A dashboard needs a human to open it, read it, and retype what mattered. That human is the bottleneck. A nicer chart does not remove a person from the loop — it gives them a nicer screen to transcribe from.
archie mcp is a read-only stdio MCP server. The next session asks
directly, in the same turn it needs the answer, and gets structured data back.
Claude Code
claude mcp add agentworth --scope user -- archie mcp
--scope user matters more here than for a typical MCP server. The point is a
session in any repo asking about any other repo's history — "what was I
doing in the other checkout yesterday", asked from somewhere else entirely. A
project-scoped entry is only live in one repo at a time, which defeats that.
The equivalent hand-written entry goes in ~/.claude.json under the top-level
mcpServers key, or in .mcp.json at project scope:
{
"mcpServers": {
"agentworth": {
"type": "stdio",
"command": "agentworth",
"args": ["mcp"]
}
}
}
Any other MCP client
The server is plain stdio: command agentworth, one argument, mcp. No
environment variables and no auth — it runs as your own user, the same trust
boundary as typing agentworth in a terminal.
Where that server definition goes differs per harness, and this repo does not document the registration step for Codex, OpenCode or Gemini CLI. Use each harness's own MCP configuration docs, with the command and args above.
The 12 tools
All read-only. Nothing scans, nothing writes, nothing touches the original session logs.
A client's tools/list shows 22, not 12. The extra 10 are the pre-0.1.16 tool
names, still registered as deprecated aliases forwarding to the same handlers,
so a client configured before the rename keeps working. They are removed in
v0.1.18. Use the 12 below.
| Tool | What it answers |
|---|---|
session_list |
Filter the index by adapter, model, repo, time or outcome. |
session_show |
One session with its events, tokens and outcome rung. |
repo_blame |
Which session, model and prompt produced a line of code. |
stats_usage |
Tokens and cost rolled up by day, week or month. |
window_show |
Throughput over a moving window. |
agent_list |
Which adapters are detected and what they yield. |
stats_outcomes |
Verified-outcome rate by model, adapter or repo, with an n floor. |
session_handoff |
What a session promised, decided and did not finish. |
session_carry_forward |
What the previous sessions in this repo left for this one. |
session_forgotten |
Decisions compaction dropped, returned verbatim with receipts. |
session_asks |
Where a question's answer already landed. |
repo_suspect |
Commits whose session never proved anything. |
Full schemas: the Reference.
Redaction
Redacted output is the default everywhere event or file content comes back.
include_raw is the only opt-in, and it is per call — never global.
A session's first and last tool call
Open with session_carry_forward, so the session starts knowing what the last few did.
Close with session_handoff, so the next one does. If the session compacted in
between, session_forgotten recovers what its own summaries dropped.
No tool scans. Run archie scan yourself if the index looks stale.
What to run
claude mcp add agentworth --scope user -- archie mcp
archie mcp
archie scan
archie doctor --self-test
archie mcp on its own speaks the protocol on stdin and stdout, which is
how you check the binary is reachable. doctor --self-test runs the real
workflow end to end, including an MCP round trip, and exits non-zero if any step
fails.