DevSwarm

Your AI coding assistant, now with a team.

Drop one MCP server into Codex, Amp, or Claude Code and get 37 tools for spawning parallel agents, running task pipelines, and doing multi-step code work — without leaving your existing workflow.

Quick Start · Features · All 37 Tools · How It Works · Contributing

---

The Problem

You're already using Codex, Amp, or Claude Code. It writes code, fixes bugs, answers questions. But it's still one agent doing one thing at a time.

You: "Find all the memory leaks in this codebase and fix them"

  Orchestrator decomposes the task
       │
  ┌────┼────┐
  ▼    ▼    ▼
 [W1] [W2] [W3]   ← parallel agents, each owns a subsystem
  │    │    │
  └────┼────┘
       ▼
  Synthesizer → one clean report back to you

DevSwarm is an MCP server that gives your AI assistant the ability to orchestrate itself — spawning sub-agents, running parallel workloads, and chaining multi-step task pipelines. No new UI. No new workflow.

---

⚡ Quick Start

Option 1: Download a binary (recommended)

Grab the latest release for your platform from GitHub Releases.

Option 2: Build from source

git clone https://github.com/justrach/codedb.git
cd codedb
zig build          # builds zig-out/bin/devswarm
zig build test     # run all tests

Requirements: Zig 0.15.x, codex and/or claude CLI on PATH, Git

---

Connect to your AI assistant

Claude Code

Add to ~/.claude.json:

{
  "mcpServers": {
    "devswarm": {
      "command": "/path/to/devswarm",
      "args": ["--mcp"],
      "env": { "REPO_PATH": "/path/to/your/repo" }
    }
  }
}

Then run /mcp to verify — you'll see 37 tools added to your assistant.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.devswarm]
command = "/path/to/devswarm"
args = ["--mcp"]
env = { REPO_PATH = "/path/to/your/repo" }

Amp

Add to your Amp MCP config:

{
  "mcpServers": {
    "devswarm": {
      "command": "/path/to/devswarm",
      "args": ["--mcp"],
      "env": { "REPO_PATH": "/path/to/your/repo" }
    }
  }
}

---

🚀 What You Can Do

Swarms — parallel agents on big tasks

run_swarm("Audit the entire auth system for security issues", max_agents=5)

An orchestrator breaks the task into sub-tasks. Workers run in parallel. A synthesizer combines everything. You get one answer instead of five tabs.

Task Chains — multi-step pipelines

run_task("Fix the race condition in src/queue.zig", preset="reviewer_fixer")

Built-in presets chain agents together automatically:

Preset	Pipeline
finder_fixer	find the issue → fix it
reviewer_fixer	review → fix reported issues
explore_report	deep exploration → structured report
architect_build	design → implement

Review-Fix Loops — iterate until clean

review_fix_loop("Check for memory leaks", max_iterations=3)

Runs reviewer → fixer → reviewer again, until the reviewer says NO_ISSUES_FOUND or hits the iteration cap.

Single Agents with Role + Model Routing

run_agent("Explain the PPR algorithm", role="explorer", mode="deep")

Each agent gets the right model automatically:

Role	Model	Does
finder	Sonnet	Search and locate
reviewer	Sonnet	Review for correctness
fixer	Sonnet	Apply fixes (writable)
explorer	Sonnet	Deep codebase exploration
architect	Opus	System design decisions
orchestrator	Opus	Decomposes swarm tasks
synthesizer	Sonnet	Combines agent outputs
monitor	Haiku	Lightweight checks

Mode	Use when
smart	Most tasks
rush	Quick answers
deep	Hard problems, architecture
free	Minimize cost

---

🔧 Full Tool List (37 tools)

Agents run_agent · run_swarm · run_task · review_fix_loop · run_reviewer · run_explorer · run_zig_infra

Planning decompose_feature · get_project_state · get_next_task · prioritize_issues

Issues create_issue · update_issue · close_issue · get_issue · create_issues_batch · close_issues_batch · link_issues

Git create_branch · get_current_branch · commit_with_context · push_branch · recently_changed · git_history_for

Pull Requests create_pr · get_pr_status · list_open_prs · merge_pr · get_pr_diff · review_pr_impact

Code Intelligence blast_radius · relevant_context · symbol_at · find_callers · find_callees · find_dependents

Repo set_repo

---

⚙️ How It Works

DevSwarm is a provider-agnostic runtime. When you call run_agent, it:

1. Resolves — picks backend (Claude or Codex), model tier, system prompt, and tool preamble based on role + mode + what's available on your PATH
2. Dispatches — spawns the agent on the right backend, falls back automatically if one isn't available
3. Returns — streams output back through MCP

System prompts are assembled dynamically from agency rules, role instructions, mode guidance, and auto-detected tool availability (zig tools → ripgrep → grep). No hardcoded prompts.

---

Telemetry

devswarm collects anonymous usage telemetry to help improve the project. This is enabled by default.

What's collected

• Agent roles used (e.g. "finder", "reviewer", "fixer")
• Model names (e.g. "claude-sonnet-4-6")
• Token counts (input/output per worker)
• Wall time and estimated cost
• Worker count and parallelism metrics

What's NEVER collected

• Your code, file contents, or diffs
• Prompts, task descriptions, or agent outputs
• Repository names, file paths, or branch names
• Any personally identifiable information

How to opt out

Edit .devswarm/config.toml:

[telemetry]
enabled = false

Or set the environment variable:

export DEVSWARM_TELEMETRY=false

You can opt out at any time. The telemetry preference is set during onboarding and stored in .devswarm/config.toml.

---

Contributing

Contributions are welcome. See CONTRIBUTING.md for guidelines before opening a PR.

License

MIT — see LICENSE

---

Full changelog: README-changelog.md