PlugBrain CLI Reference

The plugbrain Command Line Interface provides local administration, code indexing, question answering, multi-client configuration, hardware inspection, and MCP integration.

usage: plugbrain <command> [options]

Workspace Setup and Client Configuration

`init`

Initialize PlugBrain in the current directory or specified path with a single command.

Discovers the git repository root (or uses the directory path), registers the workspace, executes initial code indexing with a live progress report, prints the local web UI URL, detects installed AI coding assistants (Claude Code, Codex, Cursor, Windsurf, Hermes, AGY, OpenCode) to enroll PlugBrain into their MCP client configurations, and adds the PlugBrain agent-protocol block to AGENTS.md (plus CLAUDE.md when present and not already referencing AGENTS.md). Running init multiple times is idempotent.

plugbrain init [path] [--no-clients] [--no-agents-file] [--dry-run]

Flags:

[path]: Directory path to initialize (defaults to the current working directory).
--dry-run: Preview planning actions without making any changes to disk.
--no-clients: Skip detecting and updating AI coding assistant client configurations.
--no-agents-file: Skip adding or updating the agent-protocol block.

Behavior of --dry-run:

Does not create or modify plugbrain.db on disk. If a database file already exists, it opens it strictly read-only; if no database exists, it uses an in-memory database (:memory:).
Evaluates workspace registration using planWorkspaceRoot instead of inserting records, reporting whether the folder would register or is already known.
Bypasses index scanning (outputs would index <workspace>).
Inspects client configurations and outputs a unified diff (renderEntryDiff) showing exact lines that would be added to each detected client, without touching any configuration files.

Example:

plugbrain init --dry-run

`setup`

Configure local AI coding clients to use PlugBrain via MCP stdio (plugbrain mcp).

Discovers supported client configurations in the user profile directory (or PLUGBRAIN_CONFIG_HOME), updates their configuration files (JSON, TOML, or YAML) to include the plugbrain MCP server entry, and maintains automatic timestamped backups.

plugbrain setup [--all|<client_name>] [--dry-run] [--undo]

Supported Client Names: claude, codex, cursor, windsurf, hermes, agy, opencode (or --all).

Flags:

--all: Target all detected client configurations (default if no specific client is provided).
--dry-run: Display a unified diff of proposed configuration changes without modifying files.
--undo: Revert the last applied setup modifications.

Behavior of --undo:

Reverts client configuration files to their state prior to PlugBrain setup.
If a client configuration file was newly created by setup, undo removes the plugbrain entry. If no other user entries exist, the file is safely deleted and reported as restored. If the user added custom settings after setup, the file is preserved and only the plugbrain section is removed.
For pre-existing configuration files, undo restores the configuration from the latest setup backup (<file>.plugbrain-backup-<timestamp>).
Before restoring or modifying any file, undo writes a counter-backup (<file>.plugbrain-undobackup-<timestamp>) of the current state, ensuring subsequent user edits are never lost.
If no backup is found and no entry exists, undo safely reports the file as skipped without errors.

Example:

plugbrain setup --all
plugbrain setup --undo

`agents-file`

Add or refresh the instructions from docs/agent-protocol.md between managed markers. Text outside those markers is preserved. The default writes AGENTS.md and also CLAUDE.md when that file exists without an AGENTS.md reference.

plugbrain agents-file [--target AGENTS.md|CLAUDE.md|both] [--dry-run] [--undo]

--dry-run prints the proposed content without writing. --undo restores the pre-change backup, or removes the managed block when no backup exists. Duplicate or incomplete markers are refused with a repair hint.

`doctor`

doctor --agents performs read-only checks for each supported client: installation, PlugBrain MCP config, the local daemon health endpoint, current-directory workspace registration, and the managed agent block. Each row includes a repair command. Use --json for machine-readable rows. Doctor never starts the daemon or edits configuration.

plugbrain doctor --agents [--json]

MCP Server and Daemon

`mcp`

Start the Model Context Protocol (MCP) JSON-RPC 2.0 stdio server for client integration.

When invoked without --workspace, the server resolves the workspace context automatically from the current working directory:

Respects PLUGBRAIN_WORKSPACE if set.
Matches registered workspace roots containing the current directory (longest match wins).
If inside an unregistered git repository, automatically registers the git root and warms the index in the background.

plugbrain mcp [--workspace <id>] [--auth-key <key>]

Flags:

--workspace <id>: Explicitly bind the MCP server to a specific workspace ID.
--auth-key <key>: Provide authentication token for mutating tools.

