Getting started
Begin with a small result you can run and inspect.
How do I install and set up Canopy?
Download Canopy for your computer, install and sign in to a supported coding CLI, then open a project folder. The first-project guide includes a tiny practice app, a copyable task, and the checks to run before accepting the result.
Can I use Canopy without knowing terminal commands?
You can direct an agent without writing every line, but you still need a project, Git, an installed coding CLI, and the runtime your app uses. Start with a small practice task and verify the running result.
First project guide → · Noncoder path → · Founder workflow →
Does Canopy include an AI model subscription?
No. Canopy runs supported CLIs you install and sign into. The CLI and its provider control model access, usage limits, and billing.
Can I use Canopy offline with a local model?
Local project surfaces can work without Internet when their files and dependencies are present. AI inference needs an installed CLI configured to reach a local model server; test the full task after disconnecting.
How do I find the command that runs my AI-built app?
Use the project's own README and scripts to identify the working directory, prerequisites, command, and local URL. The first setup error is evidence to investigate, not a reason to ask the agent to rewrite the app.
What should I ask the agent to build first?
Choose one observable behavior with acceptance checks. Ask for a running demo, the exact tests run, changed files, and what remains uncertain.
Can I talk to a coding agent in Canopy and hear it reply?
Canopy documents on-device dictation into prompts and other text fields on supported systems. Claude Code and Copilot CLI also document their own voice input. Spoken replies and continuous hands-free conversation require a separate verified output path.
Agents, models, and context
Keep separate sessions useful and know which CLI owns each capability.
Can Claude Code and Codex work on the same project?
Yes. Give independent code changes separate worktrees and branches. The CLIs retain separate conversations and accounts; use repository artifacts and a short handoff to coordinate.
Can agents in different tabs talk to each other?
Supported Canopy sessions can exchange direct messages and project context, but that does not merge their private conversation histories. Verify delivery and task ownership before assuming the handoff succeeded.
Do file claims keep two agents from overwriting the same file?
Treat a claim as a visible overlap signal. Check each session's actual checkout and Git diff; separate worktrees isolate working files, while a person still owns integration and review.
How many coding agents should I run at once?
Count independent outcomes and your capacity to review them. More live tabs help only when checkouts, services, and final integration remain understandable.
Two-agent workflow → · Choose concurrency → · Triage live agents →
Is my coding agent stuck or waiting for me?
Check the live terminal's last event and the process it started. An approval, persistent dev server, completed turn, slow test, and failed process need different next steps.
How do I resume a task without pasting a huge transcript?
Preserve the goal, branch or PR, changed files, tests, decisions, and next action. Resume the owning CLI conversation when it is useful; a different CLI needs an explicit handoff.
Why did Claude Code forget my task after compaction?
Compaction can lose detailed instructions that lived only in conversation. Check the current branch, diff, accepted decisions, and tests before giving Claude one verified next step. Keep stable project rules in the instruction file the CLI reloads.
How do I stop an agent rebuilding a feature we already have?
Ask it to find the active page, route, helpers, callers, and tests before editing. Have it explain what can be reused and why, then check the running result and final diff for duplicate paths.
Copy the read-before-build task → · Investigate without editing →
How do I keep an agent in the right app in a monorepo?
Map the user-facing route to its owning package and shared dependencies. Give the agent a bounded initial edit area, then verify that its run command, URL, checkout, tests, and final diff refer to that app.
How can I tell whether a coding-agent skill loaded?
Check that this CLI discovered the skill, invoke it explicitly on a safe task, then try natural wording and an adjacent request that should not trigger it. Inspect the invocation and resulting artifact, not just the agent's claim.
Should I share a skill I found online with my team?
Read the entire skill folder, including scripts and referenced files. Map its tool and network access, run it against sample data in a disposable checkout, and share a reviewed version with an owner.
Run and review the result
A proposed change becomes useful when the app runs and the final diff is checked.
Can I start my website, API, and worker together?
Save each component's real development command and working directory, then verify its output and URL. A saved command makes restarting easier; it does not mean the process kept running while the host was off.
Why won't my dev server start?
Identify the failing component, checkout, command, and first error. Check its runtime, port, dependencies, and printed URL before changing application code; a running process can still serve the wrong branch.
