Kram
Open-source terminal coding agent
One agent. Every model. Routed live from the TUI.
Kram is a gateway-native agent, not a wrapper: switch provider combos with Ctrl+S, survive failures, steer mid-turn, reattach after terminal crashes, and rewind risky edits with one key.
curl -fsSL https://raw.githubusercontent.com/codexmark/kram-releases/master/install.sh | sh
Kram TUI
smart → chatgpt/browser → anthropic/api fallback armed
Built for real terminal work
Complete feature surface
A gateway-native agent, not a wrapper.
Native multi-provider gateway
The multi-provider gateway is built into the agent. Compose provider combos and switch strategies live in the TUI with Ctrl+S: smart, round-robin, priority.
Provider failure isolation
Automatic fallback uses a circuit breaker per provider, so one failing route is quarantined while the turn keeps moving.
Resume-grade retries
Long streams survive phase watchdogs. Retry resumes from the token where it stopped instead of restarting the answer.
Terminal crash survival
Turns survive a terminal crash and reattach automatically, so a network drop or closed shell does not lose the agent's work.
Steer mid-turn
Type while the agent works. Kram folds new constraints into the next step without throwing away progress.
Safe edits by default
Automatic checkpoints are created before every file change. Ctrl+G rewinds the workspace when a path goes wrong.
ChatGPT without keys
Use your ChatGPT account through browser login, without an API key. Anthropic, Gemini, OpenRouter, and other APIs work beside it.
Local-first capable
Local models are first-class providers, not a side path. Route to them alone or mix them into provider combos.
Install skills from git
Install skills from git repositories: TDD, debugging, review, and domain playbooks become repeatable terminal workflows.
Memory across sessions
Persist project decisions, constraints, and preferences as memory across sessions so future turns start with useful context.
Transparent routing UI
The TUI shows the routing trail in real time plus a directional data-flow indicator, so you can see which model is working and where data is moving.
How it works
From zero to routed agent in three steps.
- 01
Install
Run the single binary installer with curl | sh.
- 02
Connect a provider
Use browser login for ChatGPT, add API providers, or point Kram at local models.
- 03
Work
Ask, steer, switch routing with Ctrl+S, follow the routing trail, and rewind changes with Ctrl+G.
Gateway resilience
Provider failure becomes a route change, not a lost turn.
Kram tracks provider health, phase timeouts, streaming progress, and retry state. When one route stalls, the circuit breaker opens and the gateway resumes through the next viable provider.
Agent turn
request
Gateway
smart routing
smart routing
ChatGPT
Provider fail
Local model
circuit breaker → automatic fallback → resume stream from token
Skills and memory
Install repeatable playbooks. Keep useful context.
TDD
Turn red-green-refactor into an executable terminal workflow.
Debugging
Reproduce, isolate, fix, and verify with evidence at each step.
Review
Run disciplined code review checklists before shipping.
Memory
Persist project decisions and preferences across sessions.
Put a routed coding agent in your terminal.
Open-source. Multi-provider. Reattachable. Rewindable.