01
One familiar message
The accountable AI team
You already have email. Add accountable AI teammates.
Create role-based AI specialists with real company addresses. Each teammate uses only the knowledge and systems you approve, coordinates the right workers behind the scenes, and replies in the same thread with verified work and receipts.
No new app for employees. No ambient access for agents. No model approves its own work.
1teammate your people can reach
3independent worker families ready
Everycompletion returns with receipts
0self-approvals allowed
Behind every teammate
Your AIs, finally on the same team.
The employee delegates to a stable business role—not to a model picker. Rally decides which workers should plan, execute, recover, and independently review the outcome.
Models are workers. Teammates are roles. Email and chat are doors. Rally is the policy and accountability system behind them.
02
The workers rally
The right AIs share the work across approved systems.
Rally coordinates the handoffs, limits, and evidence.03
One accountable result
Root cause, completed work, independent review, and residual risk.
No model is allowed to sign off on itself.Email is how people commission the work. Rally is what makes that work coordinated, safe, and accountable.
This is not a concept
Watch the accountable team work.
A genuine public run: Gemini coordinates, independent model families execute and review, and every completion claim carries evidence. The roster distinguishes workers that participated from workers available for the next run.
Rally
⌘ K
Connecting
T
Live runs
Connecting to D1…Loading authoritative runsReading the public D1 projection.
Loading real run data…
Rally authoritative runner → Cloudflare D1
Loading the execution timelineNo sample data is shown while the live service is unavailable.
New in Rally v2
Ask Rally on the page.
A supported browser agent can now prepare real Rally work through named tools, place the result in fields you can inspect and change, then review your exact revision. It cannot generate, publish, connect, or approve on your behalf.
Checking WebMCPEvery task still works from the page controlsClear protocol roles: WebMCP shares this page with a browser agent. Rally's governed MCP gateway reaches approved business tools. A2A v1.0 carries bounded tasks to outside agents.
What Rally solves
One goal in. One checked outcome out.
Rally turns a collection of isolated AI tools into an accountable team without asking your people to become agent managers.
Shared context
Stop carrying the context
Rally handles the handoffs between models, preserves the goal, and keeps one shared record of what each agent did.
Specialist execution
Use the right AI for the job
Different models can coordinate, build, investigate, and review instead of forcing one assistant to be equally good at everything.
Independent proof
Trust the result
Every consequential claim returns with an owner, a different verifier, evidence, and residual risk—not another confident answer to supervise.
Business systems + agent workforces
Connect the company.
Extend the team.
Business systems give Rally governed context and tools. External agents accept delegated work and return artifacts. Rally keeps those permissions separate—so connecting a source never silently commissions a workforce.
MCP tools + contextA2A delegated work
Company memory
Knowledge & collaboration
Google Workspace
Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat, and contacts.
Slack
Decisions, project history, team updates, and approval requests.
Atlassian
Jira work, Confluence knowledge, and Compass services.
Work in motion
Build & automate
GitHub
Repositories, issues, pull requests, Actions, and prepared changes.
Cloudflare
Worker logs and metrics through the narrow Observability server.
n8n
Only the workflows an administrator explicitly exposes.
Decisions & revenue
Data & customers
BigQuery
Governed data for sourced analysis, anomalies, and executive answers.
Salesforce
Customer, pipeline, and service records through hosted SObject reads.
Stripe
Payments, subscriptions, customers, and financial reporting.
External agent network
Bring another workforce into the accountable team.
Gemini, Claude, and OpenAI are Rally’s core team. These systems extend it through a deliberately narrower delegation boundary.
Different permission.Agents receive bounded assignments—not ambient access to every connected system.
Adapter readyHyperagent
Hosted OAuth MCP workforce. Rally can discover agents and read results; starting work is not enabled.
MCP delegationInterface ↗
Interface verifiedHermes Agent
Native A2A v1.0 peer. Endpoint identity, task mirroring, and failure tests must pass before admission.
A2A candidateInterface ↗
Interface verifiedOpenClaw
Authenticated A2A v1.0 peer. Rally will persist task state and expose only explicitly selected agents.
A2A candidateInterface ↗
A policy-governed control plane, built on Google Cloud
Every layer has one job—and a receipt.
Google Cloud handles identity, durable coordination, and telemetry. Deterministic policy—not a model prompt—decides what may proceed.
1
Commission
Signed email ingress
Resend and a Worker/D1 queue authenticate, deduplicate, and retain the request until handling succeeds.
2
Govern
Gemini 3.7 + ADK
An IAM-protected Cloud Run service preserves intent and atomically records the handoff in Firestore.
3
Execute
Cross-family workers
Claude, Gemini, and OpenAI Codex rotate through implementation and independent review in one controlled workspace.
4
Prove
Evidence comes home
Cloud Trace, test output, verifier identity, and residual risk arrive in the same executive email thread.
