Context Mode
Context Mode explained: smaller tool responses, retrievable evidence and real boundaries
Understand what Context Mode moves out of a conversation, how retrieval differs from compression, and why context savings need a task-level check.
What you will learn
- Context Mode selects evidence; it does not expand a model window.
- Temporary processing and persistent indexing solve different needs.
- Check task accuracy alongside response size.
Before you start
- Basic JSON and MCP concepts
- An isolated test client and invented data
Explain the chapter’s actual implementation boundary and verify it with an explicit evidence fixture.
Key takeaways
- Context Mode selects evidence; it does not expand a model window.
- Temporary processing and persistent indexing solve different needs.
- Check task accuracy alongside response size.
Keep the evidence, select what reaches the model
Context Mode is an MCP integration for coding assistants. Its central idea is to process or index bulky tool material outside the conversation and return a smaller answer or selected evidence. It does not enlarge the model’s context window. Nor does a short response prove that the assistant still has every fact needed to finish its task: the next question may require another retrieval.
Two workflows deserve different names. Execution tools can compute an answer from temporary input, while ctx_index creates searchable content for repeated questions. The inspected server explicitly recommends execution over persistent indexing for one-off logs. Choose based on future reuse and retention requirements, not simply on the size of a file. Indexing sensitive text is a storage decision, even when the immediate tool response is tiny.
Separate a useful claim from a universal guarantee
The README advertises large context reductions, including an example comparing 315 KB of input with 5.4 KB entering context. Treat this as the project’s example, not our benchmark or a guaranteed token reduction. Bytes, model tokens, billed input and successful answers are different measurements. A workflow that saves response bytes but misses a rare error code can be worse for the reader’s actual debugging task.
An appropriate first evaluation uses invented material containing known facts. Ask for an ordinary fact, an unusual identifier and a negative result. Check the retrieved evidence, not only the model’s confident summary. Repeat after changing the source. This reveals whether the workflow can support investigation rather than merely produce a compact first response.
Know the integration and licensing boundary
Tool availability, automatic routing and session continuity depend on the host integration. A server can appear in an MCP tool list while the host still sends ordinary tool output directly into conversation. The deployment chapter therefore tests registration and routing separately. This review inspects a fixed source commit; it does not claim to have installed every supported host or reproduced the upstream performance demonstrations.
The inspected package identifies its license as Elastic-2.0, and LICENSE includes restrictions relevant to offering a hosted or managed service. This is source-available code, not a blanket promise of permissive reuse. It belongs in this repository-discovery column with that distinction visible. Read the actual license for your proposed use and obtain appropriate advice before a commercial service decision.
Implementation steps
- 1
Choose a harmless document with three known facts.
- 2
Decide whether repeated retrieval justifies storing it.
- 3
Compare retrieved evidence with the original before evaluating savings.
Copy-ready example
{"task":"recover a release decision","source":"invented fixture","requiredFacts":["release label","timeout","rollback rule"],"correctFacts":null,"responseBytes":null}Frequently asked questions
Is this lossless compression of the conversation?
No. Stored evidence can remain available while only selected results enter context. Retrieval quality determines which facts the assistant actually sees.
Does an MCP connection enable all host features?
No. Tool registration, hooks, routing and continuity are separate capabilities that need host-specific checks.
Sources
- README.mdSource checked 2026-09-07
- package.jsonSource checked 2026-09-07
- LICENSESource checked 2026-09-07
- src/store.tsSource checked 2026-09-07
- src/search/unified.tsSource checked 2026-09-07
- src/server.tsSource checked 2026-09-07
- src/executor.tsSource checked 2026-09-07
- src/security.tsSource checked 2026-09-07
- src/session/purge.tsSource checked 2026-09-07
- tests/store-bytecap.test.tsSource checked 2026-09-07