Structured tools
Native-loop intent, target and result stay scannable.
A coding agent for the models you run
Own your coding agent.
A native agentic loop for local and API models, plus delegated turns through supported coding subscriptions. In your terminal and your editor.
$
curl -fsSL https://vibedgc.com/install.sh | bashPython 3.10+ · no root · local, API, and subscription routes
94.4%Aider polyglot · pass@2
184 / 300SWE-bench Lite · published
0agent-round timeouts · DGC
121.7 saverage measured round
Ollamallama.cppvLLMLM StudioOpenAIAnthropicOpenRouterGroqDeepSeekClaude CodeCodexQwen CodeKimiGitHub Copilot
01 The CLI
A real agentic loop, in your terminal.
Reads, edits, runs the tests and self-corrects. Permission modes, plans, skills, hooks and a multi-agent fleet in one focused TUI.
saved run · Go · alphametics
- 01readalphametics.go · tests2 files
- 02editcolumn backtracking solverfirst round
- 03runillustrative in-session testfirst round
- 04verifysuite still failingself-correct
- 05readresume · inspect rollbackline 137
- 06editpreserve winning assignmentfocused fix
- 07verifyisolated graderpassed · 0.4 s
Reconstructed from the retained DGC result for go/alphametics. The linked evidence bundle preserves scored rounds, grader and agent timings, usage totals, result hashes, and the input hash; the in-session UI steps, diff prose, and their pacing are illustrative because raw harness output is intentionally omitted.
A small fix. A complete loop.
❯ Fix the clamp bounds and run the tests.
I’ll inspect the implementation and its tests first.
→ read_file clamp.py · test_clamp.py
✎ edit_file clamp.py
− return min(lower, max(upper, value))
+ return max(lower, min(upper, value))
$ bash python3 -m unittest -v
✓ 3 tests passed
Fixed the bounds. Values inside and outside the range now pass.
One checklist, updated as the work moves.
TASKS 0 of 3 complete
- Reproduce the failurePending
- Fix the bounds and add coveragePending
- Run the regression testsPending
❯ Fix the bug, add coverage, and verify the change.
→ todo Start by reproducing the failing case.
→ read_file clamp.py · test_clamp.py
✎ edit_file Correct the bounds and add the regression case.
$ bash python3 -m unittest -v
✓ Verification passed · checklist complete
Fixed the bug and verified the regression. No pending items.
Review the approach before the edits begin.
❯ /mode plan
❯ Plan the auth package refactor.
→ read_file src/auth.ts · src/routes.ts
Move the helpers, preserve exports, then update the callers.
→ present_plan Refactor auth into its own package
Awaiting your decision · Approve / Revise / Keep researching
Example choice: approve the proposed plan.
Plan approved. I’ll start with the helper package.
Bring the exact lines into the conversation.
❯ /diff
DIFF Working tree against HEAD
clamp.py +1 −1
− return min(lower, max(upper, value))
+ return max(lower, min(upper, value))
Space · start selection → ↓ · extend → Enter · attach
✓ 2 selected lines attached to your next prompt
❯ Why did you change these bounds? [clamp.py · 2 lines]
Project guidance, right where you work.
❯ /skills
SKILLS debug · code-review · write-tests · verify
❯ Use the debug skill to investigate this failing test.
→ skill debug
I’ll reproduce the failure before changing the implementation.
$ bash python3 -m unittest -v
Inspect the failing case → trace the cause → make a bounded fix
The lower and upper bounds are reversed. I’ll fix that expression.
Pick up the context and the pending work.
❯ /resume
SESSIONS Fix clamp bounds · current project
Choose the saved session.
✓ Conversation restored · saved task list restored
✓ Reproduce the failure
✓ Fix the bounds
○ Verify the regression
❯ Continue with the remaining verification.
Edit & test
Scripted UI illustrations · native local/API workflows.
02 The extension
The same harness, native to your editor.
