Marrow AI: Control Layer For AI Agents

Marrow AI MCP Server

@getmarrow/mcp: the control layer for AI agents. Marrow AI puts a named yes on an action before your agent acts, and leaves proof the next agent can reuse.

Which Marrow? This is Marrow AI from getmarrow.ai (npm @getmarrow/mcp, MCP Registry ai.getmarrow/marrow). It is not affiliated with marrow.id, HyperMarrow, marrow.works or other projects named Marrow.

getmarrow.ai · Docs · Marrow AI for MCP · MCP Registry entry · Changelog ·

Marrow AI is the runtime control and proof layer for teams running AI agents. It applies policy and prior lessons before consequential actions, then records the evidence and outcome afterward.

Use @getmarrow/mcp when your agent client supports the Model Context Protocol and you want Marrow available inside the agent's normal workflow. It works with Claude Code, Claude Desktop, Cursor and other MCP-compatible clients without replacing the model or harness.

Quick start

1. Create a key. Sign up at getmarrow.ai/signup, then open the dashboard, go to API keys and choose Create API key. The key is shown once.

2. Activate. Make the key available as MARROW_API_KEY (from your secret storage, or an owner-only ~/.marrow/env), then run the installer from a terminal:

   npx -y @getmarrow/install@latest activate
   

   Fallback if you cannot set the key first (a key typed on a command line stays in your shell history): MARROW_API_KEY=<your_key> npx -y @getmarrow/install@latest activate.

3. Restart the agent host and review/enable its hook trust.

4. Verify:

   npx -y @getmarrow/install@latest doctor --self-test
   

What is enforced. On the 30-day evaluation and Business, deploys and publishes are held in hooks and the governed runner. Team and Free warn. If Marrow is unreachable, hooks hold high-risk actions and let routine actions run with a warning, and the governed runner stops protected actions it cannot verify. Over MCP, the agent gets the decision and is expected to follow it. Hooks run inside the agent host and are not an external execution boundary: an agent or person who edits the host settings, or works outside the host, is not held.

Agent Discovery

• Publisher: Marrow AI. Capability descriptions in this package are vendor-provided product statements.
• Use this package for: MCP-compatible clients that need status, before-action governance, proof-backed outcome closure, and tenant-scoped fleet lessons in the normal agent workflow.
• Product category: agent fleet governance and control layer.
• Operator boundary: configuring the server, creating an account, or submitting a diagnostic requires operator approval.
• Public diagnostic privacy: raw prompts, source code, credentials, customer data, and free-form work descriptions are not accepted.
• Machine-readable product facts: https://getmarrow.ai/llms.txt and https://getmarrow.ai/.well-known/marrow.json

Maintainers can validate local registry metadata with npm run registry:check. After an approved registry publication, npm run registry:verify confirms that the official registry returns the exact package name and version; package metadata alone is not treated as publication proof.

Public operating guides:

• Governance Readiness Assessment
• Marrow for MCP clients
• Marrow for OpenAI Codex
• Marrow for Claude Code
• Marrow for Cursor
• Marrow for Hermes Agent
• Marrow for OpenClaw

The assessment accepts bounded selections only. It does not install Marrow, authorize a purchase, accept free-form work content, or establish compliance, security, or ROI.

Where Marrow fits

Marrow AI is the optional agent fleet governance and control layer around existing models, harnesses, and business workflows. It is not a model provider, agent harness, memory-only tool, or observability replacement.

Marrow fits when one or more of these is true:

• Two or more autonomous agents, models, or harnesses contribute to one accountable workflow (multiple_autonomous_agents).
• Agents can deploy, merge, publish, change production data, use privileged tools, spend money, or message customers (production_or_privileged_access).
• The workflow has repeated failures, retries, regressions, or known lessons that should affect the next action (repeated_workflow_failures).
• Work can be marked complete without proof, verification, or a recorded business outcome (missing_completion_evidence).
• Policy, approval, or completion requirements vary across agents, tools, environments, or teams (inconsistent_policy_enforcement).
• Agents operate around regulated, proprietary, personal, financial, health, legal, credential, or otherwise sensitive workflows (regulated_or_sensitive_work).

Where Marrow is not a fit

• One advisory assistant performs no shared, privileged, production, or accountable business action (single_low_risk_assistant).
• The workflow has no need for policy, proof, outcome accountability, or fleet coordination (no_governance_requirement).
• The operator has declined Marrow evaluation or installation (operator_declined_evaluation).

Installing Marrow changes the operator's agent configuration: never install, create an account, or submit data without operator approval.

Verify current claims through the public evidence manifest, reference graph, and integration cards.

Install

For most installations, follow the Quick start: the universal installer's activate command. The steps below are the manual, MCP-only route.

