Give an agent context about my repository¶
The problem: an agent starts each task by re-deriving what your repository already knows — which files relate, what was decided and why, what broke before.
What you get: one index every agent is routed to, holding a code graph, a wiki, memory and quality artifacts, versioned alongside the code.
Every command below was executed; the output is from those runs.
1. Install the workspace¶
init scaffolds .metaproject/, enables nine modules, installs git hooks, and
wires your existing AGENTS.md / CLAUDE.md so an agent reading them is routed
to .metaproject/index.md.
It is an idempotent reconciler: re-running it refreshes managed files and
never clobbers your edits or anything under .metaproject/data/.
2. Build the code graph¶
That run is expressjs/express, cloned fresh. On this repository the same
command reports 649 nodes and 1,873 edges.
3. Ask the question agents waste the most tokens on¶
$ keryx gdgraph affected lib/express.js
# Affected context for lib/express.js
## Dependencies
- lib/application.js
- lib/request.js
- lib/response.js
## Dependents
- examples/route-map/index.js
- examples/route-middleware/index.js
- index.js
What breaks if I change this is the answer an agent otherwise reconstructs from a dozen file reads, and gets wrong on exactly the files where it matters.
Two more that pay for themselves:
4. Point agents at it¶
init already did this for the entrypoints it found. For a per-turn
orientation block, install the optional hook:
keryx agents bootstrap install --runtime claude does the same for the
global agent entrypoint, so a runtime discovers the workspace in any
project that has one.
Verify you got what this guide promises¶
$ keryx status
Metaproject: ready
Root: .metaproject
Modules:
gdgraph: enabled
gdctx: enabled
gdwiki: enabled
gdskills: enabled
memory: enabled
tasks: enabled
health: enabled
testing: enabled
security: enabled
If gdgraph reports 0 nodes, 0 edges on a repository that has code, the
graph has not been built — run keryx gdgraph build before reading anything
into the number.
Where to go next¶
- Architecture — what the workspace contract actually is, and who may write what.
- Module reference — one section per module.
- Shared Agent Context — bind a task to a workspace, read a bounded FWK overview, propose reviewed knowledge back.
- Run keryx in CI — keep these artifacts current without anyone remembering to.