On native local/API routes: structured tools, readable diffs, goal-aware turns, permission decisions, and plan feedback. Subscription routes retain DGC's local goal record and presentation while the vendor owns its turn.
Actual extension surface · deterministic protocol fixture · real disposable-file edit and unittest run · not a live model session · real time, no speed adjustment.
Reviewable diffs
Additions and removals remain first-class.
Goal controls
Local record, timer, progress, pause and cancel at hand.
Native plan review
Approve, revise, or keep researching on local/API routes.
02b Specialists
Named agents, with a face.
Explorer, researcher and critic are first-class task specialists. In the editor they land as coloured chips — the mark swirls while they work, stills when they finish. Click a chip for that agent’s own page: duration, answer, files. Child greps stay off the parent thread. A background task can outlive the turn; the composer stays free and DGC wakes a new turn when it lands.
map auth working
review findings finished
options research workingMap auth and have a critic review the findings.
map auth workingreview findings finishedIdentity marks · live while working · click opens the agent page
Named types
Explorer maps. Researcher writes one findings file. Critic reviews it. Worker implements.
Living chips
Eight identity marks. Same face for the same id. Animate only while the agent is working.
Inner page
Back, duration, the child’s tools and a FILES card. Composer stays on the parent.
Background wake
A detached specialist keeps going after the turn ends. The pill stays; DGC starts a new turn when it lands.
03 Native plan artifacts
Review the plan as a real page.
On local/API routes during the current process, a DGC plan can open as a clean page on a private loopback port—not raw Markdown racing past your terminal. The plan text saves with the session, so you can reopen it in DGC after a restart.
NATIVE ROUTES · PRIVATE LOOPBACK PREVIEW · PLAN TEXT SAVED WITH THE SESSION
Refactor auth into its own package
Move token and verification helpers into one package and update the callers, without changing behavior.
- Create the package and move the helpers.
2 files - Re-export the public surface.
1 file - Update callers and run the suite.
7 imports
Approve in terminal← → views · plan text saved
Bound the change before editing
The proposed surface stays inside the auth package and its direct callers. No migration or public API change is planned.
- src/auth/token.ts · verification helpers
move - src/auth/index.ts · stable re-exports
create - 7 direct imports · preserve behavior
update
Review in terminal← → views · 3 change groups
Define done before execution
Approval starts the implementation. The proposed definition of done requires every listed check to return a real successful exit code.
- Auth unit and regression suite
npm test -- auth - Static contract check
npm run typecheck - Patch hygiene
git diff --check
Execution after approval← → views · 3 required gates
04 Providers
Point it at anything that serves a model.
Local inference, direct APIs, or a subscription you already pay for. One harness and one /connect surface.
Your machine, your weights
Native chat or an OpenAI-compatible local server.
- Ollamanative chat
- llama.cppOpenAI /v1
- vLLMOpenAI /v1
- LM Studiolocal server
Your chosen endpoint
DGC sends model-conversation context to the endpoint you configure. Optional integrations receive directed requests; language servers also receive queried document text.
- OpenAInative
- Anthropicnative
- OpenRoutermulti-model
- Groq · DeepSeekcompatible
Use the plan you have
The vendor’s binary keeps ownership of login and terms.
- Claude CodePro / Max
- CodexChatGPT
- Qwen · Kimicoding plans
- GitHub CopilotGitHub
05 Plugins, apps and MCP
And point it at anything that serves a tool.
A model on your own machine is not a model with fewer tools. Install a plugin or connect a server and its tools join the same list as the built-ins — one route, one permission prompt, one set of hooks and redaction. The model is told what a connector brokers, so it stops calling a service unreachable that it can already read.
MODELqwen3.8:27bhttp://localhost:11434 · on this machine
I’ll pull the open issues, then write the notes.
Calling MCP tool
composio · LINEAR LIST ISSUES
Calling MCP tool
composio · NOTION CREATE PAGE
Edited file
release.md
Every one of those passed the same permission prompt, redaction and hooks as a built-in.
