devkb

mcp
Security Audit
Fail
Health Warn
  • License — License: Apache-2.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Fail
  • exec() — Shell command execution in bench/exercises/adr007/rest_plain.py
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Knowledge graph of your repository for coding agents: precomputed context packs and an MCP server. 55% cheaper with better answers in a 1,138-run benchmark.

README.md

devkb

DOI

A knowledge graph of your repository for software development agents. Agents answer "what calls this, what breaks, which tests"
in one step instead of searching file by file: 55% cheaper, with better answers, in our benchmark.

Median cost per task: plain search $0.070 (quality 0.931); devkb 12 tools $0.059, 18% cheaper (0.955); devkb 7 tools $0.040, 38% cheaper (0.963); devkb pack $0.025, 55% cheaper (0.981)

1. What devkb is

  • A graph built from your source and git history: functions, calls, tests, routes, errors, configuration,
    commits, pull requests, tickets (ontology).
  • Packs: one command answers a whole question (who calls this, what breaks, which tests to run), each item with
    the source line that proves it, and a statement of what the graph could not see.
  • An MCP server giving agents the same answers as tools, for Claude Code, Codex and any MCP client.

Everything runs locally: devkb reads files and runs git, never your code (unless devkb.json opts in), and the graph never leaves your machine.

Languages

Language Status What devkb understands
Python ✅ Supported functions, classes, calls (through imports, re-exports, typed objects, base classes), tests, FastAPI/Flask routes, raised and logged errors, environment variables; every pack
TypeScript, TSX 🟡 Partial functions, components, classes, and the HTTP calls they make to backend routes; not calls between them
JavaScript (.js, .mjs) 🟡 Partial as TypeScript
Odoo XML 🟡 Partial templates and menus
Bash / shell ⬜ Not yet files and history only
C, C++, Java, Rust, Go, C#, others ⬜ Not yet files and history only

"Files and history" is the language-independent layer every repository gets: files, commits, pull requests and
tickets.

2. How much it saves

Latest benchmark: the 7 open-source repositories in bench/repos.yaml, 57 tasks, 1,138 agent
runs, the same model (Sonnet) in every arm, answers scored against truth computed from the source. Cost, turns and
tokens are the median run; saving is per task against plain search, averaged over tasks; quality is the mean score:

Plain search devkb, 12 tools devkb, 7 tools devkb pack
Cost per task $0.070 $0.059 $0.040 $0.025
Saving — 18% 38% 55%
Quality (0–1) 0.931 0.955 (+0.024) 0.963 (+0.032) 0.981 (+0.050)
Turns 4 6 5 3
Context processed (tokens) 126k 169k 147k 75k
of which cache reads 123k 163k 141k 74k
of which new input and cache writes 8.3k 4.1k 2.0k 1.3k
Output tokens 1.4k 1.8k 1.3k 0.9k

Saving and quality gain by repository (95% intervals in the results):

Repository Tasks 7 tools: saving / quality pack: saving / quality
lnbits 10 53% / +0.032 66% / +0.049
Netflix Dispatch 8 45% / +0.040 61% / +0.081
Prefect 9 41% / −0.017 59% / +0.047
lattice-scan 12 36% / +0.052 51% / +0.058
Mealie 6 18% / +0.095 42% / +0.062
devkb 4 1% / +0.004 40% / +0.004
click 8 6% / +0.016 15% / +0.023

The saving is in call chains (66–77% on tasks that trace an error or a change to the route handlers that reach it);
on one-hop questions, as in click, plain search is already good. An earlier pre-registered study (30 held-out tasks,
1,020 runs, two model sizes) found packs 41% cheaper and the lean tools 35% (study/; make verify
recomputes it). Details and corrections: bench/results/2026-11-15-registry;
add a repository: docs/BENCHMARK.md.

3. Set up

uv tool install "devkb[mcp]"      # or: pipx install "devkb[mcp]"
cd your-repo
devkb init --hooks

Install it as a tool, as above, so the devkb command is on your PATH: agents start the server from there, and a
devkb installed only inside a virtual environment never starts (devkb init warns if that is the case).

This builds the graph into .devkb/, adds a short devkb section to AGENTS.md and CLAUDE.md, registers the MCP
server in .mcp.json, and rebuilds the graph after each commit. Your own content in those files is kept;
devkb init --remove undoes it.

Ask it directly:

devkb pack --project . refactor my_function     # who calls it, which tests
devkb pack --project . error "message text"     # which route handlers reach it
devkb pack --project . file path/to/file.py     # before editing a file

Your agent

Claude Code: nothing more to do. devkb init registered the server in .mcp.json and pointed CLAUDE.md at the
instructions. Start claude in the repository, approve the devkb server when asked, and check it with /mcp.
Or skip pip and devkb init and install the plugin, which brings the server, its instructions and a hook that builds
the graph in the background (needs uv):

/plugin install devkb --marketplace namiq-ai/devkb

Cursor: add .cursor/mcp.json to the repository, then enable devkb under Settings → MCP:

{"mcpServers": {"devkb": {"command": "devkb", "args": ["mcp", "--project", ".", "--toolset", "lean-pack"]}}}

Codex: add to ~/.codex/config.toml:

[mcp_servers.devkb]
command = "devkb"
args = ["mcp", "--project", ".", "--toolset", "lean-pack"]

Codex and Cursor read the instructions in AGENTS.md; Claude Code reads them through CLAUDE.md.

What agents get by default (--toolset lean-pack): a pack tool that returns a whole answer in one call, and the
seven single-question tools. The instructions tell them to ask for a pack when a question follows a chain (a file
before editing, a change, an error) and to search as usual for a one-hop lookup, which is where the benchmark found
the saving and where it did not. Measured with agents asking for the pack themselves (lnbits and Dispatch, 400 runs):
53% cheaper than plain search with the highest quality of any arm, against 42% for the seven tools alone
(bench/results/2026-11-15-lean-pack).

4. Licence and citation

Code: Apache-2.0. The study in study/ (paper, figures, pre-registration, data):
CC BY 4.0. Benchmark results quote third-party code under its own licences
(bench/results/NOTICE). Releases: docs/RELEASING.md.

Hung T. Le. Precomputed Context, Not More Tools: A Pre-Registered Study of Process Graphs for Coding Agents. 2026.

Reviews (0)

No results found