GBrain Bootstrap — your harness as your agent

gbrain bootstrap turns a Claude Code, Codex, or opencode session into a persistent personal agent: identity files rendered from your own answers, a local PGLite brain, per-turn context, session-triggered schedules, and a private GitHub repo as the agent's durable, portable body. This guide is the full contract — what gets installed, what runs when, what it can and cannot do, and how to undo all of it.

Normative design docs: AGENT_BOOTSTRAP_DESIGN.md (scope) and AGENT_BOOTSTRAP_PLAN.md (implementation). The paste block lives in the README; the runbook your agent follows is BOOTSTRAP_FOR_AGENTS.md at the repo root, fetched at the latest-stable ref.

What gets installed, exactly

Piece	Where	Runs when
Identity files (SOUL/USER/MEMORY/AGENTS/CLAUDE/HEARTBEAT/ACCESS_POLICY/GITHUB)	your workspace folder	loaded at session start
agent.json manifest + brain/, memory/, skills/, state/	workspace	—
Local brain (PGLite)	~/.gbrain/ (never in the repo)	while a session's MCP serve is open
MCP registration (gbrain serve)	Claude Code: project scope by default; Codex: user-global (no scope flag); opencode: user-global by default (project scope is an explicit opt-in — see the degradation matrix)	spawned by your harness per session
Hooks (Claude Code, ON by default)	local installs: .claude/settings.local.json (gitignored); cloud sandboxes: the COMMITTED .claude/settings.json (PATH-resolved, fail-open commands)	each prompt; fail-open; --no-hooks opts out at install, GBRAIN_HOOKS=0 disables at runtime
Per-turn persistence	Stop hook → debounced, detached scan-gated push (per workspace; 5 min default, every turn in cloud sandboxes)	after each assistant turn; GBRAIN_STOP_PUSH=0 disables; GBRAIN_STOP_PUSH_DEBOUNCE_MIN / config hooks.stop_push_debounce_min tune it
Session persistence	SessionEnd hook → scan-gated commit+push	at session end (note: the harness never fires SessionEnd on /exit — the per-turn push is what covers that)
Compaction checkpoints	PreCompact hook → secret-scanned boundary segment banked to the corpus dir; a live serve harvests it into facts + brain:// links (see docs/guides/checkpoint-compaction.md)	at each Claude Code compaction; links render as ## Compaction checkpoints on the post-compaction session start
Push-failure visibility	next turn's context + a user-visible notice; re-announces every 30 min while failing	whenever a background push fails
Optional background job (consent-gated)	git post-commit auto-push + launchd/cron 30-min pull (pull job skipped honestly on hosts without a scheduler)	while logged in
Private GitHub repo	your account, created by bootstrap repo (or an empty repo you made yourself, adopted)	privacy verified via API
Machine receipt	~/.gbrain/bootstrap/receipt.json	uninstall is keyed to it

What does NOT run: anything while the harness is closed. Session-triggered schedules fire at turn/session boundaries only. True 24/7 operation is what a hosted brain provides — this is the honest desktop contract.

Cloud sandboxes (claude.ai/code and similar)

Cloud sessions run in a reclaimed-after-inactivity VM behind a credential-injecting egress proxy. gbrain bootstrap status --json reports execution_environment: "cloud-sandbox" there, and the install adapts:

• Hooks live in the committed .claude/settings.json with PATH-resolved, fail-open commands (no machine paths). The gitignored local settings file never survives into the next session's fresh clone, and hook config is snapshotted at session start — so hooks written mid-session go live on the NEXT session. Commit and push the file.
• The per-turn push runs every turn (debounce 0) — a reclaimed VM's tail loss is permanent, so each turn banks to the private repo.
• Repo-privacy verification falls back to pure git protocol when the proxy blocks the GitHub API (GraphQL is always pinned there; REST reaches only session-attached repos). Confirmed-public origins still always refuse.
• Repo creation is refused in cloud with the flow that works: create the private repo from a normal machine or github.com, open the cloud session ON that repo, run gbrain bootstrap attach.
• The gbrain binary installs via the environment setup script — print it with gbrain bootstrap cloud-setup-script and paste it into the environment config (npm-based; bun's package fetching is proxy-incompatible there).
• No scheduler exists — the consent-gated pull job is skipped with an honest message; event-driven pushes cover persistence.