Why won't my phone open my local dev site?
Your phone's localhost points to the phone. Use the computer's reachable LAN address and the server's actual port, then check whether the frontend, API, and browser features work from that origin. Canopy Remote controls the desktop host; app page access has its own network route.
Does a local AI-built app use my production database?
It can. A localhost page may call a hosted API, database, queue, or payment service. Identify the browser request and each server-side target before an agent runs writes, imports, or migrations.
How do I run Supabase locally with my AI-built app?
Start the Supabase CLI stack in the repository that owns its configuration, read the local URL and keys from `supabase status`, then point the app's development processes at that local project. Verify a disposable write and fresh read before trusting the setup.
Why does my AI-built app lose data when I refresh?
Check whether Save changes only the page state, writes browser storage, or completes a backend request. Define whether the data should survive a tab, a browser restart, or a different signed-in device before asking the agent to fix it.
Do two agent worktrees get separate development databases?
No. Worktrees separate checked-out files. Check each app's actual database endpoint and state; isolate mutating tests and migrations or run them sequentially, then test the combined result.
Can I run a Trigger.dev worker beside my app?
Yes, as a configured local process after Trigger.dev's one-time setup. Its dashboard remains the source for task run state; Canopy shows the local command and output.
How do I test a Trigger.dev email task without contacting customers?
Identify the Development task, app trigger key, email provider, every recipient, and other side effects first. Run one dashboard test with a provider-approved test address, then one signup through the app and match its new run ID and provider event.
Why did my Trigger.dev task send two emails?
Count task run IDs, retry attempts, and provider send records separately. Two runs point toward a repeated trigger; one run with several attempts points toward retry behavior; one attempt with two sends points toward the task or provider path. Inspect the original send before replaying it.
Why are my Trigger.dev development runs stuck in a queue?
Find the run in the right project and Development environment. Trigger.dev warns that two local dev workers on one project can split run consumption; identify every worker before stopping one, then verify a safe dashboard test and the app trigger path.
Can an agent review a PR and address comments or CI?
Canopy's workflow can stage findings and hand focused follow-up work to agents. A person should verify each finding, the latest diff, checks, and whether the original concern was actually resolved.
Why did the coding agent change a huge lockfile for a small fix?
Compare the dependency manifest, lockfile, agent install command, and repository package manager. Keep an approved dependency change, remove accidental manager output only after review, and verify a clean install at the PR head.
Can someone use an AI-built page without a mouse?
A screenshot cannot answer that. Walk the actual form and dialog by keyboard, check visible focus and accessible names, inspect error announcements, and use automated scans as supporting evidence.
How do I review an AI-generated login or permission change?
First agree on who may do what to which data. Check the server and database enforcement, then test both an allowed account and another user's or team's account against the exact object. Review the final commit before approval.
Is an AI-built Stripe checkout ready after the success page works?
Trace a sandbox payment through the signed webhook to one order or entitlement. Test decline, delayed payment, duplicate delivery, and a customer who never reaches the return page before enabling live payments.
What if my coding agent committed an API key?
Stop further sharing, identify where the credential went, and have its owner revoke or rotate it with the provider. Restore affected services and inspect usage before deciding with repository owners whether shared Git history needs cleanup.
Can I deploy an AI-generated database migration after it runs locally?
Review the exact SQL and database target, then test against representative existing data in staging. Check access rules and the production project's actual backup or restore option before an owner approves deployment.
Can I launch an app that an AI agent built?
Fix one release commit and test its real user journey, cross-user data access, recovery, and any payment or worker paths. Record who reviewed the technical risks and who owns a go, hold, or reduced-scope decision.
Which agent session introduced a regression?
Reproduce the failure in a known revision, locate the suspect change in Git, then compare its PR and checkout with the related session. A recent agent tab or blame result alone does not prove cause.
What if the agent says done but the app still fails?
Reproduce the failure, inspect the correct service and browser evidence, then give the agent a focused correction. Repeat the original check and read the revised diff before accepting it.
Why does my AI-built app work locally but fail on Vercel Preview?
Compare one local commit with the Preview deployment's exact commit. Classify the first failure as build, browser, function, or external service, then compare Preview variables and backend targets without sharing secret values.
Why does Vercel say the Git author must be a team member?
