indusagi coding agent

A terminal-first AI coding agent. Talk to it in your terminal, point it at a repo, and it reads, writes, runs, and reasons about your code. It runs entirely on the indusagi framework — one dependency for the LLM API, agent loop, terminal UI, MCP client, and tracing.

It is built from scratch: every module was written from a behavioral spec and the framework's public API, never by copying third-party application source. See CREDITS.md and NOTICE.

indus                      # start an interactive session in the current repo
indus "fix the failing test in src/auth"
indus -p "summarise what changed on this branch"   # headless, prints result, exits
indus --json               # line protocol for a driving parent process

(The bin is installed as both indus and indusagi.)

Install & build

npm install
npm run build          # esbuild bundle -> dist/entry.js  +  tsc d.ts -> dist/types
npm test               # vitest
npm run typecheck      # tsc --noEmit
npm run lineage-scan   # fails if any external-source marker appears in non-test source

Requires Node >= 20. Authenticate with an API key — set ANTHROPIC_API_KEY (or another supported provider key) in the environment, or run indus signin. API-key auth only; there is no OAuth path.

Run modes

| Invocation | Mode | What it does | |------------|------|--------------| | indus | interactive | Ink/React REPL — streaming replies, a live tool deck, slash-commands | | indus -p "…" / --print | print | One request, prints the final result, exits | | indus --json / --rpc | line protocol | Newline-delimited JSON for a parent process to drive | | indus signin / signout | credentials | Store / clear the API key in the local auth vault |

indus --help prints the full flag grammar (generated from one flag table); indus --version prints the version.

Architecture

The binary is src/entry.ts. Boot resolves the invocation, prepares the workspace, and selects a runner. Each subsystem is a self-contained module and is also exported as a namespace from src/index.ts for embedding and testing.

| Module | Responsibility | |--------|----------------| | boot | argv → invocation, workspace prep, runner selection, upgrades | | workspace | per-project + global dirs, settings, session paths | | launch | flag grammar (readInvocation / renderUsage), @file attachments, credentials, model catalog & pickers | | conductor | the conversation loop — turns, tool invocation, fault/retry, branch/resume, compaction | | window-budget | token accounting + context-window compaction policy | | capability-deck | the tool registry (read/write/bash/glob/grep/fetch/search …) | | runtime-bridge | framework wiring — models, MCP, providers | | addons | loadable extensions: skills & subagents | | console | the interactive Ink/React surface | | channels | print + line-protocol headless surfaces | | briefing | system prompt / context assembly | | transcript-export | session persistence + HTML/markdown export | | insight | tracing (Trail / Probe / Signal), sinks, replay, secret redaction | | kit | leaf helpers (image sniff, managed-binary fetch, shell quoting) |

Clean-room posture

npm run lineage-scan runs scripts/lineage-scan.mjs, which walks src/ and exits non-zero if any non-test file carries an external-source marker. Test files are excluded by design — several of them assert the absence of those markers. The scan is part of every release check.

License

See NOTICE and CREDITS.md.