Escape hatch for self-hosted git you trust (every use warns loudly): the CLI flag on sources push, GBRAIN_ALLOW_UNVERIFIED_REMOTE=1, or gbrain config set push.allow_unverified_remote true (file-plane — the only form that reaches detached hook children inside a sandbox).

Bring your own repo (create-repo-first)

By default bootstrap creates the private GitHub repo for you. If you prefer to own that step — pick the name/org-under-your-account, or just work the familiar way — create a new empty private repo under your own GitHub account (no README/.gitignore/license), clone it, open the clone in your harness, and run the bootstrap block. gbrain bootstrap repo detects the empty repo you created and adopts it: it verifies the repo is private, sets a repo-local git identity, and pushes your workspace. Two constraints, both enforced with a clear message rather than a silent failure:

• Empty. A repo that already has commits (a README, a license, an existing project) is refused — create it empty, or run gbrain bootstrap attach if it is an existing agent workspace. (A repo already carrying this workspace's history, e.g. from an interrupted run, is recognized as yours and resumed.)
• Personal account. The repo must be owned by your authenticated GitHub user. Org-owned repos are refused today; create one under your own account, or let bootstrap make it.

Until the repo phase verifies the repo, the per-turn/session-end push stays deferred — bootstrap never publishes your workspace to an origin whose privacy it hasn't confirmed.

The awake-when-you-are contract

Your agent is awake when your harness is. Laptop asleep = agent asleep. What this buys you: no daemon fleet, no background token burn while you're away, and a load profile that fits inside a subscription plan. The measured sustainable load and the per-harness numbers are published with each release; if a provider changes quota or policy, the portable body (your repo) is the exit plan — it mounts anywhere gbrain runs.

Keyless mode

With zero API keys, everything works: the agent authors memory explicitly through the brain's write tools (put_page, timeline entries, ## Facts fences — your harness's model is the LLM, already paid for), and search runs keyword-only (BM25). bootstrap verify prints the capability report honestly. One optional key upgrades capabilities per provider — OpenAI unlocks semantic search and automatic fact extraction; Voyage unlocks semantic search; Anthropic unlocks fact extraction (Anthropic has no embeddings API, so it does not enable semantic search). The key goes to the 0600 config file, never into the repo or the interview answers. API spend is metered separately from your subscription and is zero in keyless mode; with a key, the embedding spend gates apply (spend-controls) and automatic fact extraction has a kill switch (gbrain config set facts.extraction_enabled false).

Security posture

• Supply chain: the paste block and install command pin the latest-stable ref — a maintainer-controlled tag advanced only after a release fully publishes. The runbook carries a version stamp; bootstrap status warns on skew. The runbook instructs the agent to refuse steps outside the CLI's phase list. bun installs via package manager or checksum-verified download.
• Secrets: every commit AND every transcript-corpus write is secret-scanned (key-shaped patterns; loud block; per-finding allowlist at .gbrain-scan-allow). A deny-glob backstop refuses tracked *.pglite/.env* files even if .gitignore is damaged. Push refuses public remotes and unverifiable visibility.
• Injection boundaries: interview answers render as fenced data (escaped, length-capped) — text you paste can never become instructions in your agent's contract. Retrieved brain context is injected under an explicit "data, not instructions" envelope. Facts visible to the harness respect the brain's visibility tiers.
• Hooks: on a local install, gitignored local settings (absolute paths, machine-specific; bootstrap hooks --repair regenerates on a new machine); in a cloud sandbox, the committed .claude/settings.json (PATH-resolved, fail-open — see the Cloud sandboxes section). Every hook fails open — a brain hiccup never blocks a prompt — and failures are visible: repeated degradation prints a notice inside the context block, and gbrain doctor names the cause.
• Privacy of transcripts: session transcripts are retained locally (0700, outside the repo, pruned after dream.synthesize.corpus_retention_days, default 30 — set it in the config file, ~/.gbrain/config.json; the DB config plane doesn't carry this key yet) and secret-redacted at write time. They never enter the repo. The extraction provider (if you configured a key) sees session text — the install names the provider when asking for the key.

