aisync: Multi-Agent Context Management, Coordination & Observability

One local hub for developers working with multiple AI coding agents: share the right context, coordinate concurrent work, and verify what happened.

aisync is agent-agnostic at its core. It provides native hooks for Claude Code, Codex and Cursor, and exposes its hub to other agents through a CLI, MCP server, or shared instruction files. These three native integrations are not a limit on which agents can use aisync.

English · 中文 · Website · Demo status board

Note

Multiple coding agents should not have to work with isolated rules, conflicting file ownership, and disconnected histories. aisync is a local, verifiable hub (~/.ai-sync/, a git repo) designed for multi-agent workflows:

• Context management: share and 3-way-sync rules; load core context up front and inject relevant context on demand, including mid-session updates where hooks are available.
• Multi-agent coordination: declare file claims, detect overlapping claims (exit code 3), release ownership, and surface blockers and other agents' activity.
• Observability & verification: inspect agent activity, blockers, and health through the journal, status reports, and board; aisync verify checks the journal's hash chain for tampering.

File claims are cooperative, not OS-enforced locks: agents must use the claim protocol. Hash-chain verification detects changes to recorded events; it does not prove every claimed action occurred.

Who it's for

Developers running two or more AI coding agents across one or multiple repositories—especially when agents work concurrently, hand tasks off, or need a shared record of decisions and blockers. aisync is not another coding agent or a task orchestrator: it is the shared context and coordination layer those agents can use.

See it

The Agent Swarm Floor

aisync-agent-swarm-floor-with-cover.mp4

Integrations: native hooks are not the compatibility boundary

Integration	Who can use it	What to expect
Native hooks	Claude Code, Codex, Cursor	Automatic context and news delivery at supported hook events, plus access to aisync's coordination and journal commands
MCP server	MCP-compatible agents	Exposes rules, rule additions, logging, claims, releases, news, status and tidy; an agent must invoke the tools
CLI	Agents that can run shell commands	Direct access to hub operations such as aisync claim, aisync news, aisync log and aisync verify
Shared instruction files	Agents that read supported instruction files	Standing rules and context only; not automatic mid-session news or file-claim enforcement

Compatibility is not equivalent to automatic integration. For agents without native hooks, features depend on the agent's MCP/CLI usage and configuration. aisync does not automatically intercept all edits or guarantee every agent will honor file claims.

Install

Give it to your AI. It reads the page, does a dry run first and asks before it changes anything:

Read https://maxont.github.io/aisync/install and install aisync for me.

Or in a terminal. It installs with uv, pipx or pip and never edits your shell config:

curl -fsSL https://maxont.github.io/aisync/install.sh | bash

Needs Python 3.11+ and git, on macOS or Linux. Nothing else: aisync uses only the standard library.

Two agents, one file

_{A real terminal run: aisync demo (synthetic data), then two tools reach for the same file. Step by step: how it works · the same demo hub as a status board.}

Multi-agent need	What aisync provides
Context management	3-way-synced rules, layered rule loading and relevant mid-session news via hooks
Multi-agent coordination	Cooperative file claims and releases, overlap detection (exit 3), blockers and activity updates
Observability & verification	Hash-chained event journal, aisync verify, status, report, tail and a static board

aisync also does the rules part: one rules block copied into every tool's own rule file, synced 3-way, conflicts never overwritten.

Dogfooded: 6,700+ journal events from 7 AI tools over 4 weeks.

Why not just AGENTS.md?

Keep AGENTS.md for a repo's standing instructions. aisync sits beside it and covers what a single file cannot: the context that changes while agents work.

	AGENTS.md	aisync
Scope	Standing instructions for agents working in a repo	A shared hub across configured agents and repositories
When it is read	When a session starts	At start, and mid-session through hooks
Who is editing what	Not covered	Cooperative claims: an overlap returns exit 3 to the caller
Tamper evidence	Not provided by the file format	Hash-chained journal: aisync verify detects alterations to recorded events
Context per session	Whole file, every time	Triggered rules only: always-loaded rules 20,673 → 7,876 tokens*

_{* Measured once on one author's own hub (2026-11-08), always-loaded rules before and after layered loading. One machine, not a benchmark: your numbers will differ. aisync tidy reports your own rule budget and uncited rules.}

Use it

aisync demo            # a throwaway hub with a synthetic two-week journal
aisync init            # create ~/.ai-sync (or wherever AISYNC_HOME points)
aisync install-hooks   # shows every tool config change as a diff, then asks
aisync watch &         # event-driven sync loop (or run `aisync sync` by hand)