Example:

plugbrain mcp

`serve`

Start the local HTTP API and serve the embedded web dashboard.

Binds exclusively to 127.0.0.1 (default port 4310) and launches the background file-watcher daemon (unless disabled with PLUGBRAIN_NO_DAEMON=1). Automatically generates and persists an authentication token in ~/.plugbrain/auth.token if not present.

plugbrain serve [port]

Arguments:

[port]: Local TCP port to bind (default 4310).

Example:

plugbrain serve 4310

Search and Natural-Language Question Answering

`ask`

Ask questions about the project codebase in plain language (English or German).

Evaluates the question intent deterministically and queries code intelligence, git status, architecture overview, or vault notes without invoking an external model.

plugbrain ask "<question>" [--workspace <id>] [--limit <n>] [--json]

Flags:

--workspace <id>: Specify workspace ID (required only if multiple workspaces exist).
--limit <n>: Maximum number of sources cited (default 5, max 25).
--json: Output raw JSON response conforming to AskResponse.

Example:

plugbrain ask "where is the brain home resolved?"

`search`

Fast full-text token search over symbols and files in the database.

plugbrain search <query>

Example:

plugbrain search resolveBrainHome

Code Intelligence (`intel`)

`query`

Search AST symbols, file paths, and notes by concept or keyword.

plugbrain query <search_query> [--repo <name>] [--limit <n>] [--json]

Example:

plugbrain query "openStore" --limit 10

`context`

Inspect 360-degree context of a code symbol, including callers, callees, and test references.

plugbrain context <symbol_name> [--file <path>] [--repo <id>] [--json]

Example:

plugbrain context openStore --file src/store/schema.ts

`impact`

Perform blast radius analysis upstream, downstream, or bidirectionally for a symbol or file.

plugbrain impact <symbol_or_path> [--direction upstream|downstream|both] [--depth <n>] [--json]

Example:

plugbrain impact src/store/schema.ts --direction upstream --depth 2

`detect-changes`

Map git diff hunks to affected code symbols and execution flows.

plugbrain detect-changes [--path <checkout_dir>] [--workspace <id>] [--json]

Example:

plugbrain detect-changes

`cypher`

Execute graph queries against the indexed symbol dependency graph.

plugbrain cypher "<query>" [--limit <n>] [--json]

Example:

plugbrain cypher "MATCH (f:File)-[:CONTAINS]->(s:Symbol) WHERE s.name = 'openStore' RETURN f.path, s.name"

`intel-status`

Display code intelligence index health, symbol/edge counts, and staleness metrics.

plugbrain intel-status [--workspace <id>] [--json]

Example:

plugbrain intel-status

System Health, Hygiene, and Census

`hygiene`

Check git repository hygiene across all checkouts: uncommitted files, unpushed branches, stashes, and orphaned worktrees.

plugbrain hygiene [--workspace <id>] [--json] [--wip-snapshot] [--repo <path>]

Flags:

--workspace <id>: Specify workspace ID.
--wip-snapshot: Save uncommitted working tree changes into temporary git refs (refs/wip/<date>/<name>).
--repo <path>: Rescue a specific standalone repository not registered in PlugBrain.
--json: Output structured JSON report.

Example:

plugbrain hygiene
plugbrain hygiene --wip-snapshot

`machine`

Inspect hardware resources: disk space per volume, pagefile size, RAM, CPU usage, and capacity forecast.

plugbrain machine [--json]

Example:

plugbrain machine

`repos`

Machine-wide git repository census: discover registered and unregistered git checkouts.

plugbrain repos [--dirty] [--refresh] [--json]

Flags:

--dirty: Filter only repositories with uncommitted or untracked changes.
--refresh: Force a fresh disk walk ignoring cached census data.
--json: Output structured JSON report.

Example:

plugbrain repos --dirty

`disk`

Scan directory metadata, inspect the saved directory tree, rank safe-to-review storage recommendations, or create a reinstallation checklist. Scanning never reads file contents except package lockfiles, which are opened only to compute duplicate fingerprints. Links and other reparse points are reported and skipped.

plugbrain disk scan [<path>...]
plugbrain disk tree [<path>]
plugbrain disk recommend [--json]
plugbrain disk wipe-check [--json]
plugbrain disk largest [n]

scan defaults to all fixed drives and replaces the previous scan. Explicit paths are useful for a bounded, read-only inventory.
tree returns immediate children from the last completed scan.
recommend groups recoverable dependency trees, caches and build output; review candidates and work that should be kept are marked separately.
wipe-check lists uncommitted or unpushed Git work, recent user-profile folders and configuration or vault paths to back up before reinstalling.
largest reports the largest files retained in the last scan (maximum 1,000).
--json prints the structured report.

plugbrain disk scan C:\\Users\\me\\Projects
plugbrain disk recommend --json
plugbrain disk wipe-check

The scan and recommendations only report. They do not delete or move files.

