Built with Astro, hosted on Cloudflare. Domain registered in 2026.
thrum.team is built with Astro, hosted on Cloudflare. It measures traffic with Plausible Analytics. It sends email through Amazon SES.
| Technology | Version | First seen |
|---|---|---|
| JavaScript Frameworks & Libraries1 | ||
| – | 12 Nov 2026 | |
| CDN & Hosting1 | ||
| – | 12 Nov 2026 | |
| Analytics & Tag Management1 | ||
| – | 12 Nov 2026 | |
| Transactional Email2 | ||
| – | 12 Nov 2026 | |
Registration from the registry's RDAP record, the certificate the site serves, and its rank in public traffic lists.
The crew hasn’t read this site’s colors yet. We read a palette when we capture a site’s screenshot, busiest sites first. Meanwhile, browse palettes from other sites.
Counted from every stack change we've recorded: of the sites we watched running one of this site's tools, how many went on to add another. A pattern across sites, not a forecast for this one.
smtp.google.comcarmelo.ns.cloudflare.comdesi.ns.cloudflare.comgoogle-site-verification=l6LQNQKPntirLBjQB79OnMUj17uo1psKyBH-Fo-aXo4v=spf1 include:_spf.google.com ~allgoogle-site-verification=SGJmdpAv1DU-yJrd73N-M8J8zdRy7QbnDWIk2GKS3AIv=DMARC1; p=none; rua=mailto:support@thrum.team; fo=1User-agent: * Allow: / Sitemap: https://thrum.team/sitemap-index.xml
# Thrum v1.1.2
> Git-backed messaging for AI agent coordination. Real-time. Cross-machine. Cross-repo. Persistent.
## What is Thrum
Thrum is a messaging and coordination system for AI agents. Messages are stored
in append-only JSONL logs (sharded by agent) on a dedicated `a-sync` orphan
branch, accessed via a `.git/thrum-sync/a-sync` git worktree. A SQLite projection provides
fast queries. A background daemon provides JSON-RPC over Unix socket and
WebSocket, with automatic Git sync every 60 seconds. No branch switching is
needed -- all sync operations happen within the worktree.
New in v0.7.0: single-agent mode (default for new repos), cross-repo peer
communication via Tailscale, PID-based identity resolution, and a three-tier
context model for session persistence across compaction.
New in v0.7.1: tmux-managed sessions replace background listeners with instant
daemon-driven nudge delivery. Session restart extracts conversation history from
JSONL transcripts and reloads it in new sessions.
New in v0.8.0:
- Tmux command queue (`thrum tmux queue/queue-status/cancel`) — daemon-managed
FIFO dispatch with `@system` completion notifications and restart recovery.
- Worktree management (`thrum worktree create/teardown/list`) — git worktrees
with automatic thrum and beads redirect wiring.
- Orchestrator role — plan execution with review gates, agent lifecycle, and
worktree management. Includes `thrum:orchestrate` skill.
- Multi-runtime tmux — `ClaudePID` renamed to `AgentPID` (schema v17).
`PreferredRuntime` field + `--runtime` flag on `thrum quickstart`. OpenCode,
Codex, and other runtimes launch via tmux alongside Claude.
- Agent status (`thrum agent set-status working|idle|blocked`) — auto-nudge
detection when working status mismatches silent pane.
- Daemon logging — slog + lumberjack rotation. `thrum daemon logs` command.
Configurable `daemon.log_level` in config.json.
- Graceful restart flow — `thrum tmux restart` (without `--force`) asks the
agent to save its own snapshot before killing, falling back to JSONL
extraction on timeout.
New in v0.8.1: CI fix for npm publish packaging (no user-facing changes).
New in v0.8.2:
- Cursor plugin — full Cursor integration with rules, hooks, skills, commands,
and MCP config. Deployed via `cursor-plugin/local-install.sh` into `.cursor/`.
- Antigravity (agy) — converts the Claude Code plugin: `agy plugin install
https://github.com/thrum-team/thrum-pro/tree/main/claude-plugin`. No hooks or MCP
after conversion; Thrum types `/thrum:prime`.