Step 1 of 4 · A model on your own machine
Scripted interface demo · the model never leaves this machine; the tool calls do.
Install one, on the record
Every entry names its licence and, separately, what DGC actually verified — often less than “it works”.
Settings ▸ Pluginsdgc plugin install
Accounts you already have
GitHub, Notion, Linear, Sentry, Supabase, Stripe. Installed is not connected, and the model is told which.
browser sign-innamed token
Many services, one server
Composio, Zapier, Arcade, Make, or n8n on your own instance. Several run the request in their own cloud — DGC says so before you connect.
mcp_searchrevoke upstream
The wire underneath
Stdio or remote, the 2026 protocol or the legacy handshake. Bounded frames, consent before any URL opens.
/mcpmcp__server__tool
06 Bring your own subscription
Already pay for a coding model? Run it here.
Pick the vendor in /connect. DGC passes each turn to its official CLI using that vendor account route—without configuring a separate DGC API key. Vendor plan, usage, and fee terms still apply.
Your existing plan
Pro, Max, ChatGPT, and other supported first-party subscriptions.
Official route
Each vendor binary owns authentication, account policy, and its model/tool turn.
One place to start
DGC maps the mode, retains session history, runs SessionStart/Stop hooks, and presents the vendor’s result.
DGCpasses prompt + mapped mode
→Official CLIowns sign-in, model + tools
→DGCrecords and presents the result
07 Benchmark
Measured on a model you run yourself.
A measured slice of 90 tasks across six languages. Every manifest records the same model alias, caller-declared digest, endpoint URL, hardware, and 600-second allowance per round; model weights were not independently verified.
goose95.6%
DGC94.4%
pi94.4%
OpenCode88.9%
Codex CLI87.8%
80%85%90%95%100%
pass@2 · axis runs 80–100%, not from zero
0agent-round timeouts · DGC only
121.7 saverage round · fastest measured
-24.6%output tokens vs goose
85 / 90solved · rank 2, tied with pi
Accessible table view
| Harness | Solved | pass@2 | Evidence note |
|---|---|---|---|
| goose | 86 / 90 | 95.6% | direct result rows |
| DGC | 85 / 90 | 94.4% | 0 timeouts · 121.7 s/round · 293,675 output tokens |
| pi | 85 / 90 | 94.4% | direct result rows |
| OpenCode | 80 / 90 | 88.9% | direct result rows |
| Codex CLI | 79 / 90 | 87.8% | direct result rows |
Every manifest records qwen3.8:27b-orig, the same caller-declared digest, endpoint URL, and hardware. The runner verified context size; model weights were not independently verified. The saved manifest reports DGC 0.24.0; its runner commit was not recorded, so this is labelled a measured slice rather than a release-gating league.
Method, caveats, raw-source checksums and reproducible commands live on the full benchmark page →
From the founder
DGC’s bet is that the harness matters more than the model. We parse tool calls straight from the model’s text, so a local model with no function-calling API still runs the full read → edit → run → verify loop. The capability lives in the harness—not the weights.
08 Power tools
Built for real, long, autonomous work.
Optional modes for data-heavy tasks, verified unattended native-model runs, and turning your own sessions into scrubbed training data.
A Python workspace across turns
Load a dataset once, then operate on variables without repeating it into context.
Command: /code-action onRun until the check passes
On local/API routes, failures return to the next turn and completion waits for a real exit code. Delegated vendor turns bypass this gate.
Command: dgc --autonomous-gate "npm run check"Your usage becomes your data
Export heuristic-filtered sessions as bounded, credential-scrubbed JSONL; review them before training.
Command: dgc export-training --successful-onlyEvery turn is a point of return
Each turn DGC runs opens a recovery point—the exact conversation and the pre-edit bytes of every file the edit tools touch, saved before the tool runs. Rewind restores both, survives resume and compaction, and refuses a file-tool edit it cannot snapshot.
