Native agent sessions remain persistent
Each worker is a real agent process with a durable identity, observable screen, closed-loop input, and recoverable history.
- Claude Code
- Codex
- Grok Build
- Gemini CLI
Private beta
Run coding agents in one repository.
Fleet gives Claude Code, Codex, Grok Build, and Gemini CLI shared ownership, durable memory, and operator control.
Have an invite? Get access →
Use each model where it works best.
- Claude CodeClaude in its native harness
- CodexGPT in its native harness
- Grok BuildGrok in its native harness
- Gemini CLIGemini in its native harness
Fleet carries assignments and context between native agent sessions.
Engineering teams often plan with one model and build with another, copying context between terminals by hand.
Fleet keeps those sessions in their native tools while the orchestrator assigns work, builders execute it, reviewers verify it, and the operator can intervene.
Agents share one worktree without overwriting one another.
Fleet assigns active territory exclusively while leaving unclaimed files open for discovery and expansion.
Fableorchestrator / delegates AUTH-17
GPTbuilder / owns implementation
shared repository3 active claims
src/auth/session.tsGPT / claimed
src/auth/session.test.tsGPT / claimed
reviews/auth-security.mdClaude / claimed
docs/operator-guide.mdopen
Claudereviewer / owns security review
Grokpeer write blocked / owner notified
Atomic claims and explicit handoffs stop peer collisions at the tool boundary before they become merge conflicts.
Every assignment includes its scope and proof.
Every assignment carries the specification, starting files, live peer boundaries, operating laws, and proof command.
The agent knows what success means before it writes and reports evidence whenever the premise is wrong or the scope must expand.
PACKET AUTH-17claimed by builder-gpt
- Objective
- Make session renewal single-use and preserve existing login behavior.
- Files
src/auth/session.tssrc/auth/session.test.ts- Peers
- Claude owns
reviews/auth-security.md. Do not touch it. - Verify
npm test -- session- Report
- Attach test evidence to AUTH-17, then mark the item done.
Fleet exposes stalls, restarts, and lost context.
The runtime keeps work recoverable when an agent stalls, forgets, or disappears instead of assuming every session will respond and exit cleanly.
The ledger survives restarts and context loss
Claims, decisions, evidence, handoffs, blocks, and verification survive context compaction and process restarts.
Fleet detects and reports stalled sessions
Liveness probes, stall detection, beacons, stale-claim recovery, and operator alerts make unattended work observable.
The control room shows every active session.
The control room ranks sessions by what needs attention, then lets the operator inspect, message, and recover any seat.
Use the subscriptions you already pay for.
Fleet coordinates native harnesses without flattening every model into another metered API.
Agents keep their existing subscriptions and provider authentication, so Fleet does not add token markup.
- Claudethrough Claude Code
- GPTthrough Codex
- Grokthrough Grok Build
- Geminithrough Gemini CLI
Release status
Fleet is complete and undergoing beta hardening.
The working product remains gated while installer reliability and security are verified.
Complete
Fleet runtime
Native seats, orchestration, control room, ledger, ownership, and recovery are implemented.
In progress
Beta installer hardening
We are validating clean installation, update, recovery, and uninstall on real machines.
In progress
Security review
The release candidate is undergoing an adversarial review before wider access.
How Fleet handles models, files, security, and cost.
Does Fleet replace Claude Code, Codex, Grok Build, or Gemini CLI?
Fleet keeps each model inside its native harness and coordinates the sessions around it rather than replacing its tools.
Why use one worktree?
Agents see one another's changes immediately, while atomic ownership prevents destructive overlap and leaves unclaimed territory open.
Is file ownership a security sandbox?
Fleet provides a coordination boundary for supported agent tools rather than hostile-process isolation; it blocks peer collisions and records the operating history.
What remains before wider release?
The product is complete, and wider release depends on beta installer hardening and the current security review.
Will Fleet add another model bill?
Fleet adds no token markup because coding agents continue using their own subscriptions or provider authentication.
Join the Fleet beta release list.
We will email when the hardened beta is ready and only about Fleet beta updates.