A coding agent that built itself.

Commit f3fec8e is sixteen lines adding Session::len and a unit test. Deliberately small — the criterion was about provenance, not size. After that the scope grew:

• The entire TUI layer: semantic colour theme, ANSI-aware width helpers, the status line, structured tool blocks.
• Token-usage capture through the SSE pipeline.
• The web_fetch tool — including applying its own code-review feedback.
• grep respecting .gitignore.
• The MCP client: a Tool::name refactor, config sections, the stdio client, startup wiring.
• Its own README and install.sh.

commit f3fec8e10581c7c3c6288753c00667f52fff5d77
Author: jinpeng <jinpeng.dong@gmail.com>
Date:   Wed Oct 14 11:35:03 2026 +0800

    dogfood

 crates/crab-core/src/session.rs | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

The diff is crab's. The review and the commit are the human's. That distinction is the whole claim.

Observability

One log, two readers.

The session file is the single source of truth. Model calls, tool calls, token counts and content digests are appended as typed observation records — never sent to the model, never rewritten.

Two subcommands read it back: crab trace for the shape of a run, crab metrics for the money. Compaction is recorded as an event and the model's view is rebuilt by replaying it, so --continue sees exactly what the model saw.

• Four-way token usage: prompt, completion, cache read, cache write.
• Costs come from a versioned price table with peak/valley windows.
• Unpriced models show −, never a fake $0.

How to read both reports →

Providers

One wire protocol, profiles, per-field overrides.

crab talks the OpenAI chat-completions API. A provider is a profile in ~/.crab/config.toml. With no config at all it uses api.openai.com and gpt-4o-mini, so export CRAB_API_KEY=… is a working setup.

• Environment beats the profile, field by field: CRAB_BASE_URL, CRAB_MODEL, CRAB_CONTEXT_WINDOW.
• --provider mock is a built-in stub — the fastest way to debug the loop without spending tokens.
• /model <name> switches mid-session; the transcript is provider-agnostic.
• Local endpoints may omit the key; anything else without one is an error.

# ~/.crab/config.toml
default_provider = "deepseek"

[providers.deepseek]
base_url       = "https://api.deepseek.com/v1"
model          = "deepseek-chat"
api_key        = "sk-…"        # optional
context_window = 128000

[providers.local]
base_url       = "http://localhost:8080/v1"
model          = "qwen"

auto_compact_threshold = 0.8   # 0 disables

Then: crab --provider local, or /model deepseek inside a session.

MCP & skills

Extend it without an ABI.

There is no plugin format to learn. External tools arrive over the Model Context Protocol; capability packages arrive as Agent Skills.

An MCP server is a [mcp.<name>] block. Its tools are registered as <server>__<tool>, and several tools from one server share a single child process. Startup is best-effort: a server that fails to spawn warns and is skipped, never blocking crab from starting.

Skills are directories with a SKILL.md, discovered in .agents/skills/ and ~/.agents/skills/. Only name, description and path go into the prompt — the body stays on disk until the task matches.

MCP configuration →

# MCP: extra tools, no rebuild
[mcp.brave]
command = "npx"
args    = ["-y", "@modelcontextprotocol/server-brave-search"]
enabled = true
env     = { BRAVE_API_KEY = "…" }

# behind a corporate proxy, a Node server also needs:
#   NODE_OPTIONS = "--use-env-proxy"

# Skills: advertised by name, loaded on demand
.agents/skills/my-skill/SKILL.md
~/.agents/skills/my-skill/SKILL.md

The model sees brave__brave_web_search. Use /skill:<name> to inject one explicitly.