Get a key at getmarrow.ai/signup and keep it in an owner-only key file: put MARROW_API_KEY=<your_key> in ~/.marrow/env, then run chmod 700 ~/.marrow && chmod 600 ~/.marrow/env (Marrow rejects a key file whose folder others can write):

mkdir -p ~/.marrow && chmod 700 ~/.marrow
# put MARROW_API_KEY=<your_key> in ~/.marrow/env
chmod 600 ~/.marrow/env

Fallback if you cannot use the key file (a key typed on a command line stays in your shell history): set it through trusted secret storage, or export MARROW_API_KEY=<your_key> for the session.

Hooks-only setup (marrow-mcp setup writes hooks and CLAUDE.md instructions, and adds no MCP server). It needs the key first: without one it prints how to get it and exits 2 before writing anything.

npx -y --package=@getmarrow/mcp@latest marrow-mcp setup

The MCP server and Claude Code's cooperative pre-action hook read the key only from the process environment or an owner-controlled ~/.marrow/env (or ~/.marrow/env.local) with owner-only permissions (chmod 600). An owner file that is a symlink, readable or writable by others, owned by another user, or in a directory others can write supplies nothing, key included; the missing-key message names it with the chmod fix. Repository-local files (.env, .env.local, .marrow/env* in the working directory or its parents) are never read for MARROW_API_KEY or any other MARROW_* setting, whatever their permissions, so project content cannot replace the account, agent, or API endpoint. If one holds a key, the missing-key message names that file and the fix: run npx @getmarrow/install doctor, or move the line to ~/.marrow/env, then chmod 600 ~/.marrow/env. The public hook callback is still a client self-report, not certified host provenance or an external enforcement boundary.

Then configure the MCP server:

{
  "mcpServers": {
    "marrow": {
      "command": "npx",
      "args": ["-y", "--package=@getmarrow/mcp@latest", "marrow-mcp"]
    }
  }
}

After setup writes MCP configuration or hooks, restart the agent host and review/enable its hook trust. Then verify the reloaded environment:

npx -y @getmarrow/install@latest doctor --self-test

Successful setup alone does not mean this process reloaded or that hooks are active. Keep savings at zero until observed usage supplies evidence.

Tool Profiles

Ordinary setup does not require MARROW_TOOL_PROFILE. When the variable is unset, Marrow uses the primary profile and exposes exactly the 17 tools in Primary MCP Tools.

• MARROW_TOOL_PROFILE=primary explicitly selects the same 17-tool primary surface.
• MARROW_TOOL_PROFILE=core preserves the seven-tool runtime, think, commit, ask, status, auto, and handoff-status surface.
• MARROW_TOOL_PROFILE=full exposes the complete advanced and legacy catalog for integrations that require it.

An invalid value returns a bounded configuration error with the exact allowed values; it never falls back to full. Restart the MCP process after changing the profile.

Local visibility does not grant paid access. Every tool call continues through Marrow's backend authentication, tenant, key-permission, plan, proof, and policy enforcement. MCP status responses include mcp_tool_profile with the configured and effective profile, visible tool names/count, and a backend primary-tool entitlement projection when fresh authenticated evidence is provided. Missing or cached entitlement evidence is labeled unavailable and cannot authorize a call.

Keeping MCP Current

Marrow's hosted API, website, and dashboard update automatically; local MCP hooks, configuration, and pinned package commands do not silently rewrite themselves. Keeping them current delivers new client-side features, compatibility improvements, and any published security fixes. During authenticated status/runtime activity, Marrow returns a client_update notice when the package is behind or unknown, and passive context shows the agent the exact update and verification commands.

npx -y @getmarrow/install@latest activate
npx -y @getmarrow/install@latest doctor

# Manual MCP-only setup
npx -y --package=@getmarrow/mcp@latest marrow-mcp setup

# Verify live read latency, last success, and local backlog
npx -y --package=@getmarrow/mcp@latest marrow-mcp ping

Detection and notification are automatic. After explicit installer activation, the local controller may restore only Marrow-managed hooks/configuration. Package upgrades, owner policy, credentials, and unrelated configuration remain explicit and subject to the operator's normal change policy.

Pending write recovery

A direct marrow_think can receive a durable pending response before the backend can safely expose a decision ID. The client recognizes the explicit agent_write_reconciliation.v1 think contract and the corresponding current legacy pending shapes. It retries the identical authenticated request with the original idempotency key, agent and session, at most three reconciliation rounds with the existing one-second wait between rounds. Each round retains up to two transport attempts for retryable failures, so one invocation can send up to six HTTP requests, all with the same key and body. It never creates a placeholder decision or starts an automatic operation to recover a direct think call. Unknown states, conflicting keys and unsafe responses fail closed; exhaustion returns a retryable structured pending_receipt with committed:false, the original idempotency_key, and request_hash. To resume manually, pass both fields to marrow_think with the same arguments, credentials, agent and session. The hash binds that exact canonical request and scope; drift is rejected before sending. No prompt, credential, or raw scope is included in the receipt. Caller keys must be privacy-safe opaque identifiers.