- Worktree create quickstart fix — `thrum worktree create` with quickstart
flags now creates a real tmux session (kept, not destroyed) and runs
quickstart inside it via SendKeys instead of the old temp-session + queue
RPC path that rejected agentless sessions. Daemon-side retry at 5s if shell
init swallowed the command, CLI-side pane capture at 12s if the identity
file still hasn't appeared. Repo-root guard: errors out if the resolved
worktree path or `worktrees.base_path` is the repo root itself.
- Two-step create + launch — `thrum worktree create` and `thrum tmux create`
now print explicit next-step hints: "Agent is NOT running yet. Start it
with: thrum tmux launch <name>". The agent identity is registered, but the
runtime is not started until `tmux launch` runs.
- `thrum tmux launch` hard-errors on missing identity — sessions created
with `--no-agent` (or worktrees with no identity file) cannot be launched.
Returns an error telling you to run `thrum quickstart` first. Launching a
runtime without an identity is a no-op.
- Next-step guard messages — `agent register`, `worktree create` (no agent),
`purge --confirm`, `daemon stop`, and `tmux restart` now print explicit
hints about what to do next or what just broke. `daemon stop` warns that
messaging commands will fail until the daemon is restarted. `tmux restart`
warns when the snapshot has 0 lines (no conversation history).
- Purge auto-re-register message corrected — agents auto-re-register via
`RefreshLocalIdentity` wired into `getClient()`, so no manual action is
needed after `thrum purge`.
New in v0.9.0:
- Safe agent registration — `thrum tmux create` now requires quickstart flags
(`--name`, `--role`, `--module`) or `--no-agent`; runs quickstart inside the
pane automatically. New alias: `thrum tmux quickstart`.
`thrum worktree create` gains optional quickstart flags; when all three are
provided it creates a real tmux session and runs quickstart inside it.
New alias: `thrum worktree setup`. Single-identity-per-worktree enforcement.
- Monitor Jobs — `thrum monitor start/list/show/update/stop/delete/logs/restart`.
Spawn a child process, watch stdout+stderr with a regex, and emit matches as
synthetic Thrum messages to a target agent. (`thrum monitor add` is an alias
for `thrum monitor start`.) Leading-edge debounce (default 60s).
Local-only, max 100 monitors. `stop` leaves the record behind and reserves
the name — `delete` is what frees it for reuse.
- Permission prompt detection — daemon-side silence-hash poller (~20s detection
latency). Polls every 10s; stable on 2 identical hashes. Permission nudges
auto-forward to Telegram when bridge is configured and bridge user is a
`permission_supervisors` entry. Reply via threaded reply (y/yes/approve/a or
n/no/deny/d) OR fresh DM (y/n/yes/no/allow/deny — note `approve` and `a`
don't trigger fresh-DM path).
Prompts are also mailed to the `@user:` owner when email notify is on (Pro); reply by email with
approve, deny or a digit alone on the first line, above the quote. The human is told after the coordinator, once
`permission_human_delay` (default 5m) passes with the prompt still visible.
- Identity guards — 8 named enforcement checkpoints (cross_worktree, G1a/G1b
quickstart guards, non_git_bootstrap, unauthenticated_rpc,
daemon_writer_liveness, prime_ownership, dead_pid_auto_reclaim).
3 modes: strict / warn / off.
- Schema v24 (permission_nudges v21, origin_daemon col v22, daemon_identity v23,
telegram_msg_map v24).
- CLI hints (Phase B/D) — contextual hints wired on `thrum init`, `thrum send`,
`thrum tmux create`, `thrum team` (and more). Two kinds: `detection`
(condition-triggered guards; fires when condition present) and `reinforcement`
(periodic memory-aids, band-throttled by context-%). Shape B = stderr trailer
` <severity> [<code>]: <message>`; Shape C = JSON top-level `"hints"` array
with fields: code, severity, message, options[], kind. kind always present.
Suppress: --quiet or THRUM_NO_HINTS=1 drops all; THRUM_NO_REMINDERS=1 drops
reinforcement only (both stderr and JSON); --json moves hints into JSON body.
Phase D: daemon RPC responses can carry hints[] merged into top-level; AP-4
`daemon.cross-version-drift` WARN detection hint on `thrum team` (peer version
mismatch).
