The Autonomous AI Coding Agent
Engineered for VS Code.
Not another CLI wrapper in an iframe. Coda uses native Language Server Protocol intelligence and real compiler diagnostics, and only reports work as done when your project's own commands confirm it.
VS Code already open? Install directly in VS Code →
- MIT LicensedFree to use
- Local-FirstRuns fully offline with Ollama
- No Lock-InSwitch models anytime
- 35+ ProvidersCloud, local & Copilot
Verification is a Contract,
Not a Vibe.
Stop debugging your AI's hallucinations. Many assistants report success because the model says so. Coda executes authoritative compiler and test commands, validates process exit codes, and audits git state before reporting completion.
auth.ts, updated all dependent imports across your codebase, and all 42 tests pass with zero errors! Let me know if you need anything else! 🚀"
- ✕ Did not execute tests — claimed passes through pure prompt hallucination.
- ✕ Hallucinated export
verifyTokenSynccausing runtime crashes. - ✕ Broke TypeScript build:
TS2305: Module has no exported member. - ✕ Secret leakage: Printed API tokens and sensitive env vars into conversational log.
tsc --noEmit vitest run tests/auth secret scrubbing atomic diff Coda doesn't mark a task complete because the model says it is. It runs your project's own commands, checks their exit codes, audits git state, and enforces protected-path rules before it reports completion.
LSP Over Blind Grep.
Native Extension Host Intelligence.
Why do standard agents hallucinate on monorepos? Because regex text search is blind to types, inheritance, and AST scopes. Coda runs directly in the VS Code host with direct access to compiler language servers.
Semantic LSP Tools vs. Blind Regex
Instead of text search that also matches comments and dead code, Coda issues native vscode.executeDefinitionProvider, findReferences, and type hierarchy requests directly to TypeScript, Rust Analyzer, Pyright, or gopls.
Pluggable Event Lifecycle & Safety Rings
Every tool invocation passes through an interceptor pipeline (EventResult.modify). Hard blocks protect .env, .git/, lockfiles, and CI configuration.
Explorer Sub-Agents & Clean Distillations
Parent contexts degrade when inundated with 20 file dumps. Coda spawns lightweight explorer sub-agents on fast, low-cost model tiers, limited to read-only and LSP tools.
Engineered for Developers Who
Demand IDE Depth.
An AI agent shouldn't feel bolted on. Coda deeply weaves into VS Code's debugger, notebook runtime, git worktrees, and editor history.
Adapter-Aware DAP Debugger
DebuggerConnects natively to the VS Code Debug Adapter Protocol. Evaluates expressions in paused runtimes, queries call stack frames, inspects scope variables, and sets conditional breakpoints automatically during bug hunts.
Official Jupyter Engine
Notebooks
Edits .ipynb files through official vscode.NotebookEdit APIs. Never corrupts raw JSON structure, generates pre-edit snapshots, and supports markdown/code cell execution.
Mid-Run Steerability
In-Flight
Notice an agent heading down the wrong rabbit hole? Don't abort a 45-second run and discard completed work. Inject guidance ([Steer]) directly between tool rounds with atomic run-token synchronization.
Ergonomic Chat Cockpit
Cockpit
A compact activity stream shows only the most recent steps, so long runs stay readable. Includes safe "Insert at Cursor" with native editor undo stack (Ctrl+Z) and composer history navigation.
35+ Model Freedom
No Lock-In
Connect Anthropic Claude, OpenAI, Google Gemini, Groq, fully offline Ollama models, or zero-key vscode-lm (GitHub Copilot). Never be held hostage by a single cloud provider.
Background Git Worktrees
SandboxedRun autonomous multi-file refactorings and extensive test suites in isolated git worktrees. Keep typing and coding in your active editor window without unexpected file reloads or dirty checkout clashes.
Show, Don't Tell:
Watch Coda Execute & Verify.
Step through three example agent runs: a refactor, a failing-test diagnosis, and a live debugger inspection. Each one ends only when a real verification command passes.
Refactor the session cache to an LRU with TTL eviction, then verify.
- Searching workspace symbols
- LSP goToDefinition
- Editing src/cache/session.ts+7 −2
- Verification gate: npm testexit 0
Simulated session for illustration. Click inside the demo, then press 1–3 to switch workflows, R to replay, or D to view the diff.
35+ Supported Providers.
Zero Vendor Lock-In.
Bring your own developer API keys, run 100% offline with local GGUF models via Ollama or LM Studio, connect enterprise VPC endpoints, or leverage your existing GitHub Copilot subscription without configuring keys.
37 providers shown
Anthropic
RecommendedStrong agentic coding and tool use; Haiku-class models power fast explorer sub-agents.
Models:Claude Opus, Sonnet & HaikuCommercialDirect to providerOpenAI
PopularDirect developer API with streaming tool calls and reasoning effort controls.
Models:GPT and o-series reasoning modelsCommercialDirect to providerGoogle Gemini
Long ContextVery large context windows for repo-wide analysis and multimodal input.
Models:Gemini Pro & FlashCommercialDirect to providerDeepSeek
Low CostCapable open-weight reasoning models at very low per-token prices.
Models:DeepSeek chat & reasoning modelsCommercialDirect to providerxAI
Fast general-purpose and reasoning models through the xAI API.
Models:Grok modelsCommercialDirect to providerMistral AI
Code-specialised Codestral alongside general-purpose Mistral models.
Models:Codestral, Mistral Large & SmallCommercialDirect to providerCohere
Enterprise-oriented models with strong retrieval and tool use.
Models:Command modelsCommercialDirect to providerPerplexity
Web-grounded answers for live documentation lookups.
Models:Sonar modelsCommercialDirect to providerAI21 Labs
Hybrid SSM-Transformer models with long context windows.
Models:Jamba modelsCommercialDirect to provider
Free, Independent, and
Privacy-First.
Coda is free and MIT-licensed. No VC backing, no lock-in, and no telemetry proxy sitting between your code and your model provider.
Local Privacy Guarantee
Your keys, your models, your code. Coda communicates directly from your machine to your chosen AI provider. No telemetry proxy, no code indexing on third-party servers, and no hidden tracking.
☕ Support Independent Coda Development
Coda is independent and 100% free. If Coda saved you hours of debugging, caught a breaking change with its delivery ledger, or made your daily coding workflow smoother, consider buying a coffee to support maintenance, model benchmark evaluations, and continuous updates.