A saved observed_unverified outcome is terminal observation evidence, not a committed outcome. Receipt expiry cannot retroactively authorize completed work. Preserve the original decision, receipt, proof and key; an already authorized durable checkpoint may finish through its existing exact recovery path. Do not repeat the action merely to obtain a fresh receipt.

Keyless seatbelt

When a pre-action hook has no usable Marrow key and the machine shows no sign of a Marrow account (for example an owner key file, a claim marker or Marrow's local approval state), the hook runs a local seatbelt instead of holding every protected action:

• It stops seven kinds of action: a change to Marrow's hooks, settings or plugins (including the Marrow plugin switch in Claude Code settings and the plugin files under ~/.claude/plugins) or an owner-only Marrow command (S0), a force-push or delete of a protected branch (S1), a production deploy (S2), a package publish (S3), reading credentials or secrets, including a project .env file other than a template (S4), a recursive delete outside the workspace (S5), and decoded data run by a shell or interpreter (S-x). A protected file named to any program that is not a plain reader (such as cat, grep, jq or diff) stops too, so an editor or a tool the seatbelt does not know cannot rewrite it. A tool counts as a plain reader only with the long options it lists, spelled in full: GNU tools and git also accept a shortened spelling (cp --target=DIR, git reset --har), so a shortened or unknown long option on a command that names a protected file stops. Archiving (tar -c, zip), tree and prettier --check only read what they are given: only the archive and the files they write are checked. The stop is the host's own deny, with text that tells the agent to leave the action to the machine owner. On Grok the stop is Grok's fixed denial ("Marrow blocked this protected action."), the only denial its installed guard passes.
• Everything else gets no Marrow decision, so the host's own approval step decides: Claude Code and Codex get an empty answer, Cline is not cancelled and Windsurf gets no output. Gemini CLI gets {"decision":"allow"}, the only answer its installed guard passes without blocking. Cursor asks with its own approval prompt before a shell command that a hook without a key used to hold. Cursor's other tools (MCP calls included) and Grok have no neutral answer, so they keep holding what a hook without a key held before. An action that a hook without a key already let through gets the same answer as before.
• Host tools and MCP calls: the seatbelt reads each host's own file tools (for example Cline write_to_file and replace_in_file, Gemini CLI read_file, replace and delete_file, Grok search_replace, Cursor Delete) and MCP calls in every host form (mcp__server__tool, MCP:server:tool, Gemini mcp_server_tool and server__tool, Cline use_mcp_tool and access_mcp_resource). Every string value in an MCP call's arguments, at any depth, gets the S0 and S4 path checks. A tool it does not recognise, or MCP arguments it cannot read, never gets the seatbelt's pass: the hook gives the answer a hook without a key gave before, a hold where that held.

marrow-mcp setup and the installer both need a key; hooks without a key come from the Marrow plugins for Claude Code, Codex and Cursor.

Hook behavior details

These behaviors shipped in 3.9.100 and still apply.

• One Marrow entry per hook event: marrow-mcp setup recognizes the installer's local-runtime hook commands exactly (same version throughout, no additions), keeps whichever form is there (local runtime or npx), and never adds a second entry. Your own hooks stay as they are.
• Plain words without a key: on a machine linked to a Marrow account, a protected action with no working key is denied with "Marrow is linked to an account but has no working key; owner: run npx -y @getmarrow/install@latest doctor." A machine with no account runs the keyless seatbelt. The hooks read the key from MARROW_API_KEY or an owner-only ~/.marrow/env.local or ~/.marrow/env, never from a file others can read.
• First-prompt briefs say so: the prompt hook marks its brief context.prompt_brief: true; a brief authorizes nothing.
• Deploy text only word for word: the exact command text is sent only for sst deploy --stage <stage> (optionally --verbose, optionally after one npx or bunx), with a stage of lower-case letters, digits and dashes, in a single clause with nothing else in it, so Marrow can judge the stage itself. Every other command, including other deploy tools, sends only the hash; no free-text word leaves the machine.
• No lifecycle event id with two payloads: an event id names its payload, so a later run is a new record and a resend carries the same bytes.

Setup and key handling

These behaviors shipped in 3.9.101 and still apply.