- Peer drift reconciliation — `peer join --type repair` for in-place re-pairing
without new peercode. `drift_reconcile_failed` shown as inline hint in peer list.
- Permission supervisors — configure via `permission_supervisors` config key
(array of role/agent/user names; default ["coordinator"]) and `project_name`
(owner of @supervisor_<project> pseudo-agent). Detection: daemon-side
SessionPoller; pattern match + stability hash scoped to bottom 15 lines of
pane capture. Claude deny key is per-shape via DisambiguateClaudeDeny:
"3" (3-option Variant A), "2" (2-option Variant B-Bash, option 2 starts with
"No"), "Escape" (2-option Variant B-Read / unknown). Snippet truncation is
tail-biased (last 5-6 lines) with ANSI strip.
- BREAKING: `peer add --type` and `peer join --type` now MANDATORY.
New in v0.9.1:
- Peercred resolver error taxonomy — introspection failures at
Step 1 (kernel peer credentials) and Step 2 (gopsutil CWD) no longer wrap
`ErrAnonymous`. They return raw errors so the server falls through to the
legacy client-asserted identity path (pre-v0.9.0 behavior). Only Step 3
(no git root above CWD) and Step 5 (CWD's git root doesn't match any
registered worktree) are PROVABLY anonymous — those still wrap
`ErrAnonymous` and trigger the allowlist-rejection error
`anonymous caller cannot invoke "<method>": cd into a registered agent
worktree and retry`.
Diagnostic path: slog.Warn fires at both introspection-failure paths with
`step=pid failed` (kernel peercred) or `step=cwd failed` (gopsutil). Grep
daemon logs for those keys to distinguish "we couldn't classify you" from
"you are provably outside every registered worktree".
Fixes 2026-04-24 breakage where claude-code Bash subprocesses on macOS hit
gopsutil.Cwd races (subprocess exits before introspection completes) and
interactive zsh callers hit the same path — both were incorrectly rejected
as anonymous despite being legitimate callers.
- thrum setup claude-md --apply (issue #8) — implemented the
documented-but-missing CLI subcommand. External users following the
quickstart hit `Error: unknown flag: --apply` because the setupCmd stub
only suggested `thrum worktree setup`. The command now installs a
Thrum-managed block (wrapped in `<!-- BEGIN THRUM -->...<!-- END THRUM -->`
markers) into ./CLAUDE.md. Three forms: bare prints to stdout, --apply
creates-or-appends, --apply --force replaces existing block idempotently.
Errors with exit 1 if a block exists and --force is not set. Template is
static (go:embed) and intentionally minimal (~30-40 lines: what Thrum
is, 5 essential commands, docs pointer). Plugin-vs-standalone framing:
use this path ONLY when not running the Claude Code Thrum plugin —
the plugin already injects messaging instructions, so combining the two
duplicates content. The CLAUDE.md block is the alternative for other
runtimes (codex, cursor, opencode) and for Claude Code
without the plugin.
New in v0.10.1:
- Hotfix for v0.10.0 quickstart-redirect regression. `thrum quickstart`
from a worktree using `.thrum/redirect` was writing the agent's
identity file to `$THRUM_HOME/.thrum/identities/` instead of the
caller's worktree, with `worktree` field also set to `$THRUM_HOME`.
Latent bug since March; surfaced because the
wizard + worktree-create flow routes through `buildInlineQuickstartCmd`
more often, which sends `thrum quickstart` into daemon-spawned panes
that inherit `THRUM_HOME` from the daemon. Fix: PersistentPreRunE
exempts `init` and `quickstart` from `THRUM_HOME` substitution;
`BuildQuickstartCmd` now accepts and emits `--repo <path>` explicitly
on the daemon-inline path; `quickstart`'s G1a/G1b guard and Step 2.5
enrichment paths drop redundant `EffectiveRepoPath` calls.
Regression tests cover local and remote quickstart. v0.10.0 is marked prerelease; upgrade to v0.10.1.
