Keen Code

Why it's ranked here

The documentation set is unusually deep for a young command-line project: separate write-ups cover provider routing, the compaction budget math, MCP transport handling, memory size limits, and the permission model, each specified as concrete states and thresholds rather than vague prose. The project is also candid about its own gaps, stating that delegated helper agents get their tool calls approved automatically, including shell commands the tool itself flags as dangerous, and that parallel delegated work has no built-in ownership or budget enforcement. That combination of a working mechanism and an honest account of its own limits is what separates it from the many thin agent wrappers submitted to this catalogue.

What's good

The edit tool will not blindly apply a change: it fingerprints each line when it is read, then rejects the edit outright if the file drifted underneath it, closing off the silent-overwrite failure mode common to plain find-and-replace editors. Provider authentication spans plain API keys, browser-based OAuth with PKCE, and the standard AWS credential chain, so switching model backends does not mean rebuilding authentication plumbing each time. Anything the agent writes to its own persistent memory files is scanned for secret-shaped text before the write goes through, and even a pre-approved tool list cannot override the underlying filesystem denials on system directories or ignored files.

Tradeoffs

MCP support only exposes tools to the model, not resources or prompts, and there is no live reload: adding or removing a server means restarting the whole agent, and per-server timeouts are not configurable. A batch of delegated work runs every task inside it concurrently, with no dependency ordering, no file-ownership checks, and no cost cap, so two parallel workers can overwrite the same files, and any helper agent granted shell access has its commands, including ones flagged as dangerous, approved automatically with no prompt. The binary also builds a usage-telemetry reporter into every run, interactive or headless.

How to use it well

It suits an engineer who already works from the terminal and wants one agent that can point at several model backends without rewriting scripts, especially teams standardized on Bedrock or an OpenAI-compatible gateway who want the credential handling done for them. Treat helper agents as a scoped feature, not a free concurrency dial: give each one the narrowest permission set its task needs, reserve shell access for profiles that truly need to run commands, and split dependent steps into separate delegation rounds by hand, since the runtime will not order them automatically. Check the memory and MCP config locations before assuming either persists per-project.

Observed

Language

Go

Interfaces

Interactive terminal REPL plus a separate non-interactive command mode with structured JSON output

AI provider backends

Ten configurable providers, switchable per session without restarting

Built-in tools

Ten, covering file read/write/edit, search, shell execution, web fetch, user prompts, task delegation, and MCP calls

MCP transports

Streamable HTTP and stdio, with none, API key, and browser-based OAuth authentication

Telemetry

Constructs a usage-telemetry reporter on every run, interactive or headless

Memory storage

Plain markdown files with fixed size caps, and writes are rejected if they match secret-pattern text

Subagent delegation

One to ten tasks per delegation call, all approved automatically once a profile is granted permissions