Workspace and Index Lifecycle

`register`

plugbrain register <path> [name]

Example:

plugbrain register C:/code/my-project my-project

`index`

Trigger incremental or full reindexing for one or all registered workspaces.

plugbrain index [workspaceId]

Example:

plugbrain index

`progress`

View progress and status of active or recent background indexing runs.

plugbrain progress [workspaceId]

Example:

plugbrain progress

`status`

Summarize all registered workspaces, indexed files, symbol counts by kind, edge relations, and database size.

plugbrain status

Example:

plugbrain status

`planet`

Manage multi-repository workspace planets, checkouts, pruning, and deletion history.

plugbrain planet <step> [args]

Subcommands:

planet register [path] [name]: Register a multi-repo planet directory and discover checkouts.
planet select [workspaceId] --checkout <id>: Persist explicit active checkout roots.
planet scan [workspaceId]: Refresh git revisions across checkouts and trigger index.
planet status [workspaceId]: Show repository counts, checkout branches, and revisions.
planet history [workspaceId]: View historical files deleted or renamed from the planet.
planet prune [workspaceId] [--apply] [--compact] [--batch <n>]: Remove indexed rows of checkouts no longer selected.

Example:

plugbrain planet status

`compact`

Reclaim free SQLite database pages on disk using VACUUM.

plugbrain compact

Example:

plugbrain compact

Planning and Master Spec (`plan`)

`plan`

Inspect master specification ledgers, progress calculations, gate checklists, and task dependencies.

plugbrain plan [status|next|task <task_id>|gates] [--json]

Subcommands:

plan status: Show overall progress percentage, task status counts, and open decisions (default).
plan next: List immediately startable unblocked tasks.
plan task <task_id>: Show detail, requirements, gates, and dependencies for a task (e.g., M12).
plan gates: View the list of all specification gates.

Example:

plugbrain plan status
plugbrain plan next

Vault and Notes Management (`notes`)

`notes`

Manage Obsidian-compatible markdown notes and knowledge graphs.

plugbrain notes <subcommand> [args]

Subcommands:

notes list [--limit <n>] [--json]: List notes with title, type, and link counts.
notes query <filter>: Property query over frontmatter (e.g., "typ=gate UND stand=offen").
notes search <text> [--limit <n>] [--json]: Search note prose text with context snippets.
notes read <path>: Read note body, frontmatter properties, outgoing links, and backlinks.
notes write <path> --from <file>|--content <text> [--expect <hash>]: Write note with version concurrency check.
notes graph [--focus <path>] [--depth <n>] [--filter <expr>]: Compute note relation graph with color groups.
notes backlinks <path>: List all notes linking to the target note.

Example:

plugbrain notes search "architecture"
plugbrain notes read "docs/quickstart.md"

Database Backup and Recovery

`backup`

Create an online SQLite backup of the PlugBrain database.

plugbrain backup [target_path]

Example:

plugbrain backup

`restore`

Restore the database from a verified backup file.

plugbrain restore <backup_path>

Example:

plugbrain restore ~/.plugbrain/backups/plugbrain-backup-1727337600000.db

Multi-Agent Swarm and Access Layer

`attach`

Attach an agent identity to a workspace and output the session briefing.

plugbrain attach <workspaceId> <agentId>

`read`

Read a file through the audited agent access layer.

plugbrain read <workspaceId> <agentId> <path>

`write`

Write file content with attribution to the active agent.

plugbrain write <workspaceId> <agentId> <path> <content>

`who`

Display file provenance, current owner, and edit history.

plugbrain who <workspaceId> <path>

`agents`

List all registered agents and their last active timestamps.

plugbrain agents

`swarm`

The coordination desk for a team of agents: registration, turns, the task queue, messages, path claims, commit approval, quotas and machine admission. It works on the local store directly, so an agent needs no HTTP token. The per-turn rules for agents are in agent-protocol.md. The command output is in German for now.

plugbrain swarm <subcommand> [args] [--workspace <id>]

Without --workspace the single registered workspace is used.