What an agent does during a session:

aisync claim   -a codex -p api -f app/payments/ -m "SDK upgrade"  # 3 = taken
aisync log     -a codex -p api -m "SDK 5.0, retries kept"
aisync rule add -a codex --when "changing an API" --do "add a contract test"
aisync release -a codex -p api

Command	What it does
init	create the hub: shared rules, journal, config, local git repo
log	append one event (dev_log, decision, blocked, unblocked, rule_change, note)
rule add / rule list	any agent adds a shared rule (--when/--do/--not, --topic, --replaces R-012); every tool's copy gets it on the next sync
rule lint / rule rewrite / rule enforcement	rules not in the When/Do/Not shape / rewrite them by ID without losing a fact / which test or hook keeps each rule
rule enforce	say what keeps an existing rule: rule enforce R-012 --enforced-by tests/test_x.py, --judgment-only "<why>", or --planned "<the check to write>" --size S|M|L for a check that is possible and not written yet. rule enforcement then lists every open one with its next step: an enforcer file that is gone, a plan past its due days (S 7, M 30, L 90), a rule with no plan. A plan's clock only runs: a later date, a larger size or a switch to judgment-only needs --reason and stays on record
tidy	rules clerk: likely duplicates, rules nobody cited lately, core layer over budget; proposals only, you tick them
rulecheck	prove a rules cleanup lost no hard fact (paths, numbers, commands, IDs)
instinct	optional habit list grown from past mistakes, injected into every session
gate	push-time test gate: gate check runs ./check.sh --fast on a commit; the git pre-push hook calls gate pre-push (opt-in, see below)
web-inbox	import web-Claude events from the web-inbox branch now (--no-fetch uses the last fetched copy)
guard	Claude Code hook: one session per repo; denies an edit when another session holds the repo (opt-in, see below)
lease	who holds which repo (lease status), or free one by hand (lease release --key)
wt	a session's own git worktree: wt start opens it; wt land lands it on main and needs --yes (without it: a dry run)
mcp	serve the hub to any MCP agent over stdio (pip install 'aisync[mcp]')
claim / release / claims	who is working on which files right now; overlap = exit 3
news	what other tools did since this session last looked
hook <tool>	entry point for Claude Code, Codex and Cursor hooks (reads hook JSON, never fails)
install-hooks	wire the hooks, seed the rules block, add the git Agent: trailer hook; diff first, asks before writing (--dry-run, --yes)
status	health, conflicts, claims, blockers, metrics (exit 1 when something needs you)
board	write the hub as one static HTML page: claims, blockers, who did what, timeline
report	read-only self-check: silent tools, commits without an Agent: trailer, stale blockers
tail	latest journal events
verify	check the hash chain
sync / watch	run one sync cycle / the debounced, event-driven loop
resolve / resume	rule on a rule conflict / clear failure counters after a pause
demo	the throwaway hub shown above

New in 0.2.0: rules as interfaces, loaded in layers

A rules file grows until every session pays for it and nobody reads it. 0.2.0 treats a rule like an API:

• When / Do / Not. Each rule is one line with a stable ID: [R-087] When: <trigger>. Do: <action>. Not: <forbidden>. The why goes to the journal, not the rule. rule_spec = true refuses anything else; rule lint lists misfits; rule rewrite reshapes old rules by ID and refuses a rewrite that drops a path, number or command.
• Layers. Mark a tool copy layer = "core": it carries core_topics in full and every other topic as one line. When a prompt or tool call touches a topic's keywords ([shelf]), the hook injects that topic once per session. On the author's own hub this cut the always-loaded rules from 20,673 to 7,876 tokens; replaying 7 days of transcripts, 98.9% of the 189 times a folded rule was cited, it had been injected before the citation.
• A stable prefix. progress_file moves the live progress table out of the shared page, so the page stays byte-identical and prompt caches keep hitting: the next session after a journal event read 61,151 cached tokens instead of 13,060.
• A clerk, not a janitor. aisync tidy proposes (duplicates, uncited rules, budget overruns) and never edits; a replaced rule moves to rules-archive.md with its ID, nothing is deleted.
• Any agent. Three doors: the AGENTS.md page, the CLI, and aisync mcp (tools: rules, rule_add, log, claim, release, news, status, tidy). The core stays standard-library only; the MCP SDK is the optional aisync[mcp] extra.

How it works

_{Step through it at your own pace (play, pause, ← →)}

A sync cycle is harvest (new git commits in watched repos become journal events, attributed by their Agent: trailer) → rules (new entries in repos you list in rule_sources) → distribute (3-way sync of the rules block into each tool file) → render (progress page) → flush (commit the hub, optional push). Each step is isolated: one failure never blocks the others.