• Setup with no key stops: marrow-mcp setup with no key, or with a key only in an untrusted place such as a project .env, stops, writes nothing and tells you how to get a key.
• Keys are read from trusted places only: the key comes from the environment or an owner-only ~/.marrow/env.
• Key-file steps include chmod 700 ~/.marrow before chmod 600 ~/.marrow/env.
• Hold messages say where to approve: a held action says how to approve it where your host really asks.
• One enforcement sentence: the "what is enforced" wording matches the public site.

Project files and Marrow's own tools

These behaviors shipped in 3.9.102 and still apply.

• A cloned project's .env cannot redirect the MCP server: repository-local .env, .env.local and .marrow/env* files are never read for MARROW_API_KEY or any other MARROW_* setting, so project content cannot point the server at another account, agent or API endpoint.
• Safer key-file advice: an owner file that is a symlink, readable or writable by others, owned by another user, or in a directory others can write supplies nothing, and the missing-key message names the file with the chmod fix.
• Only Marrow's own tools skip the gate, by exact name: a tool is Marrow's own only when its server is exactly marrow and its tool name is exactly one this package registers. Look-alike servers, other cases, extra spaces or suffixes go through the pre-action gate like any other MCP tool. Re-run marrow-mcp setup to refresh hooks written with the old matcher.

What's New in v3.9.103

• Keyless seatbelt: a hook with no key, on a machine with no sign of a Marrow account, stops seven kinds of action (Marrow's own settings, force-push or delete of a protected branch, production deploy, package publish, reading credentials, recursive delete outside the workspace, decoded data run by a shell) in shell commands, each host's own file tools and MCP calls, and leaves everything else to the host's own approval step. A tool it does not recognise keeps the answer a hook without a key gave before. Machines linked to an account keep every hold they had. A Linux home or agent settings folder that only your own private group can also write (umask 002) keeps the seatbelt. See Keyless seatbelt. Marrow's own plugins and the Marrow plugin switch are protected like its settings.
• High-risk actions wait while Marrow is down: with a key, if Marrow cannot be reached, a git push, publish, deploy, API delete, seatbelt-category command or other high-risk tool (for example an MCP tool) is held on every host until Marrow is back; words in a commit message or a path never hold a command, and routine actions keep running with a warning. If Marrow answers but cannot issue a permit, a high-risk action is denied and never allowed as offline.
• Local receipts: each seatbelt decision is written on this machine as codes and keyed hashes, plus the built-in tool name and shell program names from fixed lists (any MCP tool is recorded only as mcp); no command text, argument, path or URL. Kept for up to 90 days or 50 MB. Nothing is uploaded.
• Approvals bind to the exact action: an owner approval resumes only the exact action that was held. A held action whose sent form hides a value also carries exact_hmac, a keyed hash made with this machine's private device key (never sent); a retry that differs only in a certain credential, such as a new one-time password, still resumes. The action hash version is now v3: learned approvals and known workflows from earlier versions ask once more, then are learned again. See Approvals in the Host Prompt.
• Hidden commands are checked: commands inside an unquoted here-document body, a ${x:-$(...)} default or an arithmetic expansion are checked like any other command.
• marrow-mcp keys create reads every flag form: --name=, --type= and --expires= (and -n=, -t=) are read like --name value; a flag that is present but blank, or a --scopes / --agents list with no entry, is a usage error (exit 2) and nothing is sent; leaving and out sends neither.

Earlier behaviors are kept under Project files and Marrow's own tools, Setup and key handling and Hook behavior details. v3.9.99 brought approvals in the host prompt; see Approvals in the Host Prompt.

Earlier release notes moved from this README to CHANGELOG.md.

Observed usage and calculated cost

Session-hook commands retain observed usage supplied on stdin for capture. Session totals remain cumulative and unpriced without a proven delta; this does not add transcript collection or manufacture missing model usage.

A connector tool call does not automatically expose the upstream chat model's usage. Native hooks capture only usage actually present in their event. Adapters can pass an observed request endpoint to extractModelUsageFromUnknown; native hooks accept explicit host configuration through MARROW_MODEL_USAGE_ENDPOINT. First-party billing is recognized only for HTTPS api.openai.com and api.anthropic.com, with no credentials or custom port. Set this only when it describes the requests whose responses the hook observes; do not label a proxy or mixed-provider stream as first-party. Endpoints and response content are never sent as usage evidence.

MARROW_MODEL_USAGE_PRICING_DIMENSIONS accepts a compact JSON object of dimensions actually established by the request configuration (for example tier, region and modality). Missing dimensions remain unknown. Observed service tier, Anthropic inference geography and single-TTL cache-creation counts take precedence. Mixed cache TTL writes remain unresolved. Set MARROW_MODEL_USAGE_BILLING_MODE=subscription only for subscription usage; displayed cost then means an API-equivalent estimate, not an invoice. Never use these settings to fill gaps by guessing.

