Local Codex operator console for observable AI worker teams.
Chat in Codex. Watch the work in Together. Route, verify, review, and recover locally.
Together is a terminal-native control plane for scoped AI work, live worker output, agent readiness, review gates, and local-first execution history.
Codex stays the preferred integrator. Worker agents stay replaceable, scoped, and observable.
Install | Quickstart | How It Works | Governance | Repo Map
Together is a local-first CLI/TUI control plane that lets developers keep working in Codex while seeing, routing, verifying, and reviewing the AI worker activity behind the scenes.
It is currently a dogfooding-stage operator console:
togetherstarts or connects to a local daemon.- The terminal UI shows Project, Tasks, Task Monitor, Live Work Feed, Agents, Needs Attention, and Chat Dock.
- Codex app / skill requests and Together chat requests flow into the same daemon event log.
- Worker selection is deterministic and observable.
- Task contracts define scope, allowed files, denied files, deliverables, and success criteria.
- Verification and approval gates block unsafe or out-of-scope work.
- Settings, theme presets, status, review, and packaging commands are available from the CLI.
Longer-term, Together aims to become an AI Department Operating System for local and team AI worker workflows. The current product surface is the CLI/TUI operator cockpit.
It is:
- a terminal-native monitor for Codex-led AI work
- an agent control plane
- a work governance layer
- a verification and review gate
- a local routing and failover layer
- a measurement loop for real developer sessions
It is not:
- another AI assistant
- a replacement for Codex, Claude, Gemini, Amp, OpenCode, or other workers
- a generic terminal multiplexer
- a claim of finished enterprise orchestration
| Task Monitor | Tasks |
|---|---|
![]() |
![]() |
| Task Detail | Settings |
|---|---|
![]() |
![]() |
These mockups and analysis visuals capture the product direction behind the current interface.
| Architecture | Governance |
|---|---|
![]() |
![]() |
| Department Flow | Operations |
|---|---|
![]() |
![]() |
| Registry | Failover |
|---|---|
![]() |
![]() |
| Hero Concept | Ready-State Model |
|---|---|
![]() |
![]() |
For customers who just want to run Together:
git clone https://github.com/kyoo-147/together_working.git
cd together_working
cargo build -p cli --release
target\release\together.exeThe terminal app auto-starts the local daemon, opens the Monitor + Chat Dock UI, and writes runtime state under .together/.
Useful CLI commands:
target\release\together.exe doctor
target\release\together.exe status --json
target\release\together.exe chat --source codex-app "create a scoped landing page task"
target\release\together.exe proposal confirm <proposal-id>
target\release\together.exe settings set --theme "Ocean Blue"Windows install/package helpers:
powershell -ExecutionPolicy Bypass -File scripts\install.ps1 -AddToPath
powershell -ExecutionPolicy Bypass -File scripts\package-windows.ps1Codex skill bridge:
python skills\together\scripts\submit-chat.py "create a scoped task for the landing page"The older Python scan/report commands remain available as compatibility tools for readiness reports.
Together is built around a simple operating model:
- Codex is the preferred integrator.
- Workers are scoped and replaceable.
- Tasks are small enough to verify.
- Runtime state is observable.
- Approval depends on evidence, not trust.
