Agentic software engineering at predator speed.

Open-source, Rust-native multi-agent coding terminal.
Four specialist agents, each in its own git worktree, under your supervision — on your machine, with your own key.

Install · Quickstart · Features · Safety · Architecture · Contributing

---

At A Glance

Agentic REPL raptor plans, edits and runs commands with your consent	Four-agent swarm frontend · backend · test · security, in parallel	Worktree isolation every agent works in its own git worktree
Supervised review review → diff → test per worktree; raptor merge --execute for task worktrees, typed confirmation	Deny-by-default tool policy catastrophic commands hard-denied, egress confirmed	Local-first your machine, your DeepSeek key, nothing hosted

How A Swarm Run Flows

                         ┌─ frontend-agent ──▶ .raptor/worktrees/<task>/frontend-agent ─┐
  raptor swarm run       ├─ backend-agent  ──▶ .raptor/worktrees/<task>/backend-agent  ─┤   raptor review <task> [role]
  "Build a SaaS   ──▶    ├─ test-agent     ──▶ .raptor/worktrees/<task>/test-agent     ─┼─▶ raptor diff   <task> [role]  ──▶  you commit and merge what you accept
   landing page"         └─ security-agent ──▶ .raptor/worktrees/<task>/security-agent ─┘   raptor test   <task>              from each role worktree, with git
                               live dashboard · tool policy · consent prompts              nothing lands without you

Swarm output stays uncommitted inside the role worktrees until you have reviewed, diffed and tested it; you then commit and merge what you accept from each worktree with git. The single-task path (raptor task) adds a supervised raptor merge <task-id> --execute that only runs after a typed confirmation.

What It Does

• raptor opens an interactive agentic REPL: after a per-session file-system consent prompt the agent inspects the project, plans, edits files, and runs commands through a deny-by-default tool policy with developer confirmations.
• raptor swarm run "mission" (or /swarm inside the REPL) launches a live four-specialist swarm — frontend, backend, test, and security agents — each working in its own isolated git worktree behind a real-time dashboard.
• raptor swarm "task" and raptor swarm implement <task-id> <role> produce supervised role plans (RAPTOR_PLAN.md) and implementation proposals (RAPTOR_IMPLEMENTATION_PROPOSAL.md) without touching your tree.
• raptor review, raptor diff, raptor test, and raptor merge inspect, test, and merge worktree results with explicit confirmation.
• raptor chat for plain model chat, raptor tui for a read-only dashboard, raptor doctor for environment diagnostics.

DeepSeek is the implemented execution provider today (default model deepseek-v4-pro). The Provider trait and provider catalog are the extension point for others; see ROADMAP.md.

Install

Prebuilt binaries for Linux (.deb), macOS (universal), and Windows are published on https://raptorlabs.dev with SHA-256 checksums. On Debian/Ubuntu:

sudo apt install ./raptor-code_0.6.5_amd64.deb
raptor --help

From source

Prerequisites: Git, and Rust 1.88 or newer with Cargo (the repository pins 1.95.0 through rust-toolchain.toml).

git clone https://github.com/RaptorLabsHQ/raptorcode.git
cd raptorcode
cargo install --path crates/raptor-cli --force --locked

The package is named raptor-cli; the installed commands are raptor and raptorcode. Build the Debian package locally with ./scripts/package-deb.sh (written to dist/).

DeepSeek Setup

Get an API key from https://platform.deepseek.com/api_keys. RaptorCode reads DEEPSEEK_API_KEY from the environment, or asks for the key once on first run and stores it in ~/.raptor/.credentials (owner-only permissions). The key is never printed, logged, or committed, and /api inside the REPL manages it without echoing it.

