pypeeker-cli

Agentic DX for Python. DX = developer experience: the tools, feedback, and ergonomics that make working in a codebase feel good. Agentic DX is the same idea, but for the AI agent doing the work instead of a human. pypeeker gives AI coding agents AST-level views of any Python codebase — API surfaces, side-effect maps, import cycles, symbol locations — surfaced through MCP in token-efficient structured answers. So your agent can navigate, reason about, and refactor Python without reading whole files.

Read-only by design. pypeeker never edits, renames, or rewrites your code. No apply, no fix, no auto-refactor surface. Every tool answers a question about the codebase; the agent acts on the answer. This is a hard boundary, not a roadmap gap.

---

Cartograph Showcase

pypeeker-cli is a premier showcase for Cartograph, a platform for reusable engineering.

Every core feature in pypeeker-cli, from AST parsing to graph cycle detection, is implemented as a standalone, validated Cartograph widget. This architecture ensures that pypeeker-cli is not just a tool, but a modular assembly of hardened building blocks that can be easily extended or repurposed.

---

What the agent gets

Three MCP tools, designed to keep the agent's context budget tight:

audit(directory, kind=...) — project-wide checks

• kind="cycles" — import cycles + a cycle_hubs ranking of the most-tangled files
• kind="missing-imports" — hallucinated or broken internal import paths
• kind="interfaces" — missing docstrings and type annotations

peek(path, mode=..., symbol=...) — file or symbol inspection

• mode="skeleton" — file/package API surface with line ranges for targeted reads
• mode="locate" / "ancestry" — find a symbol's definition or class parents
• mode="impact" — bidirectional analysis of a function: outbound (what it depends on) and inbound (who calls it). Use direction="in"|"out" to scope, depth=N + root for transitive outbound

cli(command, args) — escape hatch

Runs pypeeker <command> <args> and returns stdout. For flags the typed tools don't expose (--include-deps, custom --ignore, --format json on text-default tools).

---

Real examples

All outputs below are real, captured against psf/requests (and a small synthetic project for the cycles example). Each example shows what the agent invokes over MCP and what comes back.

audit(kind="cycles") — find import cycles

Detects runtime and TYPE_CHECKING import cycles, ranks the most-tangled files via cycle hubs.

Agent invokes: audit(directory=".", kind="cycles")

# circular imports
3 cycles found

[hubs] files in 2+ cycles:
  hub.py  3

[1] runtime cycle:
  hub.py:1  → a
  a.py:1  → hub

[2] runtime cycle:
  hub.py:2  → b
  b.py:1  → hub

[3] runtime cycle:
  c.py:1  → hub
  hub.py:3  → c

The cycle_hubs ranking tells the agent which file to refactor first — hub.py shows up in all 3 cycles, so untangling it dissolves them all. (psf/requests itself reports (none) — clean codebase.)

audit(kind="missing-imports") — catch hallucinated imports

Walks every import statement; reports the ones that don't resolve to a real file.

Agent invokes: audit(directory="requests/", kind="missing-imports")

# missing imports
utils.py:78     winreg
help.py:19      chardet
cookies.py:20   dummy_threading
compat.py:60    simplejson
compat.py:67    simplejson
__init__.py:53  chardet

In requests these are intentional optional/conditional imports — but the same scan over an agent-edited codebase catches imports the agent invented and never actually installed.

audit(kind="interfaces") — find docstring and type-hint gaps

Reports public symbols missing docstrings or annotations.

Agent invokes: audit(directory="requests/", kind="interfaces")

total gaps: 233

utils.py:126  dict_to_sequence
    missing_type_hint: d
    missing_return_type
utils.py:135  super_len
    missing_docstring
    missing_type_hint: o
    missing_return_type
utils.py:206  get_netrc_auth
    missing_type_hint: url
    missing_type_hint: raise_errors
    missing_return_type

One pass over the project; agent gets a punch list of every gap with file, line, and exact missing piece.

peek(mode="skeleton") — file or package API map

Strips function bodies; returns imports, signatures, docstrings, and the line range each symbol occupies.

Agent invokes: peek(path="requests/api.py", mode="skeleton")

from . import sessions

def request(method, url, **kwargs):  # L14-59
    """
    Constructs and sends a :class:`Request <Request>`.
    ...
    """
    ...

def get(url, params=None, **kwargs):  # L62-73
    ...

def post(url, data=None, json=None, **kwargs):  # L103-115
    ...

The agent now knows exactly which lines (L14-59) to Read if it needs the body of request(). No full-file scan to find one method.

peek(mode="locate") — pinpoint a definition

AST-aware lookup: returns scope ranges, distinguishes definitions from substring matches.

Agent invokes: peek(path="requests/", mode="locate", symbol="Session")

# locate: Session
sessions.py:356-818  class Session(SessionRedirectMixin)

Pass mode="ancestry" to also resolve parent classes:

# locate: Session
sessions.py:356-818  class Session(SessionRedirectMixin)
  ↳ sessions.py:107-353  class SessionRedirectMixin

peek(mode="impact") — bidirectional analysis of a function

Default: both directions in one call. Outbound = what this function depends on (calls, writes, globals). Inbound = who calls this function across the project.