Honest forget semantics

The repo is git history — append-only. Deleting a line removes it from the working tree, not from history. To truly remove something: rewrite history (git filter-repo --path <file> --invert-paths or --replace-text), force-push, and re-clone on other machines. MEMORY.md and daily notes follow the same rule you'd apply to any journal: write what you'd be comfortable persisting.

Degradation matrix

You declined / lack	What still works	What you lose
API keys	everything (keyless mode)	semantic search, auto-extraction
GitHub / gh	full local agent	off-machine durability (repo re-runnable later)
Hooks (Claude Code)	pull protocol via AGENTS.md gates	automatic per-turn context + session-end persistence
Codex (no wired hooks, no MCP scope flag)	pull protocol + MCP tools	per-turn push (stated plainly; not oversold — codex 0.147+ ships a hook system, but gbrain does not wire it yet) + the ability to confine MCP reach to one folder (codex mcp add is always user-global)
opencode (no wired hooks; scope INVERTED: user-global by default)	pull protocol (opencode reads AGENTS.md natively) + MCP tools; project scope available as an explicit opt-in	per-turn push (opencode ships a plugin/event system, but gbrain does not wire it yet). The project-scope default is deliberately NOT offered: opencode spawns project-config servers with no trust prompt, so a committed entry would auto-execute on every collaborator machine
Bootstrap at all (plugin-only install)	MCP tools (starter surface, --source-guard) + the curated skill set via the codex/claude plugin (docs/mcp/CODEX.md)	identity files, hooks/push protocol, the private-repo body — the plugin is the lightweight lane; bootstrap is the full agent
Second simultaneous session	first session unaffected	second session's brain tools fail politely (one live serve per brain — v1 contract)
Postgres brain (incl. harness mode)	MCP tools every session + pull protocol	per-turn hook injection (no_pglite_path: the hook IPC socket is PGLite-only today; hooks stay pre-wired and light up when the engine-uniform listener lands)

Local harness mode (gbrain bootstrap harness, #4043)

The workspace install above is built for a human's laptop. A box run by an agent framework (your OpenClaw, or anything that shells out to claude -p / codex exec) already hosts a brain and a running gbrain serve --http — and those framework-spawned sessions get zero brain access by default. Harness mode wires them in one command, with no agent.json and no interview:

