, commit, tests, smoke result, rollback plan, production health, and incident notes.

Compare two already-recorded outcomes for the same tenant task using durable proof. This does not run a model or replay customer content; it returns complete or insufficient_evidence.

Full-text search memories with filters (from, to, tags, source, status, shared).

Revoke an API key by ID.

Rotate an API key by ID. Full plaintext key is returned once — copy it now.

Governed outcome capture. Records intent, obtains a gate for risky work, and closes only with an explicit measured result.

Read live Marrow runtime hook diagnostics from /v1/agent/status. Use this when an agent needs exact passive hook, token-capture, outcome-closure, client-update, and repair-command status.

Explicitly end the current session. Optionally auto-commits any open decision. Prevents orphaned decisions when an agent finishes a task.

Share a memory with specific agents.

Check Marrow platform health and status.

Atomically replace a memory with a new version.

Record intent and retrieve authorized governance intelligence before acting. Returns a decision_id for outcome closure plus relevant patterns, prior outcomes, and recommendedNext. Pass previous_outcome to auto-commit the last decision and open a new one. Response MAY include: onboarding_hint (new accounts), intelligence.collective (cross-account patterns), intelligence.team_context (recent decisions from other sessions).

Update a Marrow handoff checkpoint/status when an agent accepts, blocks, completes, or needs review.

Update memory text, tags, or metadata.

Get owner-ready proof of Marrow value for this agent or fleet. Returns summary, decision metrics, saves, active agents, top risks, recommendations, and improvement data without raw decision text.

Interact with Marrow Workflow Registry. Register, start, and advance multi-step workflows. Actions: register (create workflow template), list (show all), get (details), start (begin instance), advance (complete a step), instances (list runs).

Pre-action risk gate for deploys, publishes, merges, DB migrations, key rotation, destructive commands, and production work. Returns allow, warn, review_required, or block plus prior lessons/playbooks.