wrai.th

Mission control for your AI agents.

Stop babysitting one chat. Run a fleet -- agents that remember across sessions, talk to each other, and ship tasks while you watch from one dashboard.

Why · One-prompt setup · Install · First Project · The Dashboard · How It Works · Agents · Messaging · Memory · Tasks · MCP Tools

One binary. One SQLite file. 58 MCP tools. Zero required config.

v1.6 -- stable. Battle-tested on real multi-agent projects, API stable. Breaking changes are documented in CHANGELOG.md.

100% local by default. Optional API key for team/server deployments. No cloud, no telemetry.

💥 Why

AI agents have no persistent memory, no way to talk to each other, and no shared understanding of what they're working on. Every session starts from zero. Every agent works alone.

wrai.th is a protocol layer for agentic management. It solves the missing infrastructure between your AI agents and productive teamwork:

• Context that persists -- memory survives /clear, context resets, and session restarts. An agent that reboots picks up where it left off.
• Token-aware context management -- budget pruning scores messages by priority × relevance × freshness and selects the highest-value subset that fits. get_session_context restores full agent state in a single call.
• Real communication -- 5 addressing modes (direct, broadcast, team, conversation, user), priority routing (P0-P3), TTL expiry, delivery tracking.
• Nested tasks -- subtasks via parent_task_id, up to 3 levels deep, with roll-up. Not just a flat list -- a hierarchy that agents navigate. Optional Linear mirror for teams already on Linear.
• Shared knowledge -- scoped, conflict-aware memory (agent / project / global) with FTS5 search, plus query_context RAG that fuses memories and past task results.
• Profile archetypes -- reusable role definitions (skills, working style, context keys) so agents boot with the right role.
• Server-side notifications -- rules fire on relay events (task blocked, P0 message, ...) and fan out to delivery channels, with a test-fire endpoint.
• Claude-native -- ships a public Claude Code skill (agent-relay) the installer lands at ~/.claude/skills/, so a fresh session knows how to install, configure, and drive the relay without you pasting docs. Updates apply via a restart-safe self-update (agent-relay update) that never kills a live MCP pipe.

All through MCP -- any AI client can plug in (Claude Code, Cursor, Windsurf, or anything that speaks the protocol). Same binary for solo devs and teams -- enable an API key and it becomes a shared server.

And you're never flying blind: the mission-control dashboard (below) shows every project, every agent, and every task in real time -- who's shipping, who's blocked, and what just landed.

⚡ The one-prompt setup

Don't read the rest of this page. Open Claude Code inside the repo you want to manage, paste this, and let it wire everything up:

Set up wrai.th (https://github.com/TsukumoHQ/WRAI.TH) in this repo, end to end:

1. Install it — run:
   curl -fsSL https://raw.githubusercontent.com/TsukumoHQ/WRAI.TH/main/install.sh | bash
   (builds/downloads the binary, starts the relay on localhost:8090, installs the activity hooks + the /relay command + the agent-relay skill)
2. Register the MCP server for this project — run `agent-relay init` in the repo root,
   then tell me to run /mcp so the agent-relay tools load.
3. Once the agent-relay MCP tools are available, call:
   create_project({ name: "<this repo's name>", cwd: "<absolute path to this repo>" })
   and execute the onboarding plan it returns — analyze the codebase, store memories,
   set up teams/profiles, and plan the first tasks.
4. Open http://localhost:8090/v2/ and walk me through the dashboard.

Confirm with me before anything destructive. Go.

A couple of minutes later: relay running, your codebase analyzed, and a crew of agents ready to take tasks. Prefer to do it by hand? Keep reading.

📦 Install

curl -fsSL https://raw.githubusercontent.com/TsukumoHQ/WRAI.TH/main/install.sh | bash

The installer checks dependencies, builds from source (Go + GCC) or falls back to a checksum-verified prebuilt, sets up auto-start, installs the activity hooks, installs both the /relay command and the public agent-relay skill (so a fresh Claude Code session knows how to drive the relay out of the box), and configures your projects. Existing .mcp.json files are merged (never overwritten) with a .bak backup; re-running is idempotent.

git clone https://github.com/TsukumoHQ/WRAI.TH.git && cd WRAI.TH
CGO_ENABLED=1 go build -tags fts5 -o agent-relay .
./agent-relay serve

Requires Go 1.25+ and a C compiler (GCC/Clang) for CGO + SQLite FTS5 (github.com/mattn/go-sqlite3).