Why this exists:
- giant contexts grow too expensive
- one worker should not own every step
- routing should be explicit
- verification should not be optional
Install full skill:
npx skills add https://github.com/kyoo-147/together_workingInstall only main skill:
npx skills add https://github.com/kyoo-147/together_working --skill "together"Skill entrypoint stays here:
skills/together/SKILL.md
Scan machine:
python skills/together/scripts/discover-agents.py --format tableWrite operator snapshot:
python skills/together/scripts/doctor.pyList changed files:
python scripts/changed-files.py --jsonValidate one task:
python scripts/validate-task.py examples/task-contract.example.yaml --mode warn
python scripts/validate-task.py examples/task-contract.example.yaml --mode strict --write-artifactsRender report:
python skills/together/scripts/render-report.pyRun local release checks:
python scripts/release-check.pypython skills/together/scripts/discover-agents.py --format tablepython skills/together/scripts/write-registry.pypython skills/together/scripts/doctor.pypython skills/together/scripts/render-report.pypython scripts/validate-json.pypython scripts/validate-registry.pypython scripts/validate-routing.pypython scripts/changed-files.py --jsonpython scripts/validate-task.py examples/task-contract.example.yaml --mode warnpython scripts/release-check.py
Thin wrappers:
bin/together-scanbin/together-reportbin/together-doctorbin/together-validate
Known Providers
β
Installed CLIs
β
Ready Agents
β
Departments
β
Routing
β
Verification
β
Report
β
Output
Stage meaning:
Known Providers: curated ecosystem Together understandsInstalled CLIs: commands found on current machineReady Agents: installed workers passing lightweight checksDepartments: planning, research, vision, engineering, review, verification, fallbackRouting: readiness-first and capability-awareVerification: scope, quality, and acceptance checks before mergeReport: operator-readable snapshot
Three layers matter:
Known Providersis ecosystem coverage, not local installationInstalled CLIsis machine stateReady Agentsis machine state plus health readiness
Health checks stay cheap:
- command exists
- help/version runs
- obvious auth/config failure detected
- permission-denied state detected
Together assumes workers need boundaries.
Rules:
- agent cannot do everything
- agent works only inside assigned scope
- review checks output quality
- verification checks contract and scope
- Codex decides merge and integration
Task contract fields:
- task id
- scope
- allowed files
- denied files
- deliverables
- success criteria
- reviewer required
- verification required
Permission model:
- Observer
- Researcher
- Implementer
- Reviewer
- Integrator
Codex defaults to Integrator.
What the current CLI/TUI product enforces now:
- task contract validation
- git diff changed-file capture
- scope guard
- file policy validation
- verification artifacts
- quality gate
- review and approval gate state
- daemon-owned settings and status
- proposal confirmation before mutation
- PTY-backed worker execution path
- degraded-agent fallback
What it does not do yet:
- distributed scheduling
- full enterprise RBAC or centralized audit
- adaptive routing trained on long execution history
- automatic agent sandboxing beyond local process/worktree boundaries
- automatic commit or PR creation
Request
β
Planning
β
Research
β
Vision
β
Engineering
β
Review
β
Verification
β
Integration
β
Output
When a preferred worker fails:
- mark degraded
- start cooldown
- route to healthy fallback worker
- probe recovery later
- return to preferred worker when healthy again
Committed, scrubbed examples:
examples/agent-registry.jsonexamples/agent-report.mdexamples/last-known-good.jsonexamples/runtime-state.jsonexamples/providers.override.example.jsonexamples/task-contract.example.yamlexamples/tasks/*
These are examples, not live runtime outputs.
skills/ installable skill entrypoint and compatibility scripts
docs/ product and system docs
examples/ committed sample outputs and templates
tests/ lightweight validation tests
bin/ thin wrappers around current Python scripts
commands/ operational command guides
src/ reusable Python helpers for validation and reporting
.github/ CI workflow
assets/ brand and screenshot notes
agents/ worker-specific guidance
Generated runtime output stays local and ignored:
.together/cache/*.together/reports/*.together/runtime-state.json
Committed templates:
.together/providers.override.example.jsonexamples/*
- PRODUCT.md
- INSTALL.md
- CONTRIBUTING.md
- docs/architecture.md
- docs/registry.md
- docs/routing.md
- docs/governance.md
- docs/enforcement.md
- docs/benchmarking.md
- docs/task-workflow.md
- docs/observability.md
- docs/release.md
- docs/distribution.md
Near-term:
- dogfood the Monitor + Chat Dock UI in real developer sessions
- improve benchmark and telemetry capture for task latency, fallback, verification, and retries
- polish review, diff, settings, and compact terminal views
- strengthen Codex skill/app bridge documentation
- expand provider adapter tests and readiness diagnostics
Later:
- team policy templates
- CI and PR integration
- richer execution history and adaptive routing
- optional sandbox profiles
- multi-machine worker pools
- registry is curated, not exhaustive
- capability hints are routing hints, not benchmark claims
- failover is lightweight cooldown state, not a distributed scheduler
- Together manages workers. It does not replace them.
MIT


