Command: /rewind09 Focus pane
Keep working while it works.
A second surface under the transcript. Open the file explorer, browse and pick files for the next prompt, watch the model’s edits land—the agent never stops streaming above it. And a measured estimate of when the turn will finish, so you can walk away.
AGENT TRANSCRIPT
I’ll read the implementation and tests, then fix the bounds.
→ read_file clamp.py · test_clamp.py
FILESAgent working above · browse below
PROJECTclamp-demo/Current workspace
FILES
README.md · preview
# Clamp example Keep a value within a range. Run the tests: python3 -m unittest -v
↑↓ browseEnter · add @pathq · return to prompt
❯Ask DGC anything…
Step 1 of 4 · Browse while it works
Scripted interface demo · explore the file names, or step through the sequence.
/files · a file explorer in the loop
Three columns, previews, copy/cut/paste, rename, trash with undo. Enter drops @path into your prompt; files the model changed this turn carry a mark.
Turn ETA · know when to come back
A range, not a promise: ~2–4 min left with the task count beside it, from this project’s own history and the turn’s task list. /eta stats reports how often it held.
/notify · walk away
A terminal notification, an editor toast, or your Stop hook when the turn ends. Your choice.
Native routes, real safety
Writes follow the permission mode and stay inside the project; nothing enters the model’s context unless you insert it. No shell, ever.
10 Live diff
See every change as it lands.
/diff opens a live view of everything the model has changed—every file with its added and removed line counts, and the diff itself—under the transcript while the turn keeps running. Select the lines you want to talk about and press Enter: they land in your next prompt as a quoted diff, so “why did you change these two lines?” quotes exactly those two lines.
AGENT TRANSCRIPT
I’ll inspect the bounds and make the smallest fix.
→ read_file clamp.py
DIFFWorking tree against HEAD · no changes yet
CHANGED FILES
No changes yet
clamp.py+1 −1
clamp.py · working tree vs HEAD
Changes appear here as the agent edits.
@@ -1,3 +1,3 @@
1
def clamp(value, lower, upper):2
"""Keep value within the bounds."""Read-only view · your working tree stays unchanged
❯Ask DGC about a change…
Step 1 of 4 · Watch changes arrive
Scripted interface demo · click the changed lines to select them.
/diff · what changed, live
Git’s own answer, read-only: the working tree against HEAD, staged and untracked files, +12 −3 per file, the whole set summed in the header. Re-read every two seconds while it is open, so edits appear as they land.
Select lines → your prompt
Space starts a selection, Enter attaches it as a fenced diff block with the file and line range; a takes the whole hunk. Nothing else about the file enters the model’s context.
Side by side, or one at a time
Wide terminals show the file list and the diff together; narrow ones swap between them with Tab. J / K step through files without leaving the diff.
Looking, not committing
It never writes, stages or runs a shell command. Approving or undoing a turn stays with /review and the editor’s Changes view; /diff is for looking and asking while the work is still moving.
11 What’s inside
A complete harness, not a wrapper.
On native local and API routes, the model proposes while DGC orchestrates boundaries, execution, persistence, verification, and the evidence you review. Subscription routes delegate the model-and-tool turn to the vendor CLI.
MODELproposes
PROPOSEstructured or text tool calls
failing test output · fed back to the next proposal
DGCorchestrates
PROMPTtask · repo · goal
PERMISSIONplan · default · edits · auto
EXECUTEfile tools · policy-gated shell · MCP dispatch
VERIFYreal tests · exit codes
OUTPUTdiffs · tests · traces
/mode setsactual contractGate the loop
Permission modes, plan review and thinking control.
permissionsplanthinking
Bound the work
Sandbox, MCP, artifacts and lifecycle hooks.
sandboxMCPhooks
Keep useful state
Standing goals, sessions, rewind and project memory.
/goalsessionsrewind
Run in parallel
Fleet sessions, isolated sub-agents and reusable skills.
fleettaskskills