Grinta in action — click the preview to watch the full demo.
Grinta plans, edits, runs commands, debugs failures, validates the result, and continues until the task is finished. Its control plane and session state stay local; inference can use a major hosted provider or a local model.
Install · Documentation · Architecture · Showcase
4h 33m autonomous run · 16,393 events · 373 tool outcomes · no additional user messages
- Runs its control plane, execution, session history, and checkpoints locally
- Works across OpenAI, Anthropic, Google, OpenRouter, Ollama, and LM Studio
- Uses real Language Server Protocol (LSP) and Debug Adapter Protocol (DAP) integrations
- Recovers from provider failures, malformed tool calls, timeouts, and context pressure
- Keeps a durable event ledger, checkpoints, reverts, and inspectable audit trails
- Exposes Chat, Plan, and Agent workflows in a focused terminal UI
Install the stable version from PyPI or source:
git clone https://github.com/josephsenior/Grinta-Coding-Agent.git Grinta
cd Grinta
pipx install -e .
grintaOr install directly from PyPI:
pipx install grinta
grintaOptional integrations are available as grinta[rag], grinta[browser], or grinta[all]. See the Quick Start for Windows, WSL2, Linux, macOS, and provider setup.
| Long-horizon execution | Failure recovery | Raft key-value store |
|---|---|---|
Ran autonomously for 4h 33m, processed 16,393 events, and reached FINISHED through provider and runtime failures. |
Read failing test output, isolated defects, edited the affected code, and reran validation without a follow-up prompt. | Built a Raft-backed key-value store, recovered from a race-condition failure, and finished with 39/39 tests passing. |
| Read the report | Inspect the case study | Watch and inspect |
flowchart TD
A["Task"] --> B["Plan and act"]
B --> C["Local tools"]
C --> D["Observe and validate"]
D -->|"More work or failure"| B
D -->|"Quality gates pass"| E["Finish"]
The SessionOrchestrator coordinates model intent, the safety and validation pipeline, local execution, and observations. A durable event stream records the session while completion gates reduce premature “done” states. Read the architecture guide for the full component map and reliability guide for recovery behavior.
This table describes Grinta's design commitments—not an unverified performance ranking against other agents.
| Capability | Grinta |
|---|---|
| Local control plane | Yes |
| Provider-agnostic inference | Yes |
| Local-model support | Yes |
| LSP integration | Yes |
| DAP debugging | Yes |
| Durable session ledger | Yes |
| Checkpoints and revert | Yes |
| Terminal UI | Yes |
| Cloud account required | No |
Grinta is designed for developers who value local execution, provider freedom, inspectable state, and recovery-oriented workflows that can survive long tasks.
Grinta runs commands with the privileges of the local user. Policy gates, secret masking, and optional process isolation reduce risk but do not make hostile code safe. Use a VM or container for untrusted repositories and read the security checklist before increasing autonomy.
Grinta is seeking contributors interested in:
- Agent reliability and recovery
- LSP and debugger integrations
- Local-model compatibility
- Terminal user interfaces
- Autonomous-agent evaluation
Start with a good-first-issue or read the Contributor Map. The full development workflow is in CONTRIBUTING.md.
- Use Grinta: Quick Start · User Guide · Settings · Troubleshooting
- Understand it: Architecture · Reliability · Support Matrix · Security
- Contribute: Contributor Map · Developer Guide · CI · Roadmap
- Read the engineering story: The Book of Grinta
Created and maintained by Youssef Mejdi. Released under the MIT License.