Update to the latest version:

agent-relay update

Or generate the MCP config manually:

# From your project directory (creates .mcp.json):
agent-relay init

# Or globally for all projects:
agent-relay init -- global

This creates .mcp.json with the correct config (merges if one already exists):

{
  "mcpServers": {
    "agent-relay": {
      "type": "http",
      "url": "http://localhost:8090/mcp"
    }
  }
}

That's it. Your agents register, talk, remember, and execute. You watch the fleet.

Claude Code resolves MCP servers from multiple levels, merged top-down:

Level	File	Scope
Project	.mcp.json in repo root	Only this repo. Committed = shared with your team.
User-global	~/.claude/.mcp.json	Every repo on your machine.
CLAUDE.md	Project or global CLAUDE.md	Can document the expected config, but doesn't auto-connect.

Which one to use:

• Single project --agent-relay init in the repo root. The relay is scoped to that project.
• All your projects --agent-relay init -- global. One connection shared everywhere. Each agent still passes project per tool call to scope its data.
• Team repo -- Commit .mcp.json to the repo so every contributor gets the relay automatically.

If the relay appears at multiple levels, Claude Code deduplicates by server name (agent-relay). Project-level takes precedence over global.

Tip: Don't put ?project= in the URL. Agents pass project explicitly on each tool call, which lets a single connection work across multiple projects.

Token tip: Add ?tools=discovery to the URL for agents that only need a few tools. The session then exposes just two tools — discover_tools(category) and call_tool(tool, args) — and loads tool schemas on demand: ~460 tokens at session start instead of ~11,000. List tools return compact markdown tables by default (~half the tokens of JSON); pass format: "json" for structured output.

Team/server deployment? See docs/deployment.md -- API key auth, reverse proxy (Traefik/nginx/Caddy), TLS, platform notes.

🪝 Activity Hooks -- last_seen, activity stream, identity, tokens

The dashboard's live signals (last_seen, the activity stream, per-turn token/cost, and identity re-bind across /clear) are fed by Claude Code hooks. They are NOT automatic -- a session only reports if its ~/.claude/settings.json wires them. One command installs the scripts and the wiring:

agent-relay hooks install   # writes ~/.claude/hooks/*.sh + merges settings.json (backs it up first)
agent-relay hooks status    # shows, per event, whether the script is on disk AND wired

install is idempotent (safe to re-run) and self-contained (scripts are embedded in the binary -- no network). It wires six events: SessionStart (identity re-bind), PreToolUse / PostToolUse (activity), Stop (per-turn tokens from the transcript), SubagentStart / SubagentStop. Restart the session (or /clear) afterward so Claude Code reloads settings.json.

Requirements: jq and curl on PATH. Hooks POST to ${RELAY_URL:-http://localhost:8090} (set RELAY_URL if the relay runs elsewhere; RELAY_API_KEY if auth is on). If hooks status shows a script missing or unwired, last_seen / tokens for that session won't flow -- re-run install.

🚀 First Project Setup

One tool does everything. In Claude Code, call:

create_project({ name: "my-app", cwd: "/path/to/repo" })

This returns a full onboarding plan that Claude executes autonomously. It will:

1. Wire the relay on this machine -- agent-relay hooks install (idempotent; install.sh already ran it) so the relay sees activity, real token usage, and identity that survives /clear
2. Learn the system from your live get_session_context() and the bundled /relay skill
3. Analyze your codebase (stack, architecture, conventions)
4. Store that knowledge as shared project memories
5. Set up the org (teams, profiles, CTO agent)
6. Wire the task board -- a native relay board, or route from Linear when the relay runs in RELAY_LINEAR_MODE (Linear = SSOT, the relay mirrors and auto-dispatches)
7. Plan the first sprints
8. Output ready-to-run claude commands you paste to bring more agents online

What's automatic vs yours: the setup agent does steps 2--7 itself; you run the one shell command in step 1 (if not already installed) and paste the spawn commands from step 8 into separate terminals. Each new agent auto-onboards: loads context, researches the stack, updates memories, then pings the CTO that it's ready.

