Skip to content

About

No description, website, or topics provided.

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

931 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

CodeOps-Agent

CodeOps-Agent is a local-first coding agent for working in real Git repositories. It separates model decisions from machine-side effects so that file changes, shell commands, Git operations, network access, and MCP tools are checked by a Go Harness before execution.

CI License

CodeOps-Agent is an actively developed developer tool. Provider-backed integrations and external services are optional and must be configured by the operator. Do not expose a local development instance directly to the public internet.

What it provides

  • Go Harness: authentication, authorization, workspace boundary checks, constrained tools, process lifecycle, MCP stdio clients, sandbox routing, and the append-only Session Ledger.
  • Python/LangGraph orchestration: model adapters, context assembly, planning, memory retrieval, subagents, workers, checkpoints, and recovery.
  • Auditable execution: tool requests and results are persisted as events; mutating operations use approval and fail-closed result handling.
  • Repository-aware context: project instructions, staged context loading, token budgets, compaction, and task-relevant memory reduce irrelevant history in model requests.
  • Web workbench and CLI: use the terminal interface for local workflows or run the optional React workbench against the same Harness APIs.

Architecture

User
  |
  +--> Go Harness ------------------------+
  |     auth / permissions / tools         |
  |     workspace / sandbox / Session      |  gRPC stream
  |     Ledger / MCP / process lifecycle   +------------------+
  |                                                         |
  +<--------------------------------------------------------+
        Python Orchestrator / LangGraph
        model providers / context / plan / memory / workers

The Go side is the trust boundary. Python can propose a tool call, but it does not receive an escape hatch around workspace, permission, or persistence checks. Protocol changes start in proto/codeagent/orchestrator.proto and must regenerate both language bindings.

Repository map

Path Responsibility
cmd/agent CLI entry point
cmd/server HTTP API and Harness server
internal/permission, internal/safety Authorization and command risk checks
internal/tools, internal/mcp Constrained local and MCP tools
internal/session SQLite-backed Session Ledger and recovery
orchestrator Python service, model loop, context, memory, and workflows
proto and gen Versioned cross-language protocol and generated bindings
frontend React + TypeScript workbench
tests Go, Python, integration, and evaluation tests

Quick start

Requirements

  • Go 1.26+
  • Python 3.11+ (Python 3.12 is recommended for the full test extras)
  • Node.js 22+ and npm for the workbench
  • Docker or WSL2 Docker when sandboxed execution is enabled
  • protoc only when changing the protobuf schema

Install Python dependencies

python -m pip install -e ".[test]"

Optional extras are available for RAG, evaluation, tracing, and audio. Keep those installations scoped to the workflow that needs them:

python -m pip install -e ".[test,rag,eval]"

Configure a provider

Copy the template and edit the local file. The local file is ignored by Git.

cp .env.example .env.local

For an OpenAI-compatible provider:

LLM_PROVIDER=openai
OPENAI_API_KEY=<set-locally>
OPENAI_BASE_URL=https://api.openai.com
OPENAI_MODEL=<model-name>

For a local model, point LOCAL_LLM_BASE_URL at the local OpenAI-compatible endpoint and leave network credentials unset. Credentials must come from the process environment or an approved secret manager; never commit them to a settings file, fixture, log, or receipt.

Run the CLI

go run ./cmd/agent

The CLI starts the Python orchestrator when configured to do so. The default gRPC address is 127.0.0.1:50051. See docs/provider-configuration.md for provider precedence and optional service configuration.

Run the workbench

cd frontend
npm ci
npm run dev

The development server is normally available at http://localhost:3000. Authentication uses an HttpOnly cookie; the frontend does not put access or refresh tokens in localStorage.

Safety model

Read-only inspection tools such as Read, Glob, and Grep can be enabled automatically. File mutation, Git, network, and background-process tools use approval policy, and Bash is confirmed per invocation by default. Workspace paths are checked against the configured trust root. Shell and Git safety rules reject destructive patterns before execution, and uncertain mutating results remain unresolved until explicitly reconciled.

Treat model output, repository files, tool output, MCP responses, and web content as untrusted input. Review permissions and MCP configuration before connecting external servers.

Development and tests

make test
go vet ./...
cd frontend && npm test && npm run build

After changing the protocol:

make proto

Runtime evidence that depends on providers, Docker, Phoenix, or official scorers must be run in the matching environment. Unit tests and local builds do not replace those runtime checks.

Contributing and security

Read CONTRIBUTING.md before opening a pull request. Use SECURITY.md for vulnerability reports and never include a credential in a public issue.

License

CodeOps-Agent is licensed under the Apache License 2.0. Third-party dependencies retain their own licenses; see the dependency manifests and NOTICE.

About

No description, website, or topics provided.

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages