What does your issue tracker cost your coding agent? I measured it.

Prateek Saxena — Mon, 09 Nov 2026 13:27:40 +0000

I run Claude Code and Codex most of the day, and most of that work is tied to issues: read the ticket, do the work, post progress, hand it off. At some point I wondered what the tracker part of that loop costs, separately from the coding itself.

Every time an agent reads an issue, posts a note or hands work off, the integration it uses costs tokens: instructions it has to load, tool definitions, and whatever the tracker sends back. Nobody seemed to measure it, so I did.

The setup

One job, done the same way through each integration:

Find the open issue about EU refunds, read it with its full comment history (8 comments), post a given progress note, and hand the issue back for the next session.

Agent: Claude Code (claude -p), Sonnet 5.5
Runs: 3 per integration, medians reported
Each tool used its official setup: tk init for Tiimi, bd setup claude for Beads, the gh CLI, and Linear's remote MCP server
Fresh state every run: a new folder and a freshly seeded issue, with only that project's settings loaded
"Tracker tokens" = the input tokens the agent processed, minus Claude Code's own fixed cost (about 16k, measured with no tracker at all)

The results

Integration	Tracker tokens	Cost per job	Time	Turns
Tiimi, local	42k	$0.051	9.7 s	3
Beads	75k	$0.066	11.2 s	4
Tiimi on GitHub Issues	44k	$0.049	18.1 s	3
`gh` CLI	63k	$0.059	15.9 s	4
Tiimi on Linear	45k	$0.053	18.1 s	3
Linear MCP	141k	$0.099	24.0 s	6–8

Head to head:

vs Linear's MCP server: ≈3× fewer tokens, about half the cost, 25% faster
vs Beads: 44% fewer tokens, 23% lower cost, 13% faster
vs the gh CLI: 30% fewer tokens and 17% lower cost (about 2 seconds slower, because Tiimi also moves the issue's status)

Every run of every integration finished the job.

Where the tokens go

Reading the session traces made the differences obvious.

Linear MCP loads its tool definitions, then the agent searches, fetches the issue, fetches the comments, posts, and updates the status: six to eight model turns. Each turn re-sends the growing context, and the responses are JSON with IDs, timestamps and metadata the agent never uses.

Beads is efficient once it's going (search, show plus comments, comment), but it pays about 2k tokens up front every session: bd prime injects its command reference, and its CLAUDE.md loads too. In return the agent knows exactly what to run, and it showed: four turns, every time.

The gh CLI is lean, but the agent has to ask for --json with the right fields, and the issue with comments came back as 4.5 KB of JSON.

Tiimi gives the agent a session-start brief of about 200 tokens: the open issues plus a one-line command card, so it never runs --help. tk show returns the issue and all comments as plain lines (1.7 KB here), and a single chained command claims the issue, posts the note and hands it back. Three turns, every run.

The pattern: fewer turns matter more than smaller responses, because every extra turn re-reads everything before it.

What else Tiimi does

The benchmark measures the tracker layer, but the reason I built Tiimi is the loop around it:

Cost per issue, from logs you already have. Run this once:

tk receipt

It reads the logs Claude Code and Codex write on your machine, prices each reply at list price, and groups the cost by issue from your branch names (eng-142-fix-login → ENG-142). No setup, nothing uploaded. On a subscription it still shows which issues eat your usage limits.

Handoffs between sessions and between agents. When an agent runs low on context it runs:

tk handoff "Rounding fixed in refunds.ts. Left: 2 failing tax.spec cases. Decimal.js everywhere was 3x slower, reverted."

The next session, Claude Code or Codex, starts from that note: what's done, what's left, and the dead ends, so it doesn't retry the same failed approach.

Your existing tracker. It works with Linear, GitHub Issues and Jira (beta), or a local store, so there's no second tracker to keep in sync.

Here's the full loop in 50 seconds, on a sample project:

Run it yourself

The prompts, seed data, raw results and the benchmark script are in the repo under bench/trackers. It's one workflow on one issue, so rerun it on your own setup; I'd like to see the numbers from other agents and bigger histories.

To try Tiimi:

curl -fsSL https://tiimi.dev/install | sh    # or: brew install tiimihq/tap/tiimi
tk receipt