- Boot-time identity reconcile. Write RPCs (`thrum send`, `thrum tmux
start`) from a registered worktree no longer fail with `anonymous
caller cannot invoke X` after a daemon restart. The peercred
resolver matches caller CWDs against `session_refs JOIN sessions
WHERE ended_at IS NULL`; that view is durable in SQLite but loses
rows on shutdown / cleanup / long quiescence, so disk truth (identity
files) and resolver truth (DB rows) drifted apart. Previously only
`thrum quickstart --force` repopulated; `thrum prime` did not. The
daemon walks
`.thrum/identities/*.json` at boot and inserts the missing
`(sessions, session_refs)` pairs through `safedb` in a per-identity
transaction with `INSERT OR IGNORE`. Local-only — direct SQL, no
JSONL events, no cross-machine sync. Same pass also restores the
in-memory `sessionCwds` pane-nudge map for live tmux sessions.
A defensive guard skips fixture stub files, and the pass is bounded to 10s.
New in v0.10.0:
- `thrum init` interactive wizard — on a TTY, `thrum init` now launches a
multi-step wizard (identity → worktrees-root → role-template → daemon).
Press enter through every prompt to accept defaults. Legacy silent path
is preserved on non-TTY stdin or with `--non-interactive`. Pre-fill any
prompt with `--name`, `--role`, `--module`, `--worktrees-root`,
`--roles=enhanced|default|skip`, `--no-daemon`. `--force` re-init
pre-seeds prompts from existing `.thrum/config.json` and identity.
Transactional: SIGINT or step-failure restores `.gitignore` and
`.git/info/exclude` byte-for-byte.
- New role template `implementer-worktree-write-only.md` — wizard's
"enhanced" choice ships an implementer preamble that pins writes to the
agent's own worktree and forbids drive-by edits to the main repo.
- Default `Worktrees.BasePath` migrated from `~/.workspaces/<project>` to
`~/.thrum/worktrees/<project>`. Repos with an explicit `worktrees.base_path`
in `.thrum/config.json` are unaffected. Pin the legacy path with
`thrum config set worktrees.base_path "$HOME/.workspaces/<project>"` to
preserve existing worktrees.
- tmux gate — if `tmux` is not on `PATH` when the wizard reaches the
daemon-start step, init exits early with an OS-appropriate install hint.
- The inbox-check hook script is now correctly excluded alongside
the startup script in `.gitignore` (and `.git/info/exclude` in stealth
mode).
New in v0.9.2:
- Role configuration system — `/thrum:configure-roles` answers
persist under a new `role_config` top-level key in `.thrum/config.json`
(per-role autonomy + scope + rendered_hash; top-level schema_version).
`thrum roles refresh` regenerates `.thrum/role_templates/<role>.md` from
saved answers + the shipped templates embedded in the binary.
`thrum prime` surfaces three drift hints: `roles.config.migration`
(rendered templates exist but no role_config), `roles.config.schema-bump`
(shipped schema_version > saved), and `roles.config.body-diff` (shipped
body_hash != saved rendered_hash). Precedence: migration > schema-bump >
body-diff.
- SessionStart identity banner + auto-load directive — Claude Code sessions
launched via `thrum tmux create` and restarted via `thrum tmux restart`
show a pane-side identity banner and a size-aware `MUST-READ` directive
pointing at the briefing. The plugin SessionStart hook also injects
`thrum prime` output via `additionalContext` so the briefing reaches the
model even when the pane is small. Restart-snapshot framing is hoisted to
the top of `additionalContext` and rephrased as a directive.
- tmux session scoping — `thrum tmux status` and `thrum tmux
connect` now scope to the current daemon via a `@thrum-thrum-dir` tag,
no longer leaking sessions across worktrees and projects. Migration:
pre-0.9.2 sessions need to be recreated via `thrum tmux create` to appear
in pass-2 output.
- tmux pty leak fix — tmux-exec migrated from `respawn-pane`
to a persistent-session pool with flock-based coordination (portable
fallback when flock is unavailable), eliminating fd-leak crashes on
long-running daemons.
- Preamble fallback fixes — `runPreambleInit`
fallback now follows `.thrum/redirect`; rendered preamble paths to
strategy files are absolute, not relative.