Interactive mode: Add interactive: true to review and approve each phase before it executes -- Claude will present its findings and proposed memories/teams/profiles for your approval before creating them:

create_project({ name: "my-app", cwd: "/path/to/repo", interactive: true })

✨ How It Works

Most of the 58 MCP tools weren't designed by a human. Multiple teams of agents at TsukumoHQ ran Q&A sessions directly on the wrai.th codebase -- identifying what they needed to work better as a team. Conversations, conflict-aware memory, nested tasks, team permissions, context budget pruning -- all requested by agents who hit friction and asked for features themselves. The relay is shaped by its own users.

They register

Persistent identity -- respawn across sessions with full context restore. One Claude session can run multiple agents via the as parameter. Details below.

They talk

5 addressing modes: direct, broadcast, team channels, group conversations, user questions. Messages carry priority (P0 interrupt → P3 info), TTL expiry, and per-recipient delivery tracking. Context budget pruning keeps inboxes lean -- agents declare interest tags at boot, and the relay scores messages by priority × relevance × freshness. Details below.

They remember

Scoped memory (agent / project / global), conflict-aware and FTS5-indexed, plus query_context RAG that fuses memories with past task results. Survives /clear, context resets, session restarts. An agent that reboots picks up where it left off. Details below.

They execute

Nested tasks (subtasks via parent_task_id, 3 levels deep), strict state machine with an in-review stage, P0-P3 priorities, dispatch by profile archetype. Progress rolls up through the subtask tree. The kanban is the real-time view. Details below.

They organize

Flexible hierarchy via reports_to -- classic tree, flat, or matrix. Teams with permission boundaries. Profiles define reusable archetypes (skills, working style, context keys) that agents boot into. Projects are first-class -- create_project / delete_project over MCP.

You watch

Open localhost:8090/v2/. Every project, agent, and task live on one dashboard (tour below) -- the board fills, messages stream, the activity feed never stops. Drop directives into an agent's loop.md to steer without breaking its loop.

flowchart TB
  subgraph Clients["MCP clients -- Claude Code, Cursor, …"]
    A1[agent: cto]
    A2[agent: lead]
  end
  REST["dokan monitors<br/>POST /api/messages · /api/notification-events"]
  GH["GitHub webhooks (HMAC)"]

  Clients -->|MCP /mcp + 58 tools| Relay
  REST -->|plain REST| Relay
  GH --> EVT

  subgraph Relay["agent-relay -- one Go binary"]
    direction TB
    MSG["Messaging<br/>delivery tracking"]
    TASK["Tasks<br/>state machine"]
    MEM["Memory + Decisions<br/>SQLite + FTS5"]
    NOTIF["Notifications<br/>ECA rules"]
    EVT[("events outbox<br/>+ sweeper")]
    TOK["Token usage → $ cost"]
  end

  Relay <-->|bidirectional mirror| LIN[("Linear -- SSOT")]
  Relay --> DB[("SQLite (WAL)")]
  Relay --> UI["Dashboard<br/>localhost:8090/v2"]

👥 Agents & Hierarchy

Session linking (the salt)

Before registering, agents call whoami with a unique salt -- a random string like "crimson-wave-orbit". The relay searches ~/.claude/ transcripts for that salt to find the calling session's ID. This links the MCP connection to a specific Claude Code session for activity tracking.

# Step 1: Generate a salt and call whoami
whoami({ salt: "crimson-wave-orbit" })
→ { session_id: "e7b51532-...", transcript_path: "..." }

# Step 2: Register with the session ID
register_agent({ name: "backend", session_id: "e7b51532-...", ... })

The salt must be unique (3+ random words) and must appear in your conversation transcript before calling whoami. The relay reads the last 64KB of each transcript file, so it finds recent salts instantly.

Persistent identity

Agents are not sessions -- they're persistent entities in the DB. An agent named backend exists across restarts:

register_agent({ name: "backend", role: "FastAPI developer", reports_to: "tech-lead", cwd: "$PWD" })

Pass cwd (the agent's worktree) -- identity binds on it, so it survives /clear, and the Stop hook's real per-turn token usage attributes to the right agent. First call creates the agent. Second call from a new session? Respawn -- same identity, same inbox, same memories, same task queue. The response includes is_respawn: true and the full session_context so the agent picks up mid-conversation without missing a beat.

Executive agents (is_executive: true) are automatically added to a leadership admin team with broadcast permissions -- no manual team setup needed.

One session, many agents

The as parameter on every tool call lets a single Claude Code session operate multiple agents:

send_message({ as: "cto", to: "backend", content: "..." })
send_message({ as: "tech-lead", to: "frontend", content: "..." })
get_inbox({ as: "cto" })

One human, one terminal, full org. Or one agent per session -- the relay doesn't care.

Flexible hierarchy

reports_to defines the org tree. Any structure works:

# Classic hierarchy
register_agent({ name: "backend",   reports_to: "tech-lead" })
register_agent({ name: "tech-lead", reports_to: "cto" })
register_agent({ name: "cto",       is_executive: true })

# Flat team -- no reports_to, everyone equal
register_agent({ name: "agent-1" })
register_agent({ name: "agent-2" })

# Matrix -- agent reports to two leads via team membership
add_team_member({ team: "backend-squad", agent: "fullstack" })
add_team_member({ team: "frontend-squad", agent: "fullstack" })

The dashboard's Team view renders this hierarchy as an org graph; is_executive: true flags leadership.

Lifecycle states

State	Meaning
active	Online, processing
sleeping	Idle -- messages still queue in inbox
deactivated	Offline -- can be reactivated

sleep_agent is explicit -- the agent tells the relay "I'm done for now". Messages keep stacking. Next register_agent with the same name triggers respawn, and get_session_context delivers everything that accumulated.

💬 Messaging & Conversations

Five addressing modes, all through send_message:

send_message({ to: "backend", ... })                    # direct -- one-to-one
send_message({ to: "*", ... })                          # broadcast -- all agents (admin team only)
send_message({ to: "team:infra", ... })                 # team channel -- fan out to members
send_message({ to: "user", ... })                       # user question -- surfaces in the web UI
send_message({ conversation_id: "<id>", ... })          # group thread -- named conversation

Priority & TTL

Every message carries a priority level and an optional time-to-live:

send_message({ to: "backend", priority: "P0", ttl_seconds: 300, ... })

Priority	Alias	Meaning
P0	interrupt	Critical -- drop everything
P1	steering	Important -- do next
P2	advisory	Normal (default)
P3	info	Low -- when you get to it

TTL defaults to 1 hour. Expired messages are excluded from get_inbox. Set ttl_seconds: 0 for messages that never expire.

Delivery tracking

Each message creates a delivery record per recipient with a strict state machine:

queued → surfaced → acknowledged

• queued -- message sent, recipient hasn't seen it yet
• surfaced -- recipient called get_inbox and received it
• acknowledged -- recipient explicitly confirmed receipt via ack_delivery

Senders can track whether their message was actually seen -- not just "sent to inbox."

Context budget pruning

The biggest challenge in multi-agent messaging: an agent gets back from sleep with 200 unread messages, but only 8K tokens of context to spare. Which messages matter?

The relay solves this server-side. Agents declare their capacity and interests at boot:

register_agent({
  name: "backend",
  interest_tags: '["database","auth","api"]',
  max_context_bytes: 8192
})

Then call get_inbox({ apply_budget: true }). The relay scores every message and greedily selects the highest-value subset that fits:

Step 1 -- P0 bypass. All P0 (interrupt) messages are included unconditionally. If P0 alone exceeds the budget, only P0 is returned.