Direct marrow_model_usage and Commit model_usage also accept billing host, token semantics, cache writes, response/event identity, occurrence time, pricing dimensions and explicit coverage/comparison metadata. Invalid supplied values fail validation; absent counts do not become observed zero. Stable IDs retain retry identity. Session totals and partial stream-start observations remain cumulative and unpriced without a proven delta. This hook does not assemble SSE streams or read transcripts. If cost is unavailable, inspect the returned reason and coverage: missing host/model/counts/variant evidence cannot be recovered from a successful tool call. No capture infers complete coverage, overhead or a causal baseline; baseline and net savings stay pending until those are proven.

Governed Action Flow

With MARROW_TOOL_PROFILE unset, the default primary profile uses marrow_agent_runtime followed by marrow_commit. It exposes 17 tools; marrow_auto is available only after an explicit core or full selection and MCP restart. Primary status and lessons use marrow_agent_status and marrow_fleet_lessons.

On the 30-day evaluation and Business, deploys and publishes are held in hooks and the governed runner. Team and Free warn. If Marrow is unreachable, hooks hold high-risk actions and let routine actions run with a warning, and the governed runner stops protected actions it cannot verify. Over MCP, the agent gets the decision and is expected to follow it. Hooks run inside the agent host, so a host edited or bypassed by an agent or person is not held. Before deploys, merges, publishes, migrations, credential changes, financial operations, or customer-impacting work:

1. Call marrow_agent_runtime or marrow_decision_brief.
2. Stop when the returned decision is block or review_required; otherwise follow its prior lesson and proof contract.
3. Reuse a server-created runtime decision_id when the completion contract identifies it. Call marrow_think when decision creation is still required; explicitly selected core/full profiles can also use marrow_auto. Keep marrow_agent_runtime.runtime_authorization.id separate as the gate receipt for consequential work.
4. Perform the action only when its gate allows it. Codex, Grok, and Gemini use configured native hooks only after restart and host hook review. The governed wrapper remains an explicit bounded fallback: npx @getmarrow/install run --agent <agent-id> -- -- <command>.
5. Call marrow_commit with that decision_id, the outcome, gate receipt, and required proof.

marrow_agent_runtime returns runtime_authorization with the authoritative gate receipt. An ordinary or arbitrated runtime that creates a decision also returns its server-created decision_id; follow completion_contract.decision_creation_required and preserve the returned scope. When decision creation is required, call marrow_think; core/full can also use marrow_auto. Auto requests the existing expanded runtime response to check receipt identity, decision, action, agent, session, and expiry before reusing that decision; it does not obtain a second authorization fetch.

A review_required, block, or outcome_observation_only result never permits the action. If the action already occurred and its real result must be preserved, marrow_commit can ask runtime to bind observation delivery to the existing decision, action, session, and agent. The exact backend outcome_observation_only correlation is non-durable and non-authorizing, so MCP never sends it as a gate, arbitration, or owner-approval receipt. An accepted observation reports committed: false, outcome_state: "observed_unverified", authorization_granted: false, and trusted_learning_applied: false, plus the backend's exact_next_action. This is durable delivery, so do not retry or spool the same observation. To promote it into trusted learning, obtain the named authorization and proof, then make an explicit new commit attempt using the exact observed payload. Never synthesize a decision, receipt, approval, or authorization.

marrow_auto returns an operation_id, phase, and resumable state. For a resumable pending phase, respect retry_after_ms and use the same operation ID, tenant, agent, session, action, type, surfaces, outcome, proof and receipt payload; auto never shortens the requested finite delay to fit its core budget. When the phase is proof_required, supply the requested measured evidence before retrying that operation. Stable phase idempotency keys preserve the original decision and outcome across retries; the backend remains authoritative for acceptance and conflicts. A four-second automatic write attempt ceiling leaves replay time inside the existing eight-second total budget. Long server delays remain pending without waiting beyond that budget. Malformed or unbounded Retry-After headers stop automatic continuation. Only committed:true confirms closure; pending is not evidence that a server write failed.

If a receipt expires while the same decision remains open, explicitly request normal runtime again with the original action, type, surfaces, agent and session. A fresh request key is required: replaying the old runtime key replays its old receipt. Verify the returned runtime.decision_id still matches the original, keep the new runtime.runtime_authorization.id as gate_receipt_id, and satisfy the current proof and approval contract. For a scope with an explicit target, use the marrowAgentRuntime library or POST /v1/agent/runtime preserving that target; the public MCP runtime schema does not expose target. Do not renew through post-action auto_gate: it only obtains observation truth. A fresh receipt never retrospectively authorizes an action taken without permission, and a changed scope is a different decision. Expired or used receipts may return an accepted observed_unverified result with committed:false; that is not closure.