Subcommands:

swarm watch [--for <agent>] [--agents <id,…>] [--dirs <path,…>] [--timeout 20m] [--json]: Wait for relevant Brain events and sparse filesystem/resource changes.
swarm register <agentId> [name]: Register a new agent worker.
swarm turn <agentId> <start|end>: Check in at a turn boundary.
swarm claim <agentId> <path> --task <taskId>: Acquire mutual exclusion lease on a path.
swarm release <agentId> <leaseId>: Release held lease.
swarm board: View active fleet status, unread messages, and worktrees.
swarm supersede <task> [--by <task>] --note <reason>: Permanently prevent a pending task from being offered again, optionally linking its replacement.
swarm reassign <task> --to <agent>: Restrict a pending task to a registered worker.
swarm priority <task> <n>: Set an integer priority; larger values are offered first.
swarm reap [--repo <path>] [--target <branch>] [--apply]: Preview or safely remove merged, clean linked worktrees; defaults to a read-only dry run.
swarm reap --auto on|off: Enable or disable automatic cleanup for this workspace.
swarm send <fromAgent> <toAgent> --body <text>: Send peer message.
swarm resources: Display host CPU, RAM, disk quotas, and admission decisions.
swarm admit <test|build|install>: Check admission gate for hardware-intensive actions. Agents
swarm register <agent> --surface <s> --account <label> [--quota-pool <pool>] [--key <resource>] [--model <m>] [--name <n>] [--worktree <path>]... [--takeover]: register an agent. Surfaces: claude-code, codex-app, freebuff, agy, native, other. --worktree binds the agent to a checkout so the board can show its branch. Workers that draw from one provider project must share the same --quota-pool, even when they carry different resource keys; it defaults to the account label.
swarm turn <agent> start [--claim]: check in at the start of a turn. Prints unread messages and the task offered to this agent; --claim takes it.
swarm turn <agent> end --state needs-task|awaiting-commit|blocked|paused [--summary <s>] [--deliver <evidence>] [--repo <worktree>] [--review]: check out with exactly one state. --deliver hands in the single held task; the evidence must be a non-empty file inside the workspace. A delivery records the agent, attempt, source revision when available, evidence path and SHA-256. Review tasks titled R-… (or explicitly marked with --review) also need a judgment (PASS, PASS_MIT_AUFLAGEN or FAIL) and a 40-character reviewed commit hash. Without a delivery, the queue task stays claimed and appears under swarm board --next.
swarm ack <agent> <messageId>: mark a message as read. Unread messages are shown at every turn start until acknowledged.
swarm retire <agent> --note <reason>: take an agent off the board.

Work

swarm enqueue <title> [--body <b>] [--to <agent>] [--plan <M00>] [--by <agent>] [--after <taskId>]: queue a task, for one agent (--to) or for whoever takes it first. --after makes it wait: it is neither offered nor claimable until the predecessor is delivered or its holder has ended a turn with awaiting-commit or needs-task; the addressee then gets one message that the task is free.
swarm supersede <task> [--by <task>] --note <reason>: permanently remove a pending task from worker offers and record the actor, optional replacement task, and reason. Superseded tasks and the latest queue changes appear on swarm board and in its JSON projection.
swarm reassign <task> --to <agent>: limit a pending task to a registered worker. The change is audited and shown on the board.
swarm priority <task> <n>: set a pending task's integer priority from -100000 to 100000. Higher values are offered first; equal priorities remain first-in-first-out.
swarm deliver <agent> <taskId> --path <evidence> [--repo <worktree>] [--review]: hand in a claimed task. The named agent must hold it and the evidence file must exist, be non-empty and resolve inside the workspace. The receipt includes the task, agent, delivery attempt, source HEAD when --repo or the selected workspace root identifies a Git checkout, normalized evidence path and SHA-256. Review tasks titled R-… (or passed with --review) require a verdict and reviewed commit hash.
swarm wave-done <waveId> [--json]: evaluate the authoritative operator wave.manifest trace (payload.waveId and payload.taskIds) against stored delivery receipts, an authoritative independent review.completed trace (payload.verdict and payload.commit) for the delivered commit, and either an integration.accepted trace or an explicit wave.owner-decision trace per task. Missing evidence returns BLOCKED (exit 1) and emits no event. A complete report is persisted and emits one WAVE-DONE event; repeating the command returns the same report without emitting another event.
swarm claim <agent> <path>... [--task <id>] [--ttl-min <n>]: take a write lease on paths. A path another agent holds is refused.
swarm release <agent> [<path>...] [--task <id>]: release leases (all of the agent's, or the given ones).
swarm send <agent> --subject <s> --body <b> [--from <agent>]: message an agent. The first argument is the recipient. The sender comes from --from or from PLUGBRAIN_AGENT; with neither, the message is refused (exit 2) — the Brain never books a message to the integrator that somebody else wrote.
swarm approve <agent> [--note <n>] [--by <agent>]: let an agent that ended with awaiting-commit commit. It arrives as a message.