That message points to a deployment access gate, often before a build starts. Check the blocked commit's author, linked Vercel account, owning team and plan, and the owner's access decision before asking an agent to edit code.
What if an agent edited the wrong branch?
Stop new edits, identify the checkout that contains the changes, preserve them, and verify the destination branch before restarting work. A tab title is not proof of the actual working directory.
Phone access and team sharing
Know what lives on the host, who can act, and what can be shared.
Can I control an agent from my phone when my computer is off?
No. Canopy Remote connects to the running desktop host. Enable its PIN-protected path when needed and check the current project before steering a session.
Should I use Claude Code Remote Control, Copilot CLI, Canopy Remote, or SSH?
Choose the view that covers your actual task: one CLI conversation, several installed agents and local services, or a general shell. All four local routes require a running host; compare their access rules and what you can verify on your phone.
Did my agent stop when my phone disconnected?
A phone connection loss does not establish the CLI's state. Reconnect to the host, find the same session and branch, and inspect the latest output and changes before sending an instruction again.
Reconnect without repeating work → · Understand process lifecycle →
What can someone with my Remote PIN do?
A person who can reach the endpoint and has the current PIN can drive permitted Remote surfaces, including agent terminals. Protect and rotate the PIN, and remember that a CLI's own permissions remain relevant.
Can teammates share my paid AI model account?
No. Team collaboration can exchange project messages, files, and review requests. Each installed CLI retains its provider credentials, models, limits, and billing.
Usage, cost, and tool choice
Use complete task outcomes and verified account data to choose a workflow.
Is Canopy's estimated session cost my provider bill?
No. Usage estimates and plan-limit snapshots answer different questions from an invoice. Check the account, time window, token categories, and provider record before drawing a spending conclusion.
What if an agent hits its usage limit mid-task?
Check the live CLI turn and the provider's exact limit, then preserve the branch, diff, tests, and next action. Decide whether to wait, use an available account option, or hand the bounded task to a different CLI.
Why does a session show so many sent tokens?
Later agent turns can include prior context, files, tools, and repeated or cached input. Separate fresh input, cache reads, cache writes, and output before deciding a session wasted money.
Should I use Haiku or Luna first, then Opus or Astra for review?
Try a bounded, read-only first pass and give the reviewer the final commit and acceptance checks. Compare complete task costs, retries, and accepted results. Model availability and billing belong to each CLI and provider account.
Why use Canopy instead of a terminal or AI editor?
Choose by the complete task. A terminal can be enough for one CLI; editors and agent workspaces add different ways to run, inspect, resume, and review work. Compare the same repo, CLI, and acceptance checks.
How should I compare AI coding workspaces fairly?
Use the same repository commit, bounded task, CLI and model where possible, and acceptance checks. Record setup, running behavior, review, resume, human effort, and the final accepted result.
Can Canopy replace Replit Agent for building an app?
Canopy can run an existing project locally with installed coding CLIs, services, preview, and Git review. Replit includes a hosted Agent and a publishing path. Compare the same feature and account for runtime, secrets, data, and deployment separately.
Can I use a local coding agent on my Lovable app and keep editing in Lovable?
Yes. Connect Lovable Git sync, work locally on a feature branch, review and merge it into Lovable's active synced branch, then check Lovable Preview and publish separately. Confirm the local app's backend before testing writes.
Can I keep using Bolt after an agent edits its GitHub repo locally?
Yes, when the Bolt project is connected to GitHub. Work on a feature branch, merge in GitHub, verify Bolt fetched the intended commit and branch, and check the published site separately. Avoid simultaneous edits to the same feature.
Why does v0 not show a local agent's merged GitHub change?
Check that the PR merged into the v0 chat's base branch. An open v0 working branch may need Pull Changes before its preview includes newer base commits; production is a separate deployment check.
How is Canopy different from Superset for parallel agents?
Both organize installed CLIs, worktrees, local previews, and PR review. Compare process persistence, app run setup, Remote access, and license terms in the versions you would use.
Is Windsurf now Devin Desktop, and how does it compare with Canopy?
Cognition renamed the editor Devin Desktop and added a command center for local and cloud agents. Compare its ACP and editor workflow with Canopy's installed CLI terminals and local project services using one finished task.