# Blazorly — full text for coding agents

Blazorly is an open-source agentic coding harness (MIT). Give it a task and it reads, writes, runs commands, searches the web, and orchestrates sub-agents. Surfaces: Blazor web UI (default port 5080), headless CLI (`blazorly run`), JSON-RPC (`serve-stdio`), Agent Client Protocol (`serve-acp`).

Install (Linux/macOS):
curl -fsSL https://raw.githubusercontent.com/deepakkumar1984/blazorly-harness/main/installer/install.sh | bash

Install (Windows PowerShell):
powershell -c "irm https://raw.githubusercontent.com/deepakkumar1984/blazorly-harness/main/installer/install.ps1 | iex"

Data: ~/.blazorly (Windows: %LOCALAPPDATA%\blazorly). No Docker, no Python, no .NET SDK required to use a release binary.

Quick start:
blazorly
# then set provider + apiKey in Settings or ~/.blazorly/settings.json

{
  "provider": "deepseek",
  "model": "deepseek-v4-flash",
  "apiKey": "sk-…"
}

Settings UI (sidebar → Settings), five tabs:
- Model & API: provider, load models from the API, base URL, key (per provider)
- Custom providers: extra OpenAI-compatible routes
- Sessions & context: jsonl vs sqlite, context window, compaction, spill, local telemetry
- Capabilities: plugin toggles (web, skills, terminals, MCP, teams, System One, tool-gate, …)
- Permissions & sandbox: default preset for new sessions; optional E2B

Session chrome: Chat tab (conversation + tool cards) and Trajectory tab (turn timeline; raw events pages the log). Terminal drawer and Stats sit next to those tabs. Header chips: provider/model, effort, permission, context meter.

Providers: DeepSeek, OpenAI, Anthropic, xAI, Ollama, LM Studio, custom OpenAI-compatible routes. Keys resolve per request; env vars win if settings field is empty.

Permission presets (switch with /permission):
- full-access (default): tools on the host
- workspace-write: writes confined to the session folder; Linux Landlock jails bash/run_code
- read-only: no mutations

CLI:
blazorly                 # UI
blazorly run "job"       # headless; cwd is workspace
blazorly sessions
blazorly init            # draft AGENTS.md (dry-run; --write to create)
blazorly eval
blazorly decisions
blazorly serve-stdio
blazorly serve-acp
blazorly update
blazorly --version

Headless exit codes: 0 completed/max-tokens, 2 error/blocked, 3 aborted/interrupted, 1 harness failure.

Tools include: bash, read, write, edit, grep, glob, web_search, web_fetch, todo_write, subagent_start/send/list/interrupt, swarm, review, job_list/output/kill, skill, session_search and traces, MCP tools you configure.

Multi-agent: swarm (planner + parallel workers + reviewer; failed tasks re-dispatch), review (pass|fail|concerns), teams (enableTeams), background sub-agents. Children render in the parent chat Agents panel. Delegated children answer ask_user_question themselves.

Durable sessions: append-only JSONL or SQLite. Turns end with completed | aborted(cause) | interrupted | error | blocked | max-tokens. Cancelled tools close. Partial assistant text is kept. Process kill reloads with in-memory repair. Measured: median 0.3 ms mid-stream cancel (in-process); 148 ms median bash process-tree kill; 132,010-event production session cold-replays in 1.29 s.

System One / JEV-1 (optional): auto-plan, risk-gate, tool-gate. Tool gate is hybrid; failure returns the full tool list.

Source: https://github.com/deepakkumar1984/blazorly-harness
Site: https://blazorly.dev/