Magnitude – Open Weights Coding Agent

TL;DR

TL;DR: Magnitude is a self-hostable coding agent (Apache-2.0) that runs frontier-class performance on open models like GLM-5.2 and DeepSeek. It deploys entirely inside your VPC — your code never touches third-party servers.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: magnitude.dev ← verified
• Source repository: github.com/magnitudedev/magnitude ← verified (34 stars, Apache-2.0)
• Pricing page: app.magnitude.dev/billing ← verified
• Docs: docs.magnitude.dev ← verified
• Terminal-Bench 2.1 results: docs.magnitude.dev/benchmark ← verified

What Is Magnitude?

Magnitude is a coding agent built around open-weight models. Unlike agents locked to GPT-4 or Claude, Magnitude runs entirely inside your infrastructure — your VPC, data center, or air-gapped environment. It currently supports GLM 5.2, Kimi K2.7 Code, and DeepSeek V4 Flash.

The product pitch is direct from the homepage: “The best coding agent for open models.” It scored 75.5% on Terminal-Bench 2.1 (89 tasks, 5 trials each) using GLM-5.2, versus Claude Code’s 78.9% on Opus 4.8 — but at roughly one-third the cost per successful trial.

The stack is TypeScript/Bun for the CLI and desktop UI, with a Rust daemon (acn) hosting the agent runtime, sessions, and display streams.

Setup Workflow

Step 1: Sign up for cloud access (or skip to Step 4 for self-hosted)

# Sign up at https://app.magnitude.dev — $5 free credits, no card required

Step 2: Install the CLI

# npm
npm i -g @magnitudedev/cli

# or bun
bun add -g @magnitudedev/cli

# or pnpm
pnpm add -g @magnitudedev/cli

Step 3: Start coding

magnitude

Paste your API key when prompted. Magnitude opens a terminal UI, indexes your project, and you interact via natural language.

Step 4: Self-hosted / VPC deployment

For enterprise deployments inside your own infrastructure:

1. Contact magnitude.dev for VPC pricing and setup
2. They deploy and tune the model serving stack inside your environment
3. SSO, RBAC, spend controls, and audit logs are included

Requirements: Linux (WSL on Windows), modern terminal (Ghostty or Kitty recommended — macOS Terminal.app has limited color support).

How It Works

clients (cli/web) → client-common → sdk → acn (daemon)

• CLI / Desktop — user-facing client, imports only from client-common and sdk
• acn — Rust daemon hosting the agent runtime, sessions, file operations, and display streams
• protocol — wire contract shared by SDK and ACN
• providers — model implementations for GLM, Kimi, DeepSeek

Session inspection

bun session list                              # list recent sessions
bun session events <id>                       # all events for a session
bun session search <keyword>                  # search across last 5 sessions
bun session projection <id> Window            # replay and dump named projection as JSON

Sessions are stored in ~/.magnitude/sessions/ with UTC timestamp folder names.

OpenTelemetry tracing

Magnitude runs a local OTEL collector at http://127.0.0.1:27686. Query it with curl:

curl http://127.0.0.1:27686/api/traces?service=<service>
curl http://127.0.0.1:27686/api/spans/<span-id>

Benchmark Results

From the official Terminal-Bench 2.1 results (445 total runs):

| Agent | Model | Success Rate | Cost / success | |---|---|---|---| | Claude Code | Opus 4.8 | 78.9% (351/445) | ~$1.20 | | Magnitude | GLM-5.2 | 75.5% (336/445) | $0.42 | | Claude Code | GLM-5.2 | 70.8% (315/445) | $0.60 | | OpenCode | GLM-5.2 | 50.8% (226/445) | $0.59 |

Magnitude on GLM-5.2 trails Claude Code on Opus 4.8 by 3.4 percentage points, but costs roughly one-third per successful trial.

Pricing

Cloud (no infrastructure needed):

• Pass-through pricing — per-token cost with no markup
• $5 free credits on signup, no card required
• Top-ups in $5 increments via Stripe

| Model | Input/1M | Cached input/1M | Output/1M | |---|---|---|---| | GLM 5.2 | $1.40 | $0.14 | $4.40 | | Kimi K2.7 Code | $0.95 | $0.16 | $4.00 | | DeepSeek V4 Flash | $0.14 | $0.04 | $0.28 |

Enterprise (self-hosted):

• Deployed inside your VPC or data center, optionally air-gapped
• SSO, RBAC, spend controls, audit logs, founder-level support
• Contact for pricing: calendly.com/tom-magnitude/30min

Practical Evaluation Checklist

• [x] Open-weight model support (GLM, Kimi, DeepSeek)
• [x] Self-hosted / VPC deployment option
• [x] Air-gap compatible (enterprise tier)
• [x] CLI-first workflow
• [x] Session inspection with bun session
• [x] OpenTelemetry tracing built in
• [x] Per-token pass-through pricing (no markup)
• [x] Free credits on signup
• [x] MIT/Apache-2.0 license on open-source components

Security Notes

• Enterprise tier: code never leaves your infrastructure
• Zero data retention on cloud tier for prompts and code
• RBAC and audit logs on enterprise

FAQ

Q: What terminal does Magnitude require? A: Works in any terminal, but renders best in Ghostty or Kitty. macOS Terminal.app has limited color support.

Q: Does Magnitude work on Windows? A: Requires WSL (Windows Subsystem for Linux). Native Windows is not yet supported.

Q: How does it compare to Claude Code? A: On Terminal-Bench 2.1, Magnitude on GLM-5.2 scores 75.5% versus Claude Code on Opus 4.8 at 78.9%. The cost per successful trial is roughly 3x lower ($0.42 vs $1.20).

Q: Can I use my own model provider? A: Currently supports GLM 5.2, Kimi K2.7 Code, and DeepSeek V4 Flash. Custom provider support is not yet documented.

Conclusion

Magnitude targets teams that want agentic coding capability without sending proprietary code to third-party APIs. The open-model approach (Apache-2.0, 34 GitHub stars) and the Terminal-Bench 2.1 numbers are compelling — particularly the cost-per-success metric. If you need Claude-class performance on open weights with VPC deployment, this is worth evaluating.

Try the cloud tier with $5 free credits at app.magnitude.dev, or contact them for enterprise self-hosted pricing.