gbrain bootstrap harness --yes

• Mints a least-privilege bearer token (scopes read+write, stored in the access_tokens.scopes column; reads span the brain's federated sources). Re-runs rotate mint-first: the previous token is revoked by id only after the new one is wired and smoke-tested, so clients are never dead mid-swap. The smoke sends a deliberately invalid credential first — an endpoint that accepts anything is not this brain's serve — and a failed smoke rolls the wiring back (fresh registrations removed, replaced ones restored, the headless pre-approval stripped) and retires the fresh mint immediately, so nothing live is ever left pointed at an unverified endpoint. Prior wiring is only cleaned up after the replacement verifies.
• Claude Code: user-scope HTTP MCP registration, mcp__gbrain pre-approved in user-scope permissions.allow (headless claude -p blocks MCP tools without it), and the five lifecycle hooks — user scope by default, or exactly the dirs you pass with repeatable --project (never both; the two would double-fire every event). --no-capture wires context injection only and skips the transcript-capture events.
• Codex: one managed [mcp_servers.gbrain] block with the bearer token INLINE in the codex config (0600) — framework-spawned codex inherits no shell profile, so the env-var lane the connect path uses would never reach it. One owner per server name: if the gbrain codex PLUGIN is also enabled, two gbrain servers exist in different layers — the wire proceeds with a loud WARNING and gbrain doctor reports the collision (plugin_lane_collision); keep one (codex plugin remove gbrain@gbrain, or --remove here).
• opencode: one managed mcp.gbrain remote entry with the bearer header INLINE in the user-global JSONC config (0600), written by the same comment-preserving editor the workspace lane uses — the {env:…} interpolation the connect path prefers would resolve empty under a framework-spawned opencode for the same no-shell-profile reason. Note: downgrading gbrain below the release that introduced opencode support after wiring it leaves the opencode entry in place for manual removal — edit the opencode config by hand, or re-upgrade and run gbrain bootstrap harness --remove.
• Honesty on Postgres brains: per-turn injection is degraded (the matrix row above); MCP is the active seam and the summary says so.
• --status [--json] probes the live truth (serve health, token validity via host-config recovery — the Claude Code lane only recovers a bearer from a registration whose URL matches the receipt; the codex managed block is read from the exact path the receipt records — and per-target states) with a cron-honest exit contract: 0 only when the serve, token, and every target verify and the rotation has converged (honest degrades count as OK); 1 on an unreachable serve, a failed token verify, failed or pending targets, an unconverged rotation, or a half-removed install whose token still awaits revocation. With no install at all it says so and exits 0 (2 under --json, so machine callers can tell absence apart). gbrain doctor carries a matching bootstrap_harness_health check. --json on the install itself emits a single machine-readable document on stdout (prose goes to stderr).
• The full flag surface lives in gbrain bootstrap --help: --url/--port point at a non-default serve (a non-loopback --url is refused unless you also pass --token, which flips into registrar mode — MCP wiring only, no hooks, nothing minted), --force replaces a foreign same-name MCP registration, --name renames the server, --harness picks the hosts, and --no-hooks skips hook wiring entirely.
• --remove tears down exactly what the machine-level receipt (<home>/bootstrap/harness.json) records — host removals are engine-free and run even while a serve is live; the token revoke defers with exact instructions if a live PGLite serve holds the brain. gbrain bootstrap uninstall removes harness wiring first, automatically.
• Everything is stated before it happens; non-interactive runs require --yes. Close active Claude Code sessions for the cleanest user-scope settings writes (the host also writes that file).

PGLite note: minting needs the single-writer lock, so on a PGLite brain either pre-mint (gbrain auth create bootstrap-harness --scopes read,write while the serve is stopped) and pass --token, or stop/re-run/restart. Postgres brains mint fine while the serve runs. A token you supply is never revoked by --remove or rotation (it is not the harness's to revoke) — retire it yourself with gbrain auth revoke when you're done with it.

Binary-downgrade note: token scoping is data-only (no migration), so a gbrain binary OLDER than the release that shipped it verifies every scoped token as FULL-ACCESS — the old verify path never reads the scopes column. If you downgrade after a harness install, revoke the scoped tokens first (gbrain auth revoke with the id flag) and re-mint once you upgrade again.

Multi-device

Clone your agent repo on machine two and run gbrain bootstrap attach — it validates the manifest, wires this machine (source registration, hooks repair, MCP), and verifies. The brain database is derived state, rebuilt from brain/ + re-ingestion; hot facts extracted only on machine one arrive via the repo's pages and fences. Simultaneous editing from two machines is ordinary git conflict territory — sources push pulls divergence-safely (commit first, rebase pull, loud on conflicts).

Uninstall

gbrain bootstrap uninstall removes exactly what this machine's install receipt records: hook wiring, MCP registrations (surgically — foreign servers and hooks survive), and bootstrap-created state. Your repo is never touched — the body remains yours. The brain database is KEPT by default; --delete-brain is offered only when bootstrap created the brain, offers a facts export first, and enumerates what it is about to remove. It refuses to run while a session's serve is live.

If something seems broken

One command: gbrain doctor. It covers hook health, push staleness, serve/lock collisions, schema state, and prints fixes. gbrain bootstrap status --json emits a support blob (versions, harness, last verify/push, hook failure rate) your agent can relay verbatim when you report a problem.