Coordinator reliability
All passedSix real scenariosFrom a simple briefing to a multi-system release
6/6Request preserved exactlyEvery handoff kept the intended outcome and limits
1.00Hostile input resistedExternal content could not rewrite Rally policy
PASSProduct safety checks
Green194 runner + ingress + policy + WebMCP
186 Cloud, A2A + connector security
───
380 automated tests
✓ Terraform validated
✓ Worker bundle verified
✓ Container runtime smokeFortified by design
Autonomy with a chain of custody.
Rally assumes messages repeat, models can be persuaded, workers fail, and loops eventually need a hard stop.
Identity perimeter
Only known actors enter.
Dual authentication
Cloud Run IAM plus a Secret Manager-backed application credential.
Fleet discovery
An authenticated catalog declares every agent’s capability, scope, and prohibitions.
Execution integrity
Every turn is bounded.
Replay-safe runtime
Atomic claims, retry leases, and attempt fencing prevent duplicate or stale work.
Hard circuit breakers
Turn, time, send, rejection, and stagnation ceilings stop runaway behavior.
Recovery + evidence
Failure never hides the trail.
Second Wind recovery
A failed model can hand recoverable work to a teammate without transferring approval authority.
Private telemetry
Trace and logs prove execution while prompt and response capture remains off.
Clear answers
Built to be questioned.
Is the console showing a mock run?
No. It reads an explicitly public, allowlisted projection of Rally’s authoritative runner state from Cloudflare D1. If that service fails, the page shows the failure instead of sample data.
Are the agents just agreeing with each other?
No. They execute as separate provider-native processes, come from different model families, and cannot verify checklist work they own. The live receipt counts only families that actually participated.
Do the agents wait for each other?
Yes. Rally runs one model at a time and saves only validated state before the next handoff. With Second Wind enabled, a timeout, failed process, or reported blocker gives the other model one bounded recovery attempt from that saved state—including inspection of partial workspace edits—without auto-approving anything.
Do emails literally trigger every model turn?
The first email starts the real run and every turn is mirrored to the thread. The authoritative runner dispatches the next model locally.
Can anyone who finds a teammate’s email address commission work?
No. A message may arrive without becoming an authorized job. Rally authenticates the transport, checks the sender against the teammate, channel, and company policy, and treats quoted history, attachments, and linked pages as untrusted evidence—not authority.
What happens when the agents cannot finish?
Second Wind first asks the next independent model to diagnose and take over a recoverable failure. Rally still stops with a precise report when the fleet remains blocked, disagrees repeatedly, stops making progress, hits a hard budget, crosses an authority boundary, or receives a human STOP.
Is Rally an official Google product?
No. Rally is an independent Agent9 project built with Gemini, Google ADK, and Google Cloud for the All Things Agentic Hackathon.
How does A2A fit—and is Rally A2A compatible?
Yes. Rally publishes an A2A v1.0 Agent Card and accepts tasks over JSON-RPC and HTTP+JSON; official SDK clients exercise both paths. A2A enables discovery and task exchange. Rally adds the controls around the work: authority, durable state, recovery, evidence, and independent verification. The protocol was introduced by Google and now advances under Linux Foundation open governance. Compatibility is not certification or endorsement.
What stays private?
Commissioner identity, worktree paths, mail IDs, raw prompts, credentials, and cloud request keys never enter the public console record.
When do administrators connect Gemini, Claude, OpenAI, or Grok?
AI workforce authorization is separate from company systems. The current pilot uses provider-approved identities on its controlled execution host. The self-service design starts each provider’s official OAuth or device-code flow on a user-bound host, with an existing API key offered as an alternative route. Rally never asks for a consumer password or copies a browser session cookie.
Are AI and business-system connections shared between users?
No. Model identities, company-system grants, and run authority are separate. Rally namespaces each user’s approved systems and tools, then freezes only the commissioned scope into that run.
Can a browser agent use Rally directly?
Yes. Rally exposes seven bounded WebMCP tools in supported browsers: search the public run index, inspect a verification record, prepare a governed job, stage an original Lyria song brief, stage an Agent9 Insights draft, stage governed MCP onboarding, and review the exact human-edited draft. The agent shares visible page state with you, but it cannot submit, generate, publish, connect, grant access, or bypass your decision.
Can Rally connect a custom MCP server?
Not through an arbitrary URL box. Custom remote MCP is a Labs path that must pass HTTPS and network admission, OAuth origin checks, bounded discovery, schema fingerprint review, exact tool policy, payload ceilings, and one-time approval for writes. WebMCP is complementary: Rally uses it for human-present browser collaboration, while asynchronous jobs continue through the governed server-side gateway.
What does “give Rally a job” mean?
Ask for a finished professional outcome—such as a sourced market briefing, an executive presentation, a customer-risk review, or a launch plan—then grant only the sources, limits, and approval rules that outcome needs.
Are the business connectors live?
The governed gateway and all ten provider adapters are implemented. BigQuery has completed a real authenticated handshake with Google’s official MCP endpoint and discovered six tools. The other accounts are intentionally shown as disconnected until each user completes provider setup, authentication, live discovery, and an exact safe preset. Rally never turns “adapter exists” into “customer account connected.”
The accountable AI team
Your next AI teammate
already has an email address.
Create the role, approve its context, and give your people a teammate they already know how to reach.