Project Status
Caliban v0.15.0 is a pre-1.0 release. The binary (caliban) is daily-usable from main; the
core agent loop, TUI, headless mode, sessions, permissions, tools, MCP, sub-agents, memory,
sandbox, telemetry, and the driveable server surfaces are all shipped. A number of parity
gaps with Claude Code remain. Per-release detail is in the Changelog.
What is shipped
The table below summarizes the major shipped areas. All items marked ✅ are available on
main today.
| Area | Status |
|---|---|
Interactive TUI (ratatui, transcript, status bar, slash menu, @file picker) | ✅ |
Headless --print / stream-json I/O protocol | ✅ |
Persistent named sessions (--session, --resume, --continue) | ✅ |
Permissions: rule grammar, six modes, caliban perms CLI, audit log | ✅ |
| Built-in tools (Read, Write, Edit, MultiEdit, Glob, Grep, Bash, BashOutput, KillShell, WebFetch, WebSearch, NotebookEdit, TodoWrite, AgentTool, Enter/ExitPlanMode, Read/WriteMemoryTopic, ToolSearch, Skill) | ✅ |
Image / vision input (engine built in caliban-images; not yet wired into the binary) | 🟡 |
| MCP client (stdio + HTTP, OAuth, elicitation, per-server permissions) | ✅ |
Sub-agents (in-process, background fleet via caliband, worktree isolation) | ✅ |
Memory tiers: CLAUDE.md ancestry, @-imports, auto-memory | ✅ |
Settings layering (Managed > User > Project > Local > --settings, deep-merge) plus a CALIBAN_* environment layer above them | ✅ |
| Model router v2 (declarative routes, fallback chains, circuit breakers, capability filters) | ✅ |
| Providers in the binary: Anthropic, OpenAI, Google Gemini (+ local OpenAI-compatible servers); Bedrock/Vertex/Azure adapters are library-only | ✅ |
Checkpoints + /rewind, including forking a new session from a checkpoint | ✅ |
Driveable server surfaces: caliban mcp serve, caliban acp serve, caliban http serve; a fleet worker can also serve ACP on its own TLS+token listener | ✅ |
caliband graceful drain + resume of daemon agents from a persisted session (a supervisor control-protocol command for orchestrators; there is no caliban agents subcommand for it) | ✅ |
| Agent-loop policy surface: turn / wall-clock / cost budgets, verification guidance, named profiles with adaptive local-vs-cloud defaults | ✅ |
Per-session permission posture (supervised / unattended) on a fleet spawn, authorized upstream and audited | ✅ |
Prebuilt macOS arm64 (aarch64-apple-darwin) binary on tagged releases | ✅ |
| Plugins, hooks, skills | ✅ |
| OS sandbox (Seatbelt on macOS, bubblewrap on Linux) | ✅ |
| OpenTelemetry + per-request cost tracking | ✅ |
What is partial or backlog
Some rows in the parity matrix are 🟡 (partial / experimental):
| Area | State |
|---|---|
| Vim editing mode in TUI | 🔴 not yet |
Cost surfacing in TUI (/cost display) | 🟡 backlog |
| GitHub Actions workflow / devcontainer feature | 🔴 planned |
| IDE extensions, GitHub App, remote control, mobile (theme N) | 🔴 parked until CLI parity |
IDE extensions, the GitHub App, claude.ai/code, iOS, Slack integration, Remote Control, Channels, Routines, Deep links, and Teleport are all tracked in the parity matrix under theme N. They are explicitly parked until the terminal/CLI feature set reaches full parity with Claude Code. Do not rely on any of these surfaces being available in the near term.
For the full up-to-date breakdown, see Parity vs Claude Code. If you hit something unexpected, see Troubleshooting.
License
Caliban is licensed under AGPL-3.0-only. See Philosophy for the rationale.