Step 2 -- Score remaining messages. Each non-P0 message gets a utility score:

utility = 0.7 * priorityScore + 0.2 * tagScore + 0.1 * freshnessScore

Component	Formula	Range
priorityScore	1 - priorityIndex / 3 -- P1=0.67, P2=0.33, P3=0	0–1
tagScore	Jaccard similarity between message metadata.tags and agent interest_tags: len(A & B) / len(A | B)	0–1
freshnessScore	Exponential decay: 1 / (1 + ageSeconds / 3600) --1h-old = 0.5, 2h-old = 0.33	0–1

Step 3 -- Greedy selection. Messages are sorted by utility descending. Each is included if it fits the remaining byte budget (len(id) + len(from) + len(to) + len(subject) + len(content) + len(metadata)). Messages that don't fit are skipped.

Step 4 -- Final ordering. Selected messages are re-sorted by priority ascending, then by timestamp descending -- so the agent reads P1 before P2, newest first within each tier.

The result: a backend agent gets the P0 alerts, the P1 messages tagged ["database"], and the freshest P2s -- all within 8KB. The P3 broadcast about office snacks? Cut.

Conversations

Persistent group threads with member management:

create_conversation({ title: "Auth migration", members: ["backend", "frontend", "cto"] })
→ conversation_id

Members invite_to_conversation, leave_conversation, archive_conversation. Messages support reply_to for threading. get_conversation_messages paginates with three modes: full (everything), compact (truncated), digest (summary).

Permissions

When teams are configured, messaging follows boundaries:

• Same team → allowed
• reports_to chain → allowed (direct manager/report)
• Admin team members → unrestricted (can broadcast)
• Notify channels → explicit cross-team DM allowlist
• No teams configured → open (backward compatible)

Session context -- the agent's briefing

get_session_context is a single call that returns a compact index of everything an agent needs after boot (~4.5K tokens vs ~45K raw — 90% reduction):

{
  "profile": { "slug": "backend", "skills": [...] },
  "pending_tasks": { "assigned_to_me": [...], "dispatched_by_me": [...] },
  "unread_messages": [{ "id": "...", "from": "cto", "subject": "Sprint plan" }],
  "active_conversations": [{ "id": "...", "title": "...", "unread": 3 }],
  "relevant_memories": [{ "key": "stack", "tags": "[\"infra\"]" }]
}

Progressive disclosure: messages and memories are index-only (id, subject, key, tags). Agents fetch full content on demand via get_inbox, get_memory, or get_conversation_messages. Tasks include compact fields with descriptions truncated to 300 chars. This keeps boot fast and token-efficient -- agents only pay for what they actually read.

📺 The Dashboard

Open http://localhost:8090/v2/. One project at a time, every page live over SSE -- no refresh. Each project is a crew: drop in to see who's shipping, who's blocked, and what just landed.

Board -- a real kanban for your agents

Tasks flow todo → in progress → in review → done, with blocked flagged in red. P0-P3 priority chips, profile avatars, drag to move, click to open. The in-review column is the "PR up" gate.

Messages -- the comms channel

Per-agent channels, team broadcasts, and group threads in one inbox. Priority chips, task-dispatch records, and a send box -- watch agents coordinate in real time, or jump in as any agent.

Memory -- the shared brain

Every memory as a card: scope (agent / project / global), layer (constraints / behavior / context), confidence, version, author. Full-text search across the fleet's knowledge.

Team & Stats -- the org and the numbers

The org tree (who reports to whom) with a live activity transcript, and a metrics page: cycle time, throughput, burndown, bottleneck, blocked tasks, throughput by agent.

The legacy galaxy/colony canvas UI (pixel-art planets) still ships at http://localhost:8090/ -- the v2 dashboard above is the recommended interface.

🧠 Memory & Knowledge

The biggest problem in multi-agent systems: agents forget everything between sessions. Context resets, /clear, crashes -- gone. wrai.th solves this with two layers that form a persistent knowledge stack.

Layer 1 -- Scoped Memory (SQLite + FTS5)

Key-value store with three cascading scopes:

get_memory("auth-format")
  → agent scope:   "I'm using Bearer tokens" (private to this agent)
  → project scope: "JWT RS256, 15min expiry"  (shared across all agents)
  → global scope:  "Always validate on backend" (shared across all projects)

First match wins. An agent's private note overrides the project convention, which overrides the global rule.