Real-agent e2e

Most bootstrap tests drive the dispatcher with PATH-shimmed claude/codex recorders — fast, hermetic, no API cost. Two additional "door" tests drive the ACTUAL binaries end to end so we catch real-world drift (a codex mcp add flag that changed shape, a harness that stopped calling our MCP server):

• test/e2e/bootstrap-real-claude.serial.test.ts — real claude -p over MCP.
• test/e2e/bootstrap-real-codex.serial.test.ts — real codex exec. It runs the keyless-init → interview → render → gbrain bootstrap hooks --harness codex path (executing the real codex mcp add into a hermetic ~/.codex/config.toml), asserts the rendered AGENTS.md carries the Gate-3 brain-first pull protocol (gbrain does not wire Codex hooks yet, so the pull protocol is its per-turn seam), then spends one live codex exec turn to prove real codex → gbrain MCP → brain → a seeded, brain-only fact (falling back to a shell gbrain query if headless stdio-MCP is unavailable).

opencode's real-binary door lives in test/e2e/install-real-opencode.serial.test.ts (its writer-parity leg handshakes gbrain's direct JSONC registration through the actual binary); docs/TESTING.md carries the full door inventory and cadence policy.

These pay real API cost and take 30s–2min per turn, so they are NOT in the PR shard. Everything is hermetic (temp HOME / CODEX_HOME / CLAUDE_CONFIG_DIR / GBRAIN_HOME per test — the operator's real ~/.claude, ~/.gbrain, ~/.codex are never touched; auth is copied read-only). Each file self-SKIPS via describe.skipIf when its binary or auth is absent, so on a machine without the tool it is a clean no-op that never fails. CI wires them into the real-agent-e2e job in .github/workflows/heavy-tests.yml (nightly + the real-agent-e2e / heavy-tests label); on a stock runner they self-skip. To actually exercise the binaries you need a runner with authed claude/codex and the provider creds (GSTACK_ANTHROPIC_API_KEY/ANTHROPIC_API_KEY, VOYAGE_API_KEY) exported.

Run locally (where both are installed + authed):

bun test test/e2e/bootstrap-real-codex.serial.test.ts

DX exploration harness (developer instrument, not a test)

The door tests prove the install WORKS; they say nothing about how it FEELS. test/helpers/tty-harness.ts spawns any CLI (gbrain, claude, codex, grok, opencode) under a real pseudo-terminal (Bun's terminal: spawn option) and records every output burst with a millisecond timestamp, so unnecessary pauses become a measurable artifact (computeStalls → stalls.md) instead of a vibe. Same hermetic env as agent-harness.ts; pure helpers are unit-tested in test/tty-harness.test.ts (zero subprocesses, PTY smokes self-skip where terminal: is unavailable). The harness itself also backs one required-CI test: test/init-picker-pty.serial.test.ts asserts the interactive gbrain init pickers under a real PTY (see the TTY decision table in docs/TESTING.md). The DX-exploration layer below stays an instrument — nothing in it asserts.

scripts/dx-explore.ts drives it to capture the fresh-user funnel as timestamped transcripts under .context/dx-runs/ (gitignored — nothing asserts, no CI):

bun run scripts/dx-explore.ts help              # comprehension surfaces (no keys)
bun run scripts/dx-explore.ts init [--keyless]  # interactive init, naive-user autopilot
bun run scripts/dx-explore.ts claude-install    # REAL claude running the paste-in bootstrap
bun run scripts/dx-explore.ts codex-install     # REAL codex, same
bun run scripts/dx-explore.ts opencode-install  # REAL opencode running the paste-in bootstrap
bun run scripts/dx-explore.ts grok-install      # REAL grok, brain-only GROK.md install (no bootstrap path)
bun run scripts/dx-explore.ts drive -- gbrain init   # manual: steer a live TUI via a file channel

drive mode is how an agent in a Conductor workspace explores a live TUI across separate tool calls: cat <dir>/session/screen.txt to watch, append {"line":"..."} / {"key":"Down"} / {"stop":true} to <dir>/session/input.jsonl to steer. Each run writes meta.json, visible.txt, frames.jsonl, and stalls.md. --keyless strips provider keys so the true no-key first-touch path is exercised (a Conductor session's ambient ANTHROPIC_API_KEY would otherwise leak in). Install scenarios pay real API cost — launch them as background tasks.