Open Source · MIT License

Autonomous agents for every project. Deployed where your team works.

CodeSpar deploys persistent AI coding agents to WhatsApp, Slack, Telegram, and Discord. Each project gets its own agent that monitors builds, investigates failures, proposes fixes, and orchestrates deploys. Self-hosted. Docker Compose up in 5 minutes.

terminal
6Agent types
5Channels supported
L0–L5Autonomy levels
10Security layers
<5 minTo first agent
Open Source on GitHubMIT LicenseTypeScript · Docker · Claude Agent SDKSelf-hosted · No vendor lock-in

What is CodeSpar

Not a chatbot. Not a code assistant. A multi-agent platform.

CodeSpar gives every project a persistent AI agent that lives in your messaging channels. Agents accumulate codebase context over time, monitor CI/CD events, and coordinate with specialized sub-agents. You @mention them. They handle the rest.

Monitor

SB
Squad Backend
you, @codespar, +3 others
@codespar status build
09:14
code<spar> agent-gw
Build #348 — api-gateway (main)
  142/142 tests | 87.5% cov | 3m12s
09:14

Your agent tracks every build, test run, and deploy.

Fix

# backend
<>
agent-gwAPP2:14 AM
Build #349 broken. Regression in auth.ts:47. Found fix.
Diff: +2 -1 in src/middleware/auth.ts

Agents investigate failures, propose fixes, and wait for your approval.

Deploy

#deployments
<>
CodeSparBOT
Deploy Approval Required
Deploy api-gateway to staging?
Approval: 1/2 required (quorum)

Production deploys require quorum. Approvals work across channels.

Architecture

Six agent types. One orchestration layer.

Every project gets a persistent Project Agent. When work needs doing, it spawns specialized ephemeral agents.

Project Agent
Persistent

Always-on. Monitors repo, CI/CD, channels. Maintains codebase context.

Task Agent
Ephemeral

Executes coding tasks in isolated Docker containers.

Review Agent
Ephemeral

Analyzes PRs. Checks code quality. Auto-approves low-risk per policy.

Deploy Agent
Ephemeral

Orchestrates deploys. Pre-checks, approvals, health monitoring, rollback.

Incident Agent
Ephemeral

Investigates production errors. Correlates with recent changes.

Coordinator
Persistent

Cross-project orchestration. Cascading deploys, shared locks.

Trust Model

Agents earn autonomy. You set the pace.

L0
PassiveOnly responds when addressed
L1
NotifyMonitors and alerts. Default.
L2
SuggestProposes actions proactively
L3
Auto-LowAuto-executes low-risk actions
L4
Auto-MedAuto-executes medium-risk
L5
Full AutoAutonomous within policy bounds

Regardless of autonomy level, agents never auto-execute: production deployments, data migrations, security-sensitive changes, or infrastructure modifications. These always require human approval.

Channels

Same @mention. Every platform.

W
WhatsApp
S
Slack
T
Telegram
D
Discord
C
CLI

Your team uses WhatsApp for quick coordination, Slack for structured workflows, Discord for community. CodeSpar works in all of them. Same syntax: @codespar [command]

WhatsApp as a first-class channel. No Business API required.

Get Started

Three steps. Five minutes. First agent live.

01

Deploy

One command. PostgreSQL, Redis, agent supervisor — all running.

$ docker compose up -d
02

Connect

Link a WhatsApp group (QR scan), Slack channel (OAuth), or any channel. A persistent Project Agent spawns automatically.

03

@mention

Type @codespar in any linked channel. Monitor builds. Instruct fixes. Approve deploys. From your phone.

Open Source

MIT License. No asterisks.

Infrastructure this critical should be auditable, extensible, and owned by the teams that run it. CodeSpar is fully open source. No open-core restrictions. No feature gates. The entire multi-agent platform — every agent type, every channel adapter, the supervisor, the policy engine, the audit system — is in the repo.

Fork it. Self-host it. Extend it. Write a custom agent in 300 lines of TypeScript.

repo structure
codespar/
├── agents/
│ ├── supervisor/ # Lifecycle, health, scaling
│ ├── project/ # Persistent, per-project
│ ├── task/ # Ephemeral, Claude Code exec
│ ├── review/ # PR analysis, code quality
│ ├── deploy/ # Deploy orchestration
│ ├── incident/ # Error investigation
│ └── coordinator/ # Cross-project
├── channels/
│ ├── whatsapp/ # Baileys + anti-ban
│ ├── slack/ # Bolt.js + Block Kit
│ ├── telegram/ # grammy + inline KB
│ ├── discord/ # discord.js + embeds
│ └── cli/ # Terminal adapter
├── core/ # Router, policy, audit
└── docker-compose.yml

Security

10 defense layers. Agents inherit permissions, never own them.

Every action with side effects requires verification proportional to its impact. Agents act on behalf of users, never independently. The security model is layered.

01
Message Filter
02
Channel Config
03
Identity Resolution
04
RBAC (6 roles)
05
ABAC Policies
06
Agent Sandbox
07
Prompt Injection Defense
08
Execution Sandbox
09
Output Validation
10
Audit Trail

Built for teams from day one.

RBAC (6 roles)ABAC PoliciesCross-channel ApprovalImmutable Audit TrailsKill Switch

Enterprise SSO, SOC 2 compliance path, and hosted deployment are on the roadmap. Self-host the open source version today with full RBAC, ABAC, and audit capabilities.

FAQ

Common questions

No. CodeSpar deploys persistent, autonomous agents per project. They monitor, investigate, propose, and execute. A chatbot responds to questions. CodeSpar agents take action on your behalf.

WhatsApp connects via Baileys (linked device protocol) with anti-ban strategies: random delays, typing simulation, rate limits. Risk is real (~30%/year for aggressive usage). That’s why CodeSpar is multi-channel: if WhatsApp goes down, Slack, Telegram, and Discord continue. Your agents never go offline.

Devin is closed-source, single-channel, and $500/month. CodeSpar is open source (MIT), multi-channel (including WhatsApp), multi-agent (6 agent types), self-hosted, and free. Different product, different philosophy.

CodeSpar is in active development. MVP supports WhatsApp + CLI with Project Agents and Task Agents. We ship in public and the roadmap is on GitHub. Production-grade features (RBAC, approval quorum, audit trails) are designed from day one.

Yes. The architecture is modular by design. Each agent type and channel adapter is a separate package. Writing a custom agent takes ~300 lines of TypeScript. See the ‘Writing a Custom Agent’ tutorial in the docs.

Your projects deserve dedicated agents.

Deploy your first agent in 5 minutes. Open source. Self-hosted. MIT license.

code<spar>