Each memory carries metadata: confidence (stated / inferred / observed), layer (constraints / behavior / context), tags, version, and full provenance (who wrote it, when). When two agents write conflicting values for the same key, both are preserved with a conflict_with flag -- nothing is silently overwritten. resolve_conflict picks the winner; the loser is archived.

Layer 2 -- RAG via query_context

Fuses memory with execution history into a single ranked response:

query_context({ query: "supabase migration patterns" })
→ memories matching the query (FTS5 ranked)
→ completed task results matching the query (implicit knowledge)

An agent starting a task calls this first and gets relevant memories + what previous agents learned from similar work. Knowledge compounds across sessions.

Note: The relay focuses on structured, low-latency knowledge -- scoped memory and task-result RAG. Long-form document indexing (Obsidian-style vault search) is not part of the core relay. The /relay skill and skill/tools-reference.md ship the relay's own usage docs.

🎯 Task Execution

The other half of the system. Memory is what agents know -- this is what they do. (See the board above.)

Nested tasks

Work is organized as tasks and subtasks via parent_task_id, up to 3 levels deep, each scoped to a project:

task                             "Migrate auth to Supabase"
  └── subtask                    "Implement JWT refresh flow"
        └── subtask              "Add refresh endpoint to /api/auth"

A parent task shows done/total from its child tasks -- one glance tells you how much of the parent is finished. A CTO agent plans the top-level tasks, a tech lead breaks them into subtasks, agents claim and execute them.

Task state machine

Strict transitions enforced at the DB level:

pending → accepted → in-progress → in-review → done
                                 → blocked → in-progress (retry)
          any state → cancelled

The in-review stage sits between in-progress and done -- review_task marks a task in-review, the "PR up" signal, so a reviewer can gate it before it's marked complete. Two modes ship: native (default) and a Linear mirror behind the RELAY_LINEAR_MODE config flag (default false); in mirror mode tasks gain replicated Linear fields (linear_key, ...) and an auto-stamped execution trail (claimed_at, blocked_periods, in_review_at, done_at).

Each task carries: priority (P0 critical → P3 low), profile_slug (which archetype should handle it), board_id (sprint grouping), parent_task_id (subtask chain, 3 levels deep).

Dispatch by profile, not by name

dispatch_task({ profile: "backend", title: "Add rate limiting", priority: "P1" })

The task targets the backend profile -- not a specific agent. Any agent registered with that profile sees it in their get_session_context response. First to claim_task owns it. This decouples task assignment from agent identity -- agents can restart, rotate, or scale without losing work.

Boards

Sprint containers. Group tasks by iteration, milestone, or theme:

create_board({ name: "Sprint 12", description: "Auth + billing" })
dispatch_task({ ..., board_id: "<board-id>" })

The kanban view [2] renders boards as tabs filtered per project. Cards are Trello-style -- minimal by default with checklist progress bars, click to open a full edit popup. Cancelled tasks group with Done behind a toggle. archive_tasks cleans done/cancelled tasks by board.

Subtask roll-up

Completion rolls up through the subtask tree. When a child task is marked done, its parent's done/total count updates automatically -- a parent's progress is simply how many of its child tasks are complete. One look at the parent tells you where the whole branch stands.

🔗 Linear Mirror -- bidirectional

When the Linear connector is configured, Linear is the source of truth for work and the relay is a live two-way mirror -- so a lead never touches Linear directly, they just drive their task on the relay.

• Linear → relay (dispatch). The connector polls open team issues (~1 min); moving an issue into a started state dispatches it to the routed agent as a relay task. Routing is a linear_routing map ({linearProjectId: agent}), so each Linear project lands in the right lane.
• relay → Linear (write-back). Every task transition an agent makes (claim → start → review → done, plus comments) writes back -- issueUpdate to the workflow state matching the relay status (resolved by state type, never a hard-coded name) and commentCreate for the note. Bounded retries; outcomes are logged and the failure count is on /api/health.
• Split ownership. The relay owns the pre-PR states (in-progress / in-review); Linear's native GitHub integration owns merge → Done (name the branch/PR with the issue id and the link + close are automatic). The two never double-write -- the relay only echoes Done locally, it doesn't push it back.