Overview and machine

swarm board [--git] [--json] [--next]: every agent once with account, turn state, unread messages, task, leases and worktree; --git adds branch, head and uncommitted files, plus overlaps where two agents write in one worktree. A working worker without Brain contact is marked silent and shown with its minutes (still? seit 47 min ohne Kontakt).
swarm watchdog [--json] [--dry-run]: one watchdog cycle. It reports every silent worker, tells the integrator once per case (--dry-run only reads), and releases waiting tasks whose predecessor has arrived. A worker is silent when it has been working longer than the workspace threshold (default 45 minutes) without any Brain contact — turn, claim, release, ack, message, delivery or quota. Contact resets the reading; the Brain never calls a worker dead.
swarm watchdog silent-after <minutes>: set that threshold for the workspace.
swarm review-pool set <agent>...: replace the workspace's ordered reviewer list.
swarm review-pool auto on|off: switch automatic review routing (default off). When on, a turn that ends with awaiting-commit queues one review task for the first reviewer whose account differs from the author's.
swarm review-pool show [--json]: pool, routing switch and threshold. uncommitted files, plus overlaps where two agents write in one worktree. The board appends an Als Nächstes section with ranked suggestions, reasons and commands; --next prints just those suggestions.
swarm chronik [--since <iso|2h>] [--json|--md]: show stored queue, message, lease and quota events in timestamp order. Markdown is the default and can be used as a handoff note. Missing turn history is named explicitly until a version that stores turn history is available; the chronicle does not invent events. Reads are capped at 500 rows per source and output at 1,000 events; capped results say that older events were omitted.
swarm resources [--json]: free disk and RAM, admission per kind of work, reported quotas and configured pool limits.
swarm quota <account> <remaining> <percent|credits|requests|rpm|tokens> [--resets <iso>] [--note <n>]: report how much of an account's quota is left. Numbers only, never keys.
swarm quota-pool set <pool> --max-concurrent <n> [--rpm <n>]: configure one shared provider/project pool. Reservations for worker attempts count against the same concurrency and optional rolling RPM limit. swarm quota-pool show [--json] reports active and settled attempts; unreported usage remains unknown rather than being counted as zero. A provider 429 applies a bounded, pool-wide cooldown using Retry-After when available.
swarm admit <edit|test|index|build|install|worktree>: exit 0 means there is room on this machine for that kind of work, exit 5 means there is not.

Starting a worker: swarm run

A CLI worker (a Codex tab, any other command-line agent) can be started and watched by the Brain itself, so a lane needs no handwork at its turn boundary.

swarm runner set <agent> --cmd <exe> [--model <m>] [--effort <e>] [--sandbox bypass|workspace-write|read-only|danger-full-access] [--search] [--args "<template>"] [--cwd <dir>]: store how this worker is started. No key is stored — Codex brings its own login. --cmd codex builds the codex exec command line (-m, -c model_reasoning_effort=, --sandbox, -c tools.web_search=, --json, -o); any other program is built from --args, whose {prompt}, {promptFile}, {log}, {last}, {cwd}, {model}, {effort} and {sandbox} are filled in per run. A token that resolves to nothing is dropped.
swarm runner show <agent>: the stored profile.
swarm run <agent>: start a detached refill loop by default. It atomically claims the next eligible queue task, starts the worker with that task in its prompt, and waits for the run to end before starting another attempt. The loop stops after 24 idle checks (five minutes each), on an authentication failure, or when a task needs a lead decision. --max-idle-checks, --idle-ms, --max-attempts (1–5) and --poll-ms are available for controlled operation and tests. A failed attempt is retried at most twice by default; quota errors use a bounded 30-minute cooldown, and auth errors stop immediately. Each task attempt has a fencing token; stale attempts cannot issue further swarm state changes. The token is generated locally and is not a credential.
swarm run <agent> --once: start one process detached with an empty stdin and a JSONL log at <PLUGBRAIN_HOME>/runs/workers/<agent>-<stamp>.jsonl (stderr beside it, the last message at …-last.md, the rendered prompt at …-prompt.md). The prompt comes from a template in the source and names the protocol documents the registered workspace actually has — protocol, newest lane rules and the Brain's own wrapper under <workspace>/koordination — so nothing owner-specific is baked in. The run row is written before the process starts, and the working directory defaults to the workspace root.
swarm run <agent> --stop-supervisor: ask the refill loop to stop after its current worker process settles.
swarm run <agent> --status: pid, start time, log path, last log event and the worker's turn state, plus the tail of the log.
swarm run <agent> --stop: stop the process (on Windows the whole process tree of a .cmd shim) and set the worker's turn to paused — unless it had already ended its turn, which is kept.
swarm board shows runner: pid … running since … and the last log event per worker. Looking at the board is also what settles a run whose process is gone: a worker that died without ending its turn is booked blocked with the last log lines as its summary, and the integrator gets a message. A worker that ended its turn first is left alone. Liveness is only ever process.kill(pid, 0); a quiet log never means dead.
Nothing is committed and no approval is bypassed: a started worker still claims its paths, admits its tests and ends its turn with a state like any other.

