Start with the job, then find an open-source answer

AutoPTZ is a cross-platform desktop app that detects people in camera feeds and drives pan, tilt, and zoom controls to keep a selected subject framed. Users click a person to lock the target, while optional face recognition and appearance re-identification can recover identity after occlusion. It supports multiple cameras and PTZ backends including VISCA over USB or IP, ONVIF, and NDI. Tracking features require the appropriate models; when they are unavailable, affected controls are disabled while live preview remains available where possible.

AgentTrace reads local session logs from Claude Code, Codex CLI, Gemini CLI, Cursor, OpenCode, and other agent tools, then surfaces token usage, estimated cost, elapsed time, failures, latency, context pressure, and anomalies in a terminal UI. It can compare against a local baseline, export JSON, Markdown, or HTML reports, and fail CI on unhealthy runs. The inspection is local-first, with no hosted backend required for prompts, code, or logs.

OpenHarness connects to local models such as Ollama or to cloud APIs, then reads, edits, and searches code, runs commands, uses MCP servers, delegates to sub-agents, and operates headlessly in CI. Permission modes, command-risk analysis, file checkpoints, automatic Git commits, and post-edit verification make agent changes easier to inspect and undo. It requires Node.js 18+.

VibePod gives you one CLI for launching Claude, Codex, Gemini, and other coding agents in separate Docker or Podman containers. It records usage and HTTP traffic locally, then exposes a dashboard for side-by-side comparisons. A compatible container runtime is required.

ORCH deploys teams of Claude, Codex, Cursor, Pi, or arbitrary command-line workers, turns goals into tasks, and coordinates parallel execution with waiting, retries, reviews, and status tracking. Coding agents work in isolated Git worktrees and branches so changes do not reach the main branch until explicitly approved. Its shell adapter also allows the same orchestration engine to run editorial, analytics, operations, or other CLI-based processes.

AgencyBench V2 covers six capabilities across game, frontend, backend, code, research, and MCP work, with 32 real-world long-horizon scenarios and 138 tasks. A user-simulation agent supplies iterative feedback while a Docker sandbox runs functional and visual checks using rule-based, vision-based, and LLM judges. Scenarios require model and evaluator configuration, and game or frontend cases also need the remote Docker sandbox.

Codex Infinity extends the OpenAI Codex CLI with automatic next steps, testing, new improvement ideas, and chained goals. It can continue after each response or start another goal when one finishes, while detecting model providers from available credentials. Use it when unattended continuation is intentional, since the agent can keep making changes.

DvalinCode targets regulated teams that need AI coding under explicit policy, local data controls, and reviewable run records. Chat stays read-only, Cowork previews and asks for each write, while Code can run routines; the runtime also imports scanner findings into isolated remediation worktrees and supports OpenAI-compatible or local models.

Octomind is a Rust agent runtime that runs packaged specialist roles and can activate skills, MCP servers, or sub-agents during a session. It uses adaptive compaction to preserve important context, tracks costs with request and session caps, and enforces deterministic guards, hooks, and validators from scripts. The same binary supports interactive CLI, JSONL, daemon, WebSocket, and ACP modes.

ax ingests local agent sessions, tool calls, plans, skills, Git history, and related signals into a typed graph. It lets you recall past work, inspect skill and tool usage, estimate costs, surface friction, and review proposed changes to project instructions. A session-end retro loop feeds accepted lessons into later work. Ingest is local by default; publishing profiles or sessions is an explicit opt-in.

Letta Evals organizes agent evaluation from dataset and target through extractors, graders, rewards, and stored results. It supports JSONL or CSV data, multi-turn samples, multiple model handles, cached re-grading, deterministic or model-judge graders, and custom agent factories. Runs target Letta Code through a self-hosted or cloud server, so Python 3.11+, a running Letta service, and provider credentials are required.

Crab Code is a Rust-native coding agent that supports an interactive terminal UI, one-shot prompts, and non-interactive execution. It follows Claude Code-like tool and permission patterns while working with Anthropic, OpenAI, DeepSeek, Bedrock, Vertex, and custom providers. The project is under active development and is generally built from source with Cargo.

Agentplane adds an inspectable operational layer around local coding agents, recording task intent, plans, verification, context, and an Agent Change Record in repository artifacts. It supports fast work in the current checkout as well as per-task branches, worktrees, and PR handoffs for stricter review. It targets developers and teams that need reproducible agent work and requires Node.js 24+, Git, and a terminal.

agent-lsp orchestrates existing language servers and exposes their code intelligence to AI agents through MCP. It keeps language servers warm between sessions, combines many routine calls into batch operations, and supports workflows such as caller analysis and speculative edits. The project documents 65 tools, 24 agent workflows, and CI-verified support for 30 languages.

ActPlane applies labeled information-flow and causal-ordering rules below the tool layer, following an agent’s entire process tree, file edges, and network connections. It can block, kill, or notify on violations and feed a reason back to the agent; it requires Linux 5.8+ with BTF and elevated capabilities for the enforcement engine.

skill-optimizer provides both an agent skill for authoring eval suites and a local CLI for executing them in Docker against OpenRouter models. Cases and suites can be repeated across trials to benchmark skill behavior, debug failures, and compare reliability across models. The workbench can also expose hidden services such as MCP servers to the agent during a controlled test.

Use handoff to run execution work in a background CLI, return a stable result path to the current conversation, and resume the same task later with its prior context. It can route simple work to DeepSeek and request a second opinion from Codex or Opus. Claude Code or Codex must already be installed and logged in, and DeepSeek needs its own token configuration.

San is a native Go terminal harness distributed as a single binary of roughly 12 MB. It lets you switch models and search backends, attach plugins or MCP servers, and control actions through ask, auto-accept, or autopilot modes. Sessions can resume or fork, and a local inspector can replay transcripts; provider credentials are still required for the services you choose.