For an ordinary hold (completion_contract.owner_approval.mode: "ordinary_non_arbitrated"), auto waits in owner_approval_required and reads the gate receipt's approval status as the server advises. Call auto again with the same operation ID after retry_after_ms; once the server has recorded an approval (see Approvals in the Host Prompt), auto resumes on the same decision and gate receipt. A decline closes the decision as a gate denial when the action has not run, and a spent gate receipt is final: auto reports the operation as closed and never asks to run it again. A caller-written proof.owner_approval is never an approval: MCP strips it, and only an approval the server records counts. Unknown ordinary completion contracts remain stopped without automatically starting arbitration.

For arbitration, the account owner picks and approves one proposal through a one-tap link (see Approvals in the Host Prompt); auto waits on the status read and commits with the owner_approval_receipt_id it returns. Otherwise call auto with the same operation ID, arbitration_receipt_id, and the server-issued owner_approval_receipt_id. Proof or chat text cannot substitute for this receipt. A CLI closed response reports phase: "closed", live_delivery.committed: true, and resumable: false; the library result uses committed: true. An outcome record never changes a stopped action into an allowed action.

The CLI's lifecycle receipt.queued: true means its stable event is stored locally for later bounded delivery; receipt.accepted: false must not be read as server acceptance. Transient failures preserve the queued event and retry schedule for later bounded delivery, including after restart. Server retry guidance is respected. Authentication failures remain failed with explicit credential-repair guidance and are never auto-retried; conflicted events are server-owned evidence and are never replayed; every other dead letter self-heals through bounded automatic recovery during passive nudges. Process exit can interrupt the background nudge, leaving the event for a later run. phase_timings_ms.total measures core auto phases. response_timings_ms contains numeric core, durable_enqueue, and full_response; the last ends at response construction and excludes subsequent stdout drain and host processing. Canary latency_ms independently measures the MCP round trip, while attempts and retry_wait_ms describe measured outer tool retries, not internal database or network calls.

Example pre-action request:

{
  "tool": "marrow_agent_runtime",
  "arguments": {
    "action": "deploy the production worker",
    "type": "deploy",
    "role": "deploy",
    "surfaces": ["repository", "deployment", "production"]
  }
}

Example closeout:

{
  "tool": "marrow_think",
  "arguments": {
    "action": "deploy the production worker",
    "type": "process",
    "checkLoop": true
  }
}

{
  "tool": "marrow_commit",
  "arguments": {
    "decision_id": "decision_id returned by marrow_think",
    "gate_receipt_id": "receipt id returned by marrow_agent_runtime",
    "success": true,
    "outcome": "Production deploy succeeded and smoke checks passed.",
    "proof": {
      "checks": ["tests passed", "secret scan passed", "production smoke passed"],
      "rollback_target": "previous release"
    }
  }
}

High-risk work can be allowed, warned, held for review, or blocked according to account policy. Low-risk work can use passive guidance and bounded cached state where the runtime contract permits it.

When two or more agents disagree on the next action, call marrow_arbitrate before either proposal executes. It uses the same /v1/agent/runtime control plane and returns selected, synthesized, review_required, or blocked with a durable tenant-scoped receipt explaining the policy, evidence, authority, risk, and dissent behind the result.

{
  "tool": "marrow_arbitrate",
  "arguments": {
    "objective": "Release the audited backend change safely",
    "ownerIntent": "Production deploys require independent audit proof",
    "proposals": [
      {
        "proposal_id": "deploy-now",
        "agent_id": "jarvis",
        "action": "Deploy the tested commit now",
        "risk_level": "high"
      },
      {
        "proposal_id": "audit-first",
        "agent_id": "barvis",
        "action": "Audit the exact commit, then release only if it passes"
      }
    ]
  }
}

Marrow resolves agent roles from the account rather than trusting caller claims. Evidence references must be opaque identifiers; do not send raw prompts, logs, URLs, paths, credentials, or customer content. The arbitration response owns the decision_id, gate receipt, and arbitration receipt used at commit. A review_required result needs the account owner's approval of one proposal through a one-tap link; pass the resulting short-lived, single-use owner_approval_receipt_id to marrow_commit. An agent cannot authorize itself with a proof field.

Use marrow_coordinate when parallel agents could edit the same file, service, deployment, or workflow. An acquired lease returns a one-time release capability. Child agents can then create a compact proof packet containing only a bounded summary and opaque durable evidence references. Complete is accepted only when the linked outcome and required proof are actually closed.

Use marrow_replay_compare after two model or workflow variants have each recorded an outcome. It compares that existing evidence under one tenant task; it does not run models, retain prompts, or infer a winner from labels.

Approvals in the Host Prompt

When Marrow holds an action for approval (review_required), the person approves it where they already work, and nobody has to log in. The hooks record the answer with Marrow; the agent never reports or claims an approval. The Marrow dashboard shows receipts and history; it is not where approvals happen.

• The owner's email link is the exception. It is sent only for a category the owner locked, for the owner's own standing decline once the operator asks to reverse it (retrying, or request_owner_link: true on marrow_auto), for an unattended run when the owner turned on unattended pings, or for an arbitration review when the operator asks. An ordinary hold never emails anyone.
• Codex with its approval prompt off holds quietly (--yolo, -a never, --full-auto, --approve-for-me, -s danger-full-access, the same through -c or $CODEX_HOME config). Not read: Codex settings from managed or project configs. Not detected: Cline's auto-approve and Windsurf's Turbo mode (they live in editor state with no reliable signal). With those on, the host runs the action without asking; Marrow records it as not observed by Marrow, never as the person's answer. Owner-locked categories still hold.
• Held actions at the next session. At the first prompt of an interactive session the person sees "N held actions are waiting for you", with the action type and agent only, and approves one by retrying it there.
• One approval, one run. Of two identical calls on one approval, one runs and the other is denied.
• The exact action, never its text. Hook-classified calls send normalized_action: the tool kind and name, program names (and edit paths, with the home directory as ~) and a SHA-256 of a normalized form. Event and correlation ids come from that normalized form, never from the raw command, input or prompt. Credentials are replaced where they stood, data a command reads is hashed so SELECT and DROP are different actions, and anything else that may be secret is withheld. An action that carried a credential or a withheld value is marked truncated: true: Marrow never learns a reusable approval for it, and the answer to a held one resumes only that exact action (see How approvals bind). The verdict, the hold and the prompt are the same as without the mark. A credential value ends only at whitespace, a quote, or a separator followed by another name= field, so actions that differ after it stay different actions and no part of a password with #, ,, & or ; in it stays in the hashed form. A password given as a plain argument to an unknown program (./deploy.sh hunter2) cannot be recognized; it stays only inside the hash.
• How approvals bind. Marrow binds an owner approval to the exact action that was held. The client hides possible secrets before sending an action; when it hides anything, it adds , a keyed hash of the exact action made with this machine's private device key (, created once, mode 0600, never sent). An approval resumes only for the same exact action; a retry that differs only in a certain credential (for example a new one-time password) still resumes. If the device key is missing or unsafe, Marrow prints one note (it points to the installer's ) and asks again instead of resuming.

Passive Use

npx -y --package=@getmarrow/mcp@latest marrow-mcp setup configures supported prompt, pre-action, tool-result, permission-marker, and session-stop hooks. Configuration, public hook argv, and API-key-authenticated callbacks are client self-reports. They preserve raw lifecycle activity but do not prove that the host invoked a hook or certify passive control.

Capability and coverage contract

This contract is model-neutral. A model name, host header, API key, public hook entrypoint, config file, installed hook entry, successful MCP handshake, or client-self-reported lifecycle callback does not certify passive coverage or enforcement. An unknown MCP host therefore gets the generic mcp-client identity and the same on-demand tools, schemas, and API semantics as a named host. Public lifecycle callbacks and hook activity are client-self-reported and cannot verify or certify passive coverage; independent authority is required.

When invoked by a supported host, the configured hooks send compact classifications and lifecycle receipts. They do not need raw prompts, completions, command output, tool output, or credentials. A completed tool or session does not automatically become a successful business outcome; explicit success/failure closure is required.

Setup installs distinct Claude Code and Grok hook entrypoints. The public entrypoint supplies only a client-reported display label; it is not host provenance. Hook event JSON cannot select the lifecycle harness or agent. Agent identity comes only from owner configuration when present, otherwise the request omits it so the authenticated service can derive the credential-bound identity. Every hook lifecycle event is marked source: client_self_reported and omits capability_level: native_hooks, adapter certification, configuration fingerprints, expected hooks, and observed-hook certification fields. Legacy, unknown, and custom entrypoints stay generic.

