Skip to main content

MCP server that captures and retrieves architectural decisions and context for Claude Code sessions

Project description

Contexer

PyPI version License: MIT Python 3.12+ Discord

contexer.ai · Discord · Docs

Contexer is a lightweight MCP server for Claude Code that automatically captures decisions made during coding sessions and surfaces them at the start of every future session — so Claude never starts blind.

The problem

Every Claude Code session starts with no memory of the previous one. CLAUDE.md files require manual maintenance and go stale. When Claude works autonomously, the reasoning behind decisions disappears when the session ends. Teams end up re-explaining the same constraints, conventions, and architecture choices every time.

Contexer solves this by capturing decisions as they happen — silently, automatically, in the background — and replaying them as project rules at session start.


Quick start

Install takes under two minutes. See docs/install.md for full steps, verification, and uninstall.

Step 1 — install the package:

uv tool install contexer

Step 2 — wire it into Claude Code:

contexer install

Step 3 — restart Claude Code and open any git repo. On your first message, Claude will offer a quick bootstrap setup — answer yes to capture initial decisions, or skip to start immediately.


How it works

You work normally. Contexer runs silently in the background.

  • Session start — all your constraints and conventions are injected as project rules before you type anything
  • As you work — Claude captures significant decisions automatically; you never have to do it manually
  • "Why" questions — if you ask about rationale or past decisions, Contexer auto-fetches the relevant ones
  • Context window limit — decisions are saved before compaction and restored after, so nothing is lost

You never call anything directly. If Claude misses something, say "store that decision" and it will be captured immediately.


Decision types

Every stored decision has a type that controls when it is surfaced.

Type What it captures Surfaced at session start?
constraint Rules that must always apply — "never commit untested code" Yes — always
convention Team or project standards — "use uv not pip", "conventional commits" Yes — always
architecture Structural decisions — "chose REST over GraphQL for this service" No — fetched when relevant
pattern Recurring implementation approaches — "always validate at the boundary" No — fetched when relevant

Constraints and conventions apply to every task, so they load immediately. Architecture and pattern decisions are fetched on demand when the task requires them.


Managing decisions

All operations use natural language.

Store a decision

"store that as a constraint"
"save this as a convention: always use uv not pip"
"remember this architecture decision"

Note: Your first prompt each session is captured as the current task description, not as a constraint. If you open a session with an instruction like "always update docs before committing", say "store that as a constraint" to make sure it is saved correctly.

Store a global decision

Some conventions apply across every repo — commit style, branch naming, code review process. Store these once and they appear in every session automatically:

"store that globally as a convention"
"save this as a global constraint: always use conventional commits"

Only constraint and convention subtypes can be stored globally. Architecture and pattern decisions are always repo-specific.

To see what is stored globally:

"show all global conventions"
"show global constraints"

Query decisions

"show me all constraints"
"what decisions did we make about postgres?"
"show everything stored for this repo"

Update a decision

"update the uv decision — we switched back to pip"
"correct the constraint about commit format"

The old entry stays alongside the new one. If the revision is too similar to the original, rephrase it to include what changed.

Remove a decision

"delete the postgres decision"
"remove all outdated constraints"

You can also edit the store file directly — it is plain JSON at ~/.contexer/.


Token cost

Contexer is front-loaded by design. Cost is paid once at session start; everything after is near-zero.

What gets injected at session start:

Only constraint and convention decisions are pre-loaded. architecture and pattern decisions cost 0 tokens at session start — they are fetched on demand when relevant.

Decisions stored Pre-loaded rules Tokens at session start
10 5 ~125
20 10 ~250
50 25 ~625

Cost is flat at ~25 tokens per pre-loaded rule, regardless of store size.

What that compares to:

~250 tokens (20 decisions) is 0.3–2.5% of a typical Claude Code session. Without Contexer, Claude spends 200–500 tokens re-establishing each rule through back-and-forth — every session. A single corrected mistake costs 500–2,000 tokens. The tool pays for itself after one avoided repetition.

Rationale injection — only fires on "why" questions:

Event Tokens added Latency
Hit — relevant decision found ~45–80 ~0.06ms store lookup
Miss — no rationale keyword 0 0.000ms

Latency: Contexer's store operations take 0.03–0.27ms. The perceived delay is the MCP hook call round-trip (~50–200ms), which runs before Claude responds — not added on top of it. Sessions with 500 stored decisions have the same sub-millisecond retrieval as sessions with 10.

These numbers were measured on Claude Sonnet 4.6. Haiku will process injected tokens faster; Opus will take slightly longer but produce higher-quality responses with the same context. The store lookup times are model-independent — they are pure Python/disk operations.


Troubleshooting

Claude isn't storing decisions automatically. Say "store that decision" and Claude will capture it immediately.

A decision was stored but isn't appearing. Constraints and conventions load at session start. If you added one mid-session, it will appear from the next session onward.

A decision is outdated or wrong. Say "delete the X decision" or edit the store file directly.

A new decision wasn't saved — it looks like a duplicate. If the content is too close to an existing decision, it is silently skipped. Rephrase it to include what specifically changed.

No context appeared at session start. This happens the first time you open a repo. Claude will offer bootstrap setup — answer yes to capture your initial decisions. Once done, all future sessions will have context.


Contributing

Bug reports, fixes, and documentation improvements are welcome. See CONTRIBUTING.md for development setup, code style, commit format, and the PR process.

Questions or ideas? Join the community on Discord.


License

MIT — see LICENSE for full terms.

The Contexer name and logo are trademarks of Contexer.ai. The MIT license does not grant rights to use the Contexer name, logo, or brand in any way that implies official affiliation.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

contexer-0.2.0.tar.gz (111.9 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

contexer-0.2.0-py3-none-any.whl (21.5 kB view details)

Uploaded Python 3

File details

Details for the file contexer-0.2.0.tar.gz.

File metadata

  • Download URL: contexer-0.2.0.tar.gz
  • Upload date:
  • Size: 111.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for contexer-0.2.0.tar.gz
Algorithm Hash digest
SHA256 f45b63de7c0126de0872dfa4ff92f1618dd86df228830e6f3468b4775662d6ba
MD5 56ede9b801f769086eb5897bbff25b34
BLAKE2b-256 d2771b2c9b51480ce587c2889b120014171867f88155f6356c4afa9f730ba432

See more details on using hashes here.

Provenance

The following attestation bundles were made for contexer-0.2.0.tar.gz:

Publisher: publish.yml on bhargavamin/contexer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file contexer-0.2.0-py3-none-any.whl.

File metadata

  • Download URL: contexer-0.2.0-py3-none-any.whl
  • Upload date:
  • Size: 21.5 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for contexer-0.2.0-py3-none-any.whl
Algorithm Hash digest
SHA256 5f6215cd52978348d289ef96c80b64bf0b737791aa19147ecd641f6f93178e67
MD5 c16dfa34c0c195f6ff08cd2f487ad308
BLAKE2b-256 2f2f62a07c980ff1e3606b43d6dd2501d631bc989f28d6cddeaba9ac577619a2

See more details on using hashes here.

Provenance

The following attestation bundles were made for contexer-0.2.0-py3-none-any.whl:

Publisher: publish.yml on bhargavamin/contexer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page