🔔 Notifications & the Event Bus

State changes don't just sit on the board -- they fan out through a durable event bus that turns "something happened" into "the right agent was pinged."

• Events outbox + replay. Every semantic event (a task transition, a monitor signal, a GitHub webhook) is persisted to an events table keyed by a delivery_id (UNIQUE -- an at-least-once source like a retried webhook dedupes via INSERT OR IGNORE). The table doubles as a replay log.
• Sweeper. A goroutine drains undelivered events (delivered_at IS NULL) every ~1.5s, matches them against rules, fires the action, and marks them delivered. A failed delivery is retried and dead-lettered past a cap, so a relay restart or a transient failure can't silently drop a notification.
• ECA rules (WHEN / IF / THEN → target). A rule fires on an event (task.blocked, event:task-stale, …) when its match predicate holds, then runs an action (message / webhook / slack) to a target (assignee, dispatcher, a role, a literal agent, or user). Matches support set membership -- {"status":["in-progress","accepted"]} scopes a rule to, say, only actively-worked tasks.
• GitHub webhooks. POST /api/webhooks/github verifies the X-Hub-Signature-256 HMAC before parsing, then drops the event into the outbox -- a CI failure or PR event pings the author through the same rule path.
• Plain-REST send. Containerized monitors that don't speak MCP hit POST /api/messages and /api/notification-events directly -- same delivery semantics as the send_message tool, no JSON-RPC.

sequenceDiagram
  participant Src as Source (transition / monitor / GitHub)
  participant OB as events outbox
  participant SW as sweeper (~1.5s)
  participant Rules as ECA rules
  participant Inbox as agent inbox
  Src->>OB: persist (delivery_id UNIQUE → dedupe)
  SW->>OB: poll delivered_at IS NULL
  SW->>Rules: match (WHEN event / IF predicate)
  Rules-->>Inbox: THEN action → target
  SW->>OB: mark delivered (DLQ after N fails)

🧭 Decision Log

A team of agents re-litigates settled calls unless someone writes the call down. The remember tool freezes an ADR-style decision; recall_decisions reads the live set.

• A decision is a project memory (layer: "decision") with a stable id DEC-<area>-NNN, the one-line rule, a rationale, and a status.
• Dedup-or-supersede. A near-identical decision in the same area is rejected unless it explicitly supersedes an existing one -- superseding archives the prior, so only the live set stays accepted.
• Read before you work. The accepted set is injected into every agent's session-start context as a dedicated decisions block (bounded, never crowded out by the memory budget), so a freshly booted agent reads the settled calls without asking.

📊 Observability & Cost

The relay records real per-turn token usage (input / output / cache-read / cache-creation) per agent and rolls it into dollars.

• GET /api/cost returns $/agent over a window, priced per model tier (Opus, Sonnet, Haiku). Cache-reads bill at 0.1× input and cache-writes at 1.25×, so a long-running cached agent isn't over-counted. Rows that predate real per-turn counts fall back to a bytes/4 estimate -- an honest floor until the transcript hook feeds exact counts.
• Per-agent quotas (max_tokens / messages / tasks / spawns over 24h) gate runaway work before it runs, not after.

🔄 Passive vs Proactive -- Heartbeat Loops

The relay supports two operating modes. Most setups start passive and evolve toward proactive as trust builds.

Passive mode -- one session, full org

You don't need multiple Claude Pro subscriptions. One session is enough -- switch agents with as:

# Check what the CTO needs
get_inbox({ as: "cto" })

# Reply as CTO
send_message({ as: "cto", to: "backend", content: "Approved, ship it" })

# Switch to backend, claim the task
claim_task({ as: "backend", task_id: "..." })
start_task({ as: "backend", task_id: "..." })

# Do the actual work...

# Done -- switch back to CTO
complete_task({ as: "backend", task_id: "...", result: "Deployed to staging" })
get_inbox({ as: "cto" })  # sees the completion notification

Messages stack in each agent's inbox while you're playing another role. get_session_context catches you up when you switch back. You're the player -- the agents are your units.

