Requirements & timing
python3.11+
git2.30+
agentclaude code
osmac · linux · wsl
install~15s
auth~15s
init<30s
first gainturn 1
Data sent
namesyes
pathsyes
edgesyes
sourcenever uploaded
promptsvia proxy
Quickstart

Running in under two minutes.

Four commands take you from install to your first structurally-guided Claude Code session. No configuration files, no infrastructure, no workflow changes — and the same Max plan starts stretching 1.6–2× further (a 37–50% token cut per capped window) from the very first prompt.

Talon installs as a regular Python package, authenticates against your existing Cogmeta account, indexes your repository locally, and starts feeding the right context to Claude on the next turn. No network round-trips during retrieval, no source code leaving your machine, and no changes to how you already use Claude Code.

  1. 1Install · pipx install cogmeta-talon — pulls the CLI and the local indexer.
  2. 2Authenticate · talon auth — opens a browser, signs you in, no API key juggling.
  3. 3Initialise · talon init — configures Claude Code; then talon index parses your repo and builds the graph.
  4. 4Run · open Claude Code as normal. Talon serves context behind the scenes from turn one.
● alpha·Python 3.11+·Claude Code compatible
STEP 01

Install Talon

Install the client wheel via pipx — recommended for CLI tools because it isolates dependencies in its own venv. Requires Python 3.11+. Installs the talon CLI plus the indexer and MCP setup.

~ · install
$ pipx install cogmeta-talon installed package cogmeta-talon 0.8.0 These commands are now available: talon · talon-setup · talon-index ✓ done

No pipx? pip install cogmeta-talon works the same way; pipx just keeps the install isolated.

STEP 02

Authenticate

Create your Cogmeta account and link your local installation. The CLI opens a browser for the OAuth handshake. Source code never crosses the wire — the OAuth scope grants workspace metadata access only, and never your Anthropic API key.

~ · auth
$ talon auth login Opening browser to app.cogmeta.ai/auth/cli … Waiting for confirmation ✓ Authenticated as [email protected] ✓ Machine linked · workspace "default"
STEP 03

Initialise your project

Navigate to your project root and run talon init to configure the MCP server for Claude Code, then talon index to parse your repository and build the structural graph.

your-project · init
~/projects cd your-project && talon init → Scanning repository /Users/ada/projects/your-project → Detected: Python, TypeScript (2 of 4 supported) ✓ Structural graph built ✓ Index cached locally ✓ MCP server configured · ~/.claude.json ● Talon ready.

Timings vary by repo size. A medium repo (~2k files, ~5k commits) typically completes in under 30 seconds.

Optional sanity check · run talon status to confirm the index, the workspace, and the MCP reachability before opening Claude Code:

your-project · status (optional)
~/projects/your-project talon status workspace default graph indexed (5 languages) mcp reachable · api.cogmeta.ai/cogz/mcp ✓ Ready for Claude Code
ESTIMATE

Know your band before you pay

Optional, and free. Run talon estimate — a read-only scan (no LLM calls, ~30s) that characterises how hard your code is to search and returns the saving band it correlates with, before you spend a cent:

your-project · estimate
~/projects/your-project talon estimate → Scanning repository · 1,842 files · read-only · no LLM calls · ~28s → Discovery difficulty: HIGH (large, heavily indirected) ✓ Estimated saving band: 35–50% A screen, not a quote — it correlates with our sealed results; your run may land outside the band.

Small, easily-grepped repos come back low — sometimes below the floor where Talon costs more. We publish that too, in /research.

STEP 04

Start coding

Open Claude Code and work normally. Talon operates transparently — before the agent's first turn, it serves the right structural context, and steers the session clear of redundant re-reads in real time.

your-project · session (illustrative)
~/projects/your-project claude Claude Code · model claude-sonnet-4-6 ● Talon active · serving context you add a circuit breaker to the payment retry path [talon] 7 files served · task-relevant context [agent] reading src/payments/retry.ts … [talon] redundant re-read suppressed ✓ Patch ready · fewer turns vs vanilla baseline

Sample shape only — the corpus-aggregate cost reduction lives at /research; your numbers will depend on session shape.

What happens next

After your first session.

The dashboard shows per-session scoring breakdowns, context trends, and cost comparisons against your vanilla baseline — so you can see which projects benefit most.