The rules block is synced 3-way. If only the truth changed, copies are overwritten (with a backup). If one copy was edited and nobody else changed anything, the edit is adopted everywhere. If both sides changed, nothing is overwritten: it becomes a conflict for you to resolve.

Reliability and observability: timeouts and retries with backoff on every git call, failure isolation per step and per project, auto-pause after fail_threshold failures in a row with an ALERT.md and a desktop notification, a per-cycle run log (state/runs.jsonl), hook call stats and aisync status for all of it.

Configuration

~/.ai-sync/config.toml (set AISYNC_HOME to put the hub anywhere):

roster = ["claude-code", "codex", "cursor"]   # tools `report` expects to hear from
scan_dirs = ["~/projects"]                    # every git repo inside is watched
rule_sources = []                             # repos whose AGENTS.md "## Meta rule" feeds shared rules
remote = ""                                   # e.g. "origin" to also push the hub

progress_file = "progress.md"                 # keep agents.md byte-stable; hooks inject progress at session start
rule_spec = true                              # rule add only accepts When/Do/Not rules
require_enforcer = false                      # true: every new rule names --enforced-by or --judgment-only
core_topics = ["Safety", "Meta"]              # topic headings a layered copy carries in full
instincts_path = ""                           # optional habit file injected at session start
tidy_daily = false                            # the watcher starts one `aisync tidy --write` a day
tidy_usage_globs = []                         # files whose R-### mentions count as a rule being used
gate_hook = false                            # opt-in: the git pre-push hook runs the test gate (gate.py)
guard_hook = false                           # opt-in: Claude Code runs the one-session-per-repo guard (guard.py)
web_inbox_secs = 0                           # 0 = off; else poll the web-inbox branch this often

[shelf]                                       # folded topic -> keywords that pull it into a session
Deploys = ["deploy", "terraform", "rollback"]

[[target]]                                    # files carrying the shared-rules block
label = "claude-code"
path = "~/.claude/CLAUDE.md"
layer = "core"                                # optional: fold non-core topics, inject them on demand

To keep aisync watch running, start it from your login items, a systemd user unit or a launchd agent that runs aisync watch; create ~/.ai-sync/STOP to stop it cleanly.

Opt-in features (all off by default). gate_hook, guard_hook and web_inbox_secs change behavior outside the hub, so nothing turns on until you set them. The guard can deny edits: if a session is locked out, run aisync lease status, and once the holder session has ended, aisync lease release --key <repo>. To undo the guard, delete the PreToolUse and SessionEnd entries whose command ends in guard from ~/.claude/settings.json.

Standing on the shoulders of

aisync: Multi-Agent Context Management, Coordination & Observability. AGENTS.md was the starting idea.

Project	How it shaped aisync
	The starting idea: one plain-text instruction file every agent reads, no SDK. aisync keeps that shape for shared rules (one rules block copied into each tool's own AGENTS.md / CLAUDE.md) and adds what a static file cannot say: who is editing what right now, and what changed since the session started.
	Mid-session news rides on Claude Code hooks: aisync install-hooks adds SessionStart, UserPromptSubmit and PostToolUse entries to ~/.claude/settings.json, and aisync hook claude-code answers with hookSpecificOutput.additionalContext only when another tool changed something.
	Codex reads ~/.codex/AGENTS.md, which is where aisync seeds the shared rules block, and its hooks.json takes the same SessionStart / UserPromptSubmit / PostToolUse events, so aisync hook codex reuses the Claude Code reply format.
Cursor	Cursor hooks (~/.cursor/hooks.json) deliver the rules and progress at sessionStart and the news after each postToolUse through additional_context, because Cursor's User Rules live inside the app instead of a file aisync could sync.
	The journal borrows git's hash chain: every event stores the id of the one before it and its own id is a SHA-256 of that link plus its body, so aisync verify spots any edited, lost or reordered line. The hub itself is a local git repo.

aisync is an independent project, not affiliated with the projects above.

Development

python3 -m unittest discover -s tests   # includes wheel -> venv -> demo

The images are generated: the terminal GIF is a recorded real run of .github/refine/demo-terminal.json, the how-it-works animation comes from .github/refine/how-it-works.stages.json, and the credits above from .github/refine/lineage.json. All demo data is synthetic.

MIT licensed.

More Videos if you like!

aisync-agent-swarm-docking-18s.mp4

aisync-agent-swarm-signal-18s.mp4

aisync-agent-swarm-cathedral-40s.mp4

aisync-agent-swarm-pneumatic-18s.mp4