Agent invokes: peek(path="requests/sessions.py", mode="impact", symbol="Session.send", root="requests/")

# impact: Session.send

## outbound (what this function reaches into)

  external calls:   16
    ValueError
    adapter.send
    dispatch_hook
    extract_cookies_to_jar
    history.insert
    history.pop
    isinstance
    kwargs.get
    kwargs.pop
    kwargs.setdefault
    next
    preferred_clock
    resolve_proxies
    self.get_adapter
    self.resolve_redirects
    timedelta
  external writes:  0
  globals modified: 0

## inbound (who calls this) — 4 match(es)

  requests/sessions.py:265
  requests/sessions.py:591
  requests/sessions.py:705
  requests/auth.py:276

Two refactor questions answered in one call:

• Outbound: Session.send depends on 16 external calls, mutates zero external state, touches zero globals. Pure dispatcher — agent can trust the body without reading it.
• Inbound: 4 callers across 2 files. If the agent changes Session.send's contract, those are the 4 lines that might break.

Scope flags:

• direction="out" — only outbound (when you only care about what the function depends on).
• direction="in" — only inbound (when you only care about callers).
• depth=N (max 5) — transitive outbound. Walks the call graph N levels and aggregates writes, globals, and unresolved calls into one surface. Static resolution only: self.X, Class.X, and statically-imported names are followed; dynamic dispatch is reported as unresolved, not silently followed. Cycle-safe.

cli(command, args) — escape hatch

Runs an arbitrary pypeeker <command> <args> and returns its stdout. Use when you need a flag the typed tools don't expose.

Agent invokes: cli(command="impact", args=["Session.send", "requests/sessions.py", "--depth", "2", "--root", "requests/"])

Returns raw pypeeker stdout in a {stdout, stderr, exit_code} envelope. Prefer the typed tools when you can — they're easier for agents to reason about — and reach for cli when you need a one-off flag.

---

The token math

Same task — find where requests handles redirect auth — measured against actual agent tool output, including the Read tool's line-number prefixes. Tokens counted with OpenAI's o200k_base tokenizer (used by GPT-4o, o1, o3, and GPT-5).

Workflow	Tokens
Read sessions.py (full file)	9,063
skeleton sessions.py + 2 targeted Read calls (L128-158, L282-300)	3,218

The agent ends up reading only the two methods that actually matter, instead of 833 lines hoping the relevant code is in there — and the skeleton's line ranges (# L128-158) made the targeting possible.

For project-scale orientation, mapping the full requests package (18 files):

Operation	Tokens
Read every .py file in requests/	59,901
skeleton requests/	19,508

The entire library API surface — every public class, signature, docstring, and line range — in one call.

---

Install

Two steps: install the package, then register it with your agent. The package ships an MCP server (pypeeker mcp); your agent connects to that server and gets the tools described above.

pip install pypeeker-cli

After install, register with whichever agent you use:

Gemini CLI

Install as a native extension:

gemini extensions install https://github.com/benteigland11/pypeeker

Claude Code

Add as a persistent MCP server:

claude mcp add pypeeker-cli -- pypeeker mcp

Codex

Add as a global MCP server:

codex mcp add pypeeker-cli -- pypeeker mcp

Cursor and Roo Code

Pypeeker is pre-configured for automatic detection via .cursor/mcp.json.

Continue

The server is pre-configured via .continue/mcpServers/pypeeker.json.

Windsurf

The server is pre-configured via .windsurf/mcp_config.json.

Claude Desktop and Aider

Copy the JSON from claude_desktop_snippet.json into your global claude_desktop_config.json.

Project config

Add a [tool.pypeeker] section to your pyproject.toml to extend the default ignore list project-wide:

[tool.pypeeker]
ignore = ["legacy", "vendor"]

These directories are added to the defaults (venv, __pycache__, dist, node_modules, .mypy_cache, .tox, build). Active skip list is returned in every audit response under meta.skip_list.

Update notifications

When the MCP server starts, it checks PyPI once per session (cached 24h, 2-second timeout, fail-quiet) for a newer release. If your installed version is stale, agents see a one-line pip install -U pypeeker-cli notice prepended to the server instructions; up-to-date sessions see nothing. Set PYPEEKER_CLI_NO_VERSION_CHECK=1 to disable the check entirely (e.g. on airgapped networks).

---

Reusable Core

Built using the following Cartograph Widgets (found in cg/):

• cg-infra-agent-cli-python: Machine-first declarative CLI framework.
• data-ast-import-parser-python: Root-aware static import analysis.
• data-ast-skeleton-parser-python: API signature extraction.
• data-ast-symbol-locator-python: Surgical symbol pinpointing and ancestry.
• data-ast-interface-validator-python: API gap detection.
• data-ast-impact-analyzer-python: Side-effect and dependency mapping.
• infra-mcp-server-readiness-python: FastMCP version stamping + CLI escape-hatch primitives.
• infra-mcp-manifest-generator-python: Automated distribution scaffolding.
• universal-agent-response-python: Standardized JSON schema.
• universal-list-paginator-python: Result set pagination.

---

License

Licensed under the Apache License, Version 2.0. See the LICENSE file for details.