Quickstart

raptor                      # startup screen, then the interactive REPL
raptor doctor               # check the local environment
raptor chat "Explain this project"
raptor swarm run "Build a SaaS landing page with auth"
raptor review <task-id>     # inspect what the swarm produced
raptor diff <task-id>
raptor test <task-id>
raptor merge <task-id> --execute
raptor tui                  # read-only dashboard

From a source checkout, the same commands run through Cargo:

cargo run -p raptor-cli --bin raptor -- --help
cargo run -p raptor-cli --bin raptor -- doctor

Command Surface

raptor
raptor --version
raptor init
raptor tui
raptor provider list
raptor agent list
raptor agent show <id>
raptor agent create --role <role> [--task <task-id>] [--provider <provider-id>] [--model <model-id>]
raptor agent policy
raptor chat "message"
raptor task "task description"
raptor task list
raptor task show <id>
raptor task abandon <id>
raptor task cleanup <id>
raptor swarm "task description"
raptor swarm implement <task-id> <role>
raptor swarm run "mission description"
raptor review <task-id> [role]
raptor diff <task-id> [role]
raptor test <task-id> [-- <command>]
raptor merge <task-id>
raptor merge <task-id> --execute
raptor doctor

Safety Model

RaptorCode treats model output and user-provided identifiers as untrusted.

• Agent work happens in isolated git worktrees under .raptor/worktrees/; task, swarm, and agent metadata live under .raptor/tasks/, .raptor/swarms/, and .raptor/agents/. .raptor/ is git-ignored.
• Task ids, role names, branch names, and paths are validated; path traversal is rejected and writes stay inside the worktree boundary (symlink-safe).
• Git and test commands use structured process arguments, never shell interpolation.
• Tool execution runs through a deny-by-default policy: catastrophic commands are hard-denied, privileged or network-egress commands and writes outside the worktree ask for confirmation, and shell wrappers cannot hide a denied command.
• The planning (raptor swarm) and proposal (raptor swarm implement) paths never execute model output or apply patches.
• raptor chat sends only your message; it does not send repository contents, .env files, secrets, task metadata, or worktree files.
• The dashboard is read-only. Merging requires --execute and an exact typed confirmation.

Developers remain responsible for reviewing plans, proposals, diffs, tests, and merges. The full design is in docs/security.md.

Current Limitations

• Automatic patch application and fully autonomous loops outside the consented REPL and swarm paths.
• Providers other than DeepSeek, MCP, a desktop UI, cloud sync, team accounts, or an enterprise policy engine.
• OS-level sandboxing of the local agent (namespaces, seccomp, landlock); the userspace tool policy is a backstop, not a sandbox.

Persisted agent sessions and permission policy modeling exist as early local metadata foundations and are not yet an active policy enforcement system.

Repository Layout

.
|-- crates/
|   |-- raptor-agent/        frozen agent session model (see ROADMAP.md)
|   |-- raptor-cli/          the `raptor` binary: REPL, commands, swarm orchestration
|   |-- raptor-core/         shared types, validators, errors
|   |-- raptor-git/          argv-based git operations and worktrees
|   |-- raptor-providers/    provider trait, catalog, DeepSeek client
|   |-- raptor-sandbox/      reserved (empty)
|   |-- raptor-store/        atomic local state under .raptor/
|   |-- raptor-tools/        agent tool implementations
|   `-- raptor-tui/          ratatui dashboard and startup screen
|-- docs/
|-- packaging/               AppStream metainfo for the .deb
|-- scripts/                 release.sh, package-deb.sh
|-- templates/               swarm role templates
|-- CHANGELOG.md
|-- CONTRIBUTING.md
|-- LICENSE-APACHE
|-- LICENSE-MIT
|-- ROADMAP.md
|-- SECURITY.md
|-- THIRD-PARTY-NOTICES.md
`-- Cargo.toml

Documentation

• Architecture
• Security Design
• Task Workflow
• Tech Stack
• Onboarding
• Project Management
• DeepSeek API Notes
• Release Process
• Roadmap
• Changelog

Contributing

Contributions are welcome — see CONTRIBUTING.md for the development setup, the CI checks, and the commit conventions that drive the automated releases. Report vulnerabilities privately as described in SECURITY.md.

License

RaptorCode is dual-licensed under the MIT license and the Apache License 2.0, at your option. Copyright © 2026 Bob Vasic.

Releases v0.3.0 through v0.6.5 were published under a proprietary license; from this repository onward all code is MIT OR Apache-2.0. Third-party components keep their own permissive licenses; see THIRD-PARTY-NOTICES.md.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in RaptorCode by you shall be dual licensed as above, without any additional terms or conditions.