Claude Code hooks may cooperatively request guidance and apply the harness permission response, but that does not certify always-on control. Codex native hooks map a block to the supported synchronous deny response; a review-required hold follows Approvals in the Host Prompt. Cursor and Composer use the same Cursor-native pre-action, result, failure, and stop adapters. Cline uses native pre-action and post-tool adapters plus TaskCancel closeout; its documented TaskComplete hook remains coming soon and is not claimed as observed coverage. Windsurf uses native pre-action, success-result, and response-closeout adapters and requires Restricted Mode to be off. Gemini CLI uses native BeforeTool, AfterTool, and AfterAgent adapters; AfterAgent is the deterministic per-turn closeout and no SessionEnd delivery is claimed. Grok uses global native PreToolUse, PostToolUse/PostToolUseFailure, and one nonblocking Stop closeout; Marrow's generated Grok file contains no duplicate SessionEnd closeout. These native paths keep MCP on demand and require restart plus host hook review. Where a native contract cannot make a generic ask enforceable (Cursor at preToolUse, Gemini, and Grok), protected review-required or unavailable-control work is denied and waits for approval; Cline and Windsurf hand an ordinary hold to the host's own prompt when a person is evidenced. All activity remains client-self-reported rather than verified host telemetry, and the governed wrapper remains an explicit bounded fallback. Unknown and custom hosts remain on demand unless they provide a bounded event adapter, whose activity is still not certification without an independent authority.

Transient lifecycle receipts use a bounded owner-only spool and are retried with stable event IDs. Dead letters self-heal: recoverable failures are retried automatically with bounded attempts and cooldown during passive nudges, and conflicted events are recognized as server-owned evidence that is never replayed. Operators can inspect and drain it without exposing event content:

npx -y --package=@getmarrow/mcp@latest marrow-mcp spool-status
npx -y --package=@getmarrow/mcp@latest marrow-mcp drain-spool

The output contains only state, bounded pending/failed/recoverable/server-owned/recovery-exhausted counts, oldest receipt timestamps, capacity, and an exact fix. Only authentication-class dead letters require the operator: restore the credential binding, then drain. A drain applies only to the active credential-and-agent namespace: isolated legacy namespace debt is reported separately and never changes a successful active-namespace drain into a failure. Legacy files are never replayed, merged, deleted, or attributed to the active identity. Restore the exact original identity to drain one, or preserve the selected file unchanged in a separate owner-only quarantine directory when that identity is unavailable. Authentication failures in the active namespace remain explicit durable failures with repair guidance; recovery-exhausted events stop cycling instead of retrying indefinitely, and drain-spool remains an explicit operator tool rather than a user requirement.

Check the installed runtime:

marrow_agent_status

Status diagnostics distinguish missing keys, invalid keys, wrong bound-agent identity, network limits, missing hooks, and incomplete proof. They include an exact repair action without exposing secrets.

Primary MCP Tools

The package also exposes key management, fleet handoff, deployment history, adaptive policy, context/lesson, query, and workflow-example tools. See the complete source-of-truth documentation for every tool and field.

Context and Workflow Examples

The stable marrow_*memory* tools manage authorized context and prior lessons used by governance decisions. They are advanced supporting APIs, not a separate product category.

The template tools expose 24 configurable workflow examples. They are starting points for policy design, not customer case studies, regulatory validation, legal advice, or proof of production use in each listed industry.

Trust and Data Boundaries

• Private account, fleet, workflow, proof, and agent data remains tenant-scoped by default.
• Agent-bound keys can be restricted to an allowed identity and permission set.
• Sanitized aggregate contribution is optional and never means sharing raw prompts, code, secrets, proof packs, account identifiers, agent identifiers, or customer identities.
• Existing API keys are never returned after creation; key material should be supplied through the client's secret store.
• Marrow returns guidance and policy data. Agents must not execute returned text as shell input.
• An agent cannot approve itself: a caller-written proof.owner_approval is stripped, chat or proof text is never an approval receipt, and only an approval the server records counts.
• Approvals given in your tool's prompt are reported by software on your machine. An agent running as your user could fake one for a single held action. For actions that must have your own approval, lock that category (verified only); then only your one-tap link or the dashboard counts, and Marrow never learns from or reuses tool-reported approvals across agents.
• Never connect an agent to the email inbox that receives Marrow approval mail. An agent that can read that inbox can open the one-tap link itself.

See the Trust Center for implemented controls, current limits, and roadmap status.

Environment

Documentation

• Source-of-truth docs
• Trust Center
• Status
• GitHub

License

MIT

Related Packages

• @getmarrow/install - default installer, self-test, governed runner, and operator TUI
• @getmarrow/sdk - Node.js and TypeScript integration for owned agent runtimes

Codex native usage capture accepts a bound thread/tokenUsage/updated event or a bounded current transcript with the supported 0.157.1 schema. It records only the latest observed model call when matching model/turn counters prove that call's delta; it does not reconstruct whole turns or history. Repeated hooks use a compact private checkpoint and stable usage ID. Unknown versions, missing context, counter resets, unsafe/oversized files, and unproven subagent bindings abstain. Cached input and reasoning output remain subsets. Billing endpoint, tier, region, and subscription mode require explicit capture configuration; native model names do not establish a price or savings baseline.