Give the project one goal.
Describe the outcome once. The coordinator keeps the request, context and decisions in one conversation, and you choose the model, effort and approval mode.
SpecWithCodeFree & open sourceYour agents. Your machines. Your control.
One workspace for Claude Code, Codex and the agents you choose. Coordinate the work. Keep your folders, permissions and approvals in your hands.
Cloud models still receive prompts and relevant context through your chosen provider. Local execution does not mean offline AI.
Apple silicon · macOS 13+ · Early access
Not notarized yet: use System Settings → Privacy & Security → Open Anyway on first launch.
Project / tip calculatorRecorded app
Every thread stays within reach.Inspect the work behind each result.
Choose the machine.Local or privately paired.
Bring your providers.Keep eligible plans and sign-ins.
Set the boundaries.Folders, tools and approvals.
From one request to reviewed work
One conversation.
A coordinated team.
A coordinator hands parts of the job to workers on different models, then brings their results back to you. Scroll to follow a real run.
Illustration of a recorded run
YouReview the result
S
CoordinatorBoth tasks finished
◎Codexgpt-6-luna
Tip: $9.00 · Total: $59.00Reported ✓✳Claude Codeclaude-haiku-5-5
Usage section reported backReported ✓- You to CoordinatorBuild a tip calculator CLI with a usage guide.
- Coordinator to CodexWrite tip.py at the repo root.
- Coordinator to Claude CodeAdd a Usage section to README.md.
- CodexCreated tip.py · python tip.py 50 18
- Claude CodeEditing README.md, keeping existing content.
- Codex to CoordinatorTip: $9.00 · Total: $59.00. Edge inputs checked.
- Claude Code to CoordinatorREADME usage section reported back.
- Coordinator to YouBoth tasks are finished, and nothing is committed.
The coordinator splits the work.
It starts a worker thread for each part. Here Codex writes tip.py while Claude Code adds the usage guide. Each thread keeps its own provider, model and permissions on the machine you chose.
Workers run side by side.
Each worker receives the coordinator’s brief in its own thread. Open any thread to see the commands it ran, the files it changed and the output it got.
Results come back to one conversation.
Workers report to the coordinator, which checks each result against the goal and tells you what is left. Review the evidence before you accept anything: a reported result still needs verification.
The animated diagram illustrates the recorded run above; screenshots are actual app captures. Your providers, models and permissions determine what runs.
Make your AI budget go further
Put the plans you already pay for to work.
API billing meters every token. A flat coding plan, such as a $200-a-month tier, includes a usage allowance for the same kind of agent work. For heavy, sustained coding, the plan can cost a fraction of the same usage billed at API rates.
SpecWithCode puts those plans to work on machines you control: each agent runs where your files are, with the sign-in it already has. How much you save depends on your plan, model, limits and workload, so check it against your own bills.
- Sign in the native way. Claude Code and Codex use their own eligible plan sign-ins on each machine.
- See every limit in one place. Session and weekly windows for each account, before you hit them.
- Keep going at a limit. Continue the thread on another provider you have connected.
Same work, two metersAPI ratesFlat plan
3×the plan price would go to API billing for this usage, if the work fits within the plan’s limits.
Illustrative arithmetic with your own numbers, not a quote or a promised multiplier. Plans have session and weekly limits; extra usage, seat rules and provider terms can change the result. Enterprise plans can bill differently: Anthropic currently lists Enterprise as a seat fee plus API-rate usage. Check current provider pricing.
A workspace that reaches your machines
Start on one computer.
Connect the ones you trust.
Pair approved machines over your network, SSH or private HTTPS with Tailscale. Choose where work runs; each machine keeps its own pairing and native account access.
Work continues while the executing machine is awake. Connecting computers does not add subscription entitlement or grant access to an entire company network.
Your laptopChoose & supervise
Private pairing
Approved machinesExecute & retain files
Your model providers
Prompts and relevant context cross this boundary. Provider and integration data policies apply.
See the product doing the work
One goal. Two workers.
The actual recording.
Follow a coordinator, a Codex thread and a Claude Code thread through a small coding task. The later usage-limit example is simulated.
Clear boundaries.
Fewer surprises.
Does everything stay on my machine?
No. Agent processes and project files live on the selected machine, but cloud models receive prompts and relevant context, which can include code and file contents. Connected tools use their own services. Private machine connections do not make model inference local.
Can a subscription replace all API spending?
No. Use eligible plans through their supported coding agents. Limits, seat rules and provider terms still apply; API keys or extra usage may create separate charges. Enterprise billing may differ. Measure your own workload before claiming savings.
What do I need to connect another machine?
SpecWithCode running on a supported machine, a reachable private route and separate pairing permission. Tailscale is optional. Start with a small, approved setup before expanding; this page does not promise company-wide scale or universal hardware support.
What is free?
SpecWithCode is free and open source. You bring your provider accounts and pay any provider charges directly. The SpecWeave CLI keeps specs, task evidence and handoffs in git so another tool can continue the work.
Build with the agents you choose.
Keep the work on machines you control.