## Quick Start
thrum init # Set up .thrum/ in repo
thrum daemon start # Start background daemon
thrum worktree create <agent-name> # Agent work happens in a worktree; the main repo refuses new-agent binds
cd <worktree-path> # path printed by worktree create ("Worktree created at <path>")
thrum quickstart --name <agent-name> --role planner --module auth # Register + start session (run inside the worktree)
echo 'Starting work on auth module' | thrum send --to <agent-name> - # Send a message (positional body is refused)
thrum inbox # Check messages
thrum sent # Check sent items and receipts
## Global Flags
All commands accept these flags:
--role VALUE Agent role (or THRUM_ROLE env var)
--module VALUE Agent module (or THRUM_MODULE env var)
--json JSON output for scripting
--quiet Suppress non-essential output
--verbose Debug output
## All Commands
### Setup & Orientation
thrum init [--force] [--stealth] [--runtime claude|codex|cursor|antigravity|copilot|muse|opencode|cli-only|all] [--dry-run] [--skills] [--tracker beads|native] [--agent-name NAME] [--agent-role ROLE] [--agent-module MODULE]
Initialize Thrum in the current repository. Optionally generate runtime-specific configs.
--skills installs just the thrum skill without full runtime configuration.
Populates the `identity` block in config.json (daemon_id generated once via Bootstrap; idempotent).
G2 guard: refuses non-git CWD; --force bypasses.
thrum worktree setup WORKTREE_NAME
Configure a feature worktree with .thrum/redirect to share the main issue database. Alias for `thrum worktree create`.
thrum setup claude-md [--apply] [--force]
Install or manage a Thrum-managed block in ./CLAUDE.md. Default: prints template to stdout.
--apply: writes to ./CLAUDE.md. Creates file with template-only content if missing.
Appends blank-line + template at end if file exists without a Thrum block.
Errors if a Thrum block already present (exit 1, stderr message asks for --force).
--force (with --apply): replaces the existing Thrum block in place. Idempotent.
Block markers: <!-- BEGIN THRUM --> ... <!-- END THRUM --> (single space, no version, no trailing dot).
Template is intentionally minimal (~30-40 lines): what Thrum is, 5 essential commands
(whoami/team/inbox/send/reply), quickstart pointer.
Use this ONLY when NOT running the Claude Code Thrum plugin — the plugin already
injects messaging instructions via SessionStart hook + skills + slash commands.
The CLAUDE.md path is for Claude Code without the plugin, or other runtimes (codex,
cursor, opencode). Use one or the other, not both.
thrum quickstart --name <agent-name> --role ROLE --module MODULE [--display NAME] [--intent TEXT] [--runtime PRESET] [--dry-run] [--force] [--no-init] [--preamble-file PATH]
Register agent, start session, generate runtime config, and set intent in one step.
Runtime presets: claude, codex, opencode, cursor, copilot, antigravity, muse, shell.
--runtime writes `preferred_runtime` to the identity file (used by `thrum tmux launch`).
G1a guard: refuses self-rename (caller already owns an identity in this worktree); --force bypasses.
G1b guard: refuses name collision with a live foreign PID; --force bypasses.
thrum overview [--json]
Show combined status: identity, work context, team, inbox, sync.
thrum whoami [--json]
Show current agent identity (lightweight, no daemon required).
thrum version
Show version, build hash, repository URL, and documentation URL.
thrum update [--check] [--version vX.Y.Z] [--allow-downgrade] [--restart|--no-restart] [--rollback]
Update the running binary in place to the latest release (alias: thrum upgrade).
--check prints current and latest and exits 0 current, 10 update available, 11 Homebrew tap behind, 1 error.
Homebrew install: runs 'brew upgrade --cask thrum-team/tap/thrum-pro' (--version unsupported). Otherwise runs the embedded installer (checksum verified, atomic replace, old binary kept as thrum.prev); an unwritable directory prints the sudo command, thrum never runs sudo.
Exit codes: 0 ok/current, 10 (--check) update available, 11 tap behind (nothing changed), 12 licence not accepted, 1 error. --allow-downgrade is separate from --version. A sudo hint repeats --version and --allow-downgrade. A binary not named thrum, or in the Caskroom with brew unavailable, is refused.
Never accepts the licence: exits 12 and prints THRUM_ACCEPT_EULA=<version> when the new binary reports it is not accepted.
A running daemon is restarted| – |
| 12 Nov 2026 |