Example: one task through one agent

plugbrain swarm enqueue "Fix the flaky login test" --body "Repro in issue #12" --to codex-1
plugbrain swarm turn codex-1 start --claim
plugbrain swarm claim codex-1 src/auth/login.ts --task <task-id>
plugbrain swarm admit test && npm test
plugbrain swarm turn codex-1 end --state awaiting-commit --summary "Fixed the race, login tests pass" --deliver closeout/login.md
plugbrain swarm release codex-1 --task <task-id>
plugbrain swarm approve codex-1 --note "Reviewed. Commit it."

Watching coordination

plugbrain swarm turn agent-1 start
plugbrain swarm board
plugbrain swarm watch --for agent-1 --dirs ./review --timeout 20m --json

swarm watch streams coordination events from the local Brain live-event endpoint, batches events for five seconds, and checks watched Markdown directories and resource admission every 30 seconds. It reports turn changes, new messages and queued tasks for the watched agent, task deliveries, the still? watchdog, and the file/admission changes it polls itself. It exits with 0 on an event, 3 on timeout, and 1 after the live-event stream keeps failing. Set PLUGBRAIN_URL when the local Brain is served on a non-default URL.

The stream carries a monotonic event id and reconnects with bounded backoff on a dropped connection. swarm watch remembers the last id it saw in <store home>/swarm-watch-cursor and sends it as Last-Event-ID on the next (re)connection; the server replays the events still in its bounded buffer after that id. A restart therefore resumes where the previous run stopped — events older than the replay buffer are the only gap.

Example: let the Brain start the agent

plugbrain swarm runner set cx01 --cmd codex --model gpt-6-luna --effort high --sandbox bypass
plugbrain swarm run cx01
plugbrain swarm run cx01 --once     # one detached invocation, without queue refills
plugbrain swarm run cx01 --status     # pid, last log event, turn state
plugbrain swarm board                 # runner column, and settles dead runs
plugbrain swarm run cx01 --stop       # stops it, turn becomes paused
plugbrain swarm run cx01 --stop-supervisor

FilesExpand file tree

cli.md

Latest commit

History

cli.md

File metadata and controls

PlugBrain CLI Reference

Workspace Setup and Client Configuration

init

setup

agents-file

doctor

MCP Server and Daemon

mcp

serve

Search and Natural-Language Question Answering

ask

search

Code Intelligence (intel)

query

context

impact

detect-changes

cypher

intel-status

System Health, Hygiene, and Census

hygiene

machine

repos

disk

Workspace and Index Lifecycle

register

index

progress

status

planet

compact

Planning and Master Spec (plan)

plan

Vault and Notes Management (notes)

notes

Database Backup and Recovery

backup

restore

Multi-Agent Swarm and Access Layer

attach

read

write

who

agents

swarm

`init`

`setup`

`agents-file`

`doctor`

`mcp`

`serve`

`ask`

`search`

Code Intelligence (`intel`)

`query`

`context`

`impact`

`detect-changes`

`cypher`

`intel-status`

`hygiene`

`machine`

`repos`

`disk`

`register`

`index`

`progress`

`status`

`planet`

`compact`

Planning and Master Spec (`plan`)

`plan`

Vault and Notes Management (`notes`)

`notes`

`backup`

`restore`

`attach`

`read`

`write`

`who`

`agents`

`swarm`