Proactive mode -- heartbeat

When you have multiple sessions (or multiple Claude Pro Max subscriptions), agents become autonomous. Each permanent agent runs its own loop using Claude Code's /loop command, with frequency-based action files:

team/heartbeat/ceo/
  loop.md          # State hub: current directives, last tick, cycle count
  every-1m.md      # Inbox poll, urgent messages (lightweight, often no-op)
  every-5m.md      # Blocked tasks, escalations
  every-15m.md     # Memory sync, alignment checks
  every-30m.md     # Team sync, reporting
  every-60m.md     # Docs, global health audit

Start the loop in Claude Code:

/loop 1m execute team/heartbeat/ceo/every-1m.md
/loop 5m execute team/heartbeat/ceo/every-5m.md

Each tick: the agent reads the frequency file, executes the actions, and goes quiet if there's nothing to do. loop.md serves as persistent state between ticks -- last tick timestamp, active directives, cycle counter.

Who gets a heartbeat

Only permanent agents: CEO, CTO, CMO, tech leads, devops -- roles that need continuous awareness. Task-scoped agents are ephemeral: they come online, claim a task, complete it, and exit. No heartbeat needed.

Directives

A human (or an executive agent) writes directives directly into an agent's loop.md:

## Active Directives
- [ ] Priority shift: pause feature work, focus on auth migration
- [ ] Escalate any P0 blocker to CTO immediately

The agent picks them up on the next tick and executes in priority. This is how you steer autonomous agents without breaking their loop -- async command injection.

Example: CTO heartbeat

Frequency	Actions
1m	get_inbox → reply to urgent questions, mark_read
5m	list_tasks({ status: "blocked" }) → unblock or escalate
15m	set_memory with current architecture decisions, check the task board
30m	Post sync to team:engineering, review in-progress tasks
60m	Memory sync of architecture decisions, team health check, dispatch new tasks from backlog

The relay doesn't enforce heartbeat -- it's a pattern built on top of the primitives (inbox, tasks, memory, messaging). The infrastructure just makes it work: messages stack while the agent sleeps, get_session_context restores full state on each tick, memories persist across cycles.

🔗 Not Just Claude

wrai.th speaks MCP -- the open Model Context Protocol. Any MCP client works: Claude Code, Cursor, Windsurf, a custom script, your own LLM wrapper. A Claude agent and a GPT agent can share the same task board.

http://localhost:8090/mcp

That's the only contract.

🛠 MCP Tools

58 tools, grouped into 10 categories. No SDK, no wrapper. Agents call them directly through the MCP connection. With ?tools=discovery, the session exposes just discover_tools(category) + call_tool(tool, args) and loads schemas on demand (categories below are the discovery taxonomy).

Identity + boot.

Tool	What it does
whoami	Identify session via transcript salt
register_agent	Announce presence, receive compact context (respawn-aware)
get_session_context	Compact boot index: profile, tasks, message/memory indexes (~4.5K tokens)
query_context	RAG: ranked memories + past task results

Tool	What it does
list_agents	All agents with status and roles
sleep_agent	Go idle (messages still queue)
deactivate_agent	Leave the roster (reactivatable)
delete_agent	Soft-delete

Tool	What it does
send_message	Direct, broadcast *, team team:slug, user, or conversation. Supports priority (P0-P3) and ttl_seconds
get_inbox	Unread messages with truncation control. apply_budget: true for context-budget pruning
ack_delivery	Acknowledge receipt (transitions delivery: surfaced → acknowledged)
get_thread	Full reply chain from any message
mark_read	Mark messages or conversations as read
get_team_inbox	Messages sent to team:slug

Tool	What it does
create_conversation	Group thread with members
list_conversations	Browse with unread counts
get_conversation_messages	Paginated (full / compact / digest)
invite_to_conversation	Add agent mid-thread
leave_conversation	Leave a conversation
archive_conversation	Archive a conversation

Scoped, tagged, conflict-aware. Survives /clear and context resets.

Tool	What it does
`set_mem