Your coding agents get the whole product.

An agent that only sees code has to guess at everything else. Claude Code, Cursor, and Codex call Dodger over MCP and get the full picture for the task: why it exists, what was decided, and how the feature is doing with real users right now. Dodger's own coding agents work through the same door.

Book a demo

Live calls · MCP

Cursorget_task_context200 · 4 fields

→ Fix invite delivery to Microsoft domains

Why it exists14 support tickets and finding F-31 point at silent delivery failures

DecidedKeep the new sending domain from PR 191

Live stateAcceptance at 61% in production and falling

AcceptanceInvite acceptance back above 70%

Claude Codeask200 · 3 sources

→ "Which domains fail most?"

Microsoft domains. 81% of failed invites in the last 30 days went to outlook.com or a hosted tenant, from delivery logs, support tickets, and finding F-31.

Dodger Agentget_task_context200 · 4 fields

→ Add CSV export to the claims list

Why it existsFinance rebuilds this list by hand every month · 6 tickets

DecidedExport follows the on-screen filters, not the full table

Live stateClaims list opened 340 times a week

AcceptanceExport matches the filtered view to the row

Codexget_task_context200 · 4 fields

→ Speed up quote search

Why it existsSearch past 1,000 quotes feels slow · finding F-27

DecidedFix the query before adding any caching layer

Live statep95 search time at 900ms and rising

Acceptancep95 under 200ms at 10,000 quotes

Yours or Dodger's own: every agent gets the context for its task, scoped to what it's allowed to see.

It has coding agents of its own.

No agent on the task? Hand it to Dodger. Its agents get the same context yours do, and nothing merges on trust: you approve the plan before any code is written, the work runs in a sealed sandbox, a separate agent checks it, and the feature is replayed on video before you merge. The same checks run whether the code came from Dodger's agent or yours.

Add CSV export to the claims listPlan ready Building PR #64 open

1Add an Export CSV action to the claims toolbareditsrc/claims/Toolbar.tsx

2Serialize the filtered rows, not the full tablecreatesrc/claims/export.ts

3Cover filtered and empty exports with testscreatesrc/claims/export.test.ts

Approve & buildor edit the plan in plain language

Run startedfresh sandbox · acme/claims-console

GrepClaimsToolbar

Readsrc/claims/Toolbar.tsx

Editsrc/claims/Toolbar.tsx

Writesrc/claims/export.ts

Wrote 3 files, self-verifiedbuild and tests clean

Tests vs baseline6 passed · 0 failedPassed

Static analysis0 findingsPassed

Agent reviewmatches the approved planPassed

0:23

Filter to open claims0:04 Export CSV0:12 File matches the view0:19

Planned in a read-only sandbox · nothing has been written yetIsolated sandbox · none of your keys inside · cannot push to your repoEvery check is on the pull request · nothing merges without your approval

Review it like a teammate's PR.

The pull request reads like a teammate wrote it: what was built, how, and the proof. Leave a comment and the agent fixes it, replies, and updates the same branch.

Add side-by-side quote comparison#128

Open

dodger-agent wants to merge 4 commits intomainfromdodger/quote-compare

priya reviewed just nowView reviewed changes

app/quotes/carrier-search.tsx

@@ -41,6 +41,7 @@ export function CarrierSearch()

4141 const [query, setQuery] = useState("");

4242-onChange={(q) => runSearch(q)}

42+onChange={(q) => debouncedSearch(q, 250)}

4343 return (

priyajust now

Can we debounce this? It fires a request on every keystroke.

dodger-agentbotjust nowAuthor

@priya Added a 250ms debounce and a test covering rapid input. Pushed the fix to this branch.

dodger-agent addedb8d114fDebounce carrier search+38 −6

All checks have passed

4 successful checks

It learns your repo before it writes to it.

Connect a repo and Dodger studies it once. Every task after that starts already set up.

How to run your project

Dodger learns the commands that install, build, and test your code. It also notes any tests that were already failing, so those are never blamed on the agent.

A ready-made copy

Your project is set up once and saved. Every new task starts from that saved copy instead of installing everything from scratch.

A map of your code

Every file gets a short summary. The agent can search your code by what it does, not just by its name.

Repo onboarding/northwind/quotesLearningReady

indexing src/quotes/compare-table.tsx

Commands learnedbun install · bun run build · bun test

Failing tests noted2 already fail. Never blamed on the agent.

Ready-made copy savedNew tasks start from here.

Code map built0 pieces, searchable by meaning

The sandbox contract.

These rules are built into the system, not written in a policy. They hold even when the agent makes a mistake.

It never sees your keys

The sandbox gets one pass of its own. Your API keys never go in.

It cannot push code

Work leaves the sandbox only as a pull request you review.

A second agent checks it

A separate agent, one that did not write the code, checks the work before you see it.

You own the merge

Nothing lands in your codebase until someone on your team approves the pull request.

sandbox · fork of golden-snapshotrunning

Give it a feature.

Connect a repo, approve the plan, and review